Cardiac MR perfusion image processing techniques: A survey

Medical Image Analysis 16 (2012) 767–785

Contents lists available at SciVerse ScienceDirect

Medical Image Analysis journal homepage: www.elsevier.com/locate/media

Survey Paper

Cardiac MR perfusion image processing techniques: A survey Vikas Gupta a,b, Hortense A. Kirisßli a,c, Emile A. Hendriks b, Rob J. van der Geest a, Martijn van de Giessen a,b, Wiro Niessen c,d, Johan H.C. Reiber a, Boudewijn P.F. Lelieveldt a,b,⇑ a

Division of Image Processing, Department of Radiology, Leiden University Medical Center, 2300 RC Leiden, The Netherlands Vision Lab, Department of Mediamatics, Delft University of Technology, 2600 GA Delft, The Netherlands c Biomedical Imaging Group Rotterdam, Erasmus Medical Center, 3000 CA Rotterdam, The Netherlands d Imaging Science and Technology, Faculty of Applied Sciences, Delft University of Technology, 2600 GA Delft, The Netherlands b

a r t i c l e

i n f o

Article history: Received 9 May 2011 Received in revised form 14 December 2011 Accepted 15 December 2011 Available online 10 January 2012 Keywords: Cardiac MR Myocardium Perfusion Survey MRI

a b s t r a c t First-pass cardiac MR perfusion (CMRP) imaging has undergone rapid technical advancements in recent years. Although the efficacy of CMRP imaging in the assessment of coronary artery diseases (CAD) has been proven, its clinical use is still limited. This limitation stems, in part, from manual interaction required to quantitatively analyze the large amount of data. This process is tedious, time-consuming, and prone to operator bias. Furthermore, acquisition and patient related image artifacts reduce the accuracy of quantitative perfusion assessment. With the advent of semi- and fully automatic image processing methods, not only the challenges posed by these artifacts have been overcome to a large extent, but a significant reduction has also been achieved in analysis time and operator bias. Despite an extensive literature on such image processing methods, to date, no survey has been performed to discuss this dynamic field. The purpose of this article is to provide an overview of the current state of the field with a categorical study, along with a future perspective on the clinical acceptance of image processing methods in the diagnosis of CAD. Ó 2011 Elsevier B.V. All rights reserved.

1. Introduction Monitoring blood circulation through the myocardium provides valuable information about the health of myocardial tissue (Marcus et al., 1987; Kaandorp et al., 2005), and helps in diagnosing coronary artery disease (CAD) at an early stage. Myocardial perfusion is a strong indicator of the functional integrity of the myocardium. As such, imaging techniques that allow non-invasive monitoring of the coronary circulation play an important role in detecting perfusion defects. SPECT and PET have been used most commonly for myocardial perfusion assessment (Hachamovitch et al., 1998; Sabharwal and Lahiri, 2003; Hesse et al., 2005; Go et al., 1990; Schwaiger, 1994; Di Carli et al., 2007), and first-pass cardiac MR perfusion (CMRP) imaging is rapidly gaining ground. The growth of CMRP imaging can be attributed to its non-ionizing property and versatility in providing insights into the functional, as well as the anatomical aspects of myocardial perfusion. A few studies have also shown that CMRP imaging compares favorably with SPECT and PET (Schwitter et al., 2001; Wagner et al., 2003; Sakuma et al., 2005). In addition, its higher spatial resolution helps in detecting, for

⇑ Corresponding author at: Division of Image Processing, Department of Radiology, Leiden University Medical Center, 2300 RC Leiden, The Netherlands. Tel.: +31 715261882; fax: +31 715266801. E-mail address: [email protected] (B.P.F. Lelieveldt). 1361-8415/$ - see front matter Ó 2011 Elsevier B.V. All rights reserved. doi:10.1016/j.media.2011.12.005

instance, sub-endocardial layer infarcts (Bache and Schwartz, 1982; Schwaiger, 1994), which are difficult to detect in SPECT and PET images. In CMRP imaging, the changes within the myocardium are studied by monitoring the first pass of a contrast agent through the heart with Electrocardiogram (ECG)-gated MR images (Fig. 1). The myocardial perfusion reserve is, generally, estimated by comparing MR images at stress and rest; stress images are acquired by using vasodilators such as adenosine or dipyridamole (Wilson et al., 1990; Ranhosky and Kempthorne-Rawson, 1990; Kozàkovà et al., 1997; Kellman and Arai, 2007). The uptake of contrast agent by myocardial tissue induces signal intensity variation due to shortening of the T1 relaxation time, which leads to a brighter signal when a T1-weighted imaging sequence is employed. An important manifestation of this signal intensity variation is the hypo-intensity in certain regions due to lower regional blood supply that might be an indicator of myocardial ischemia (Schaefer et al., 1991). The interpretation of CMRP images is therefore essential in the early diagnosis of CAD. A CMRP scan is usually followed by a quantitative analysis of the myocardial perfusion from the acquired images in order to identify and locate the abnormal or vulnerable myocardial region(s). Although the manual interaction in this analysis allows reliable detection of a CAD in a clinical setting, it is often time-consuming, tedious, and prone to interor intra-observer measurement variability. A further complication

768

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

Fig. 1. An example of first pass cardiac MR perfusion (CMRP) images: the columns show three time-frames from a CMRP sequence representing the passage of a contrast agent through the right ventricle (RV) and the left ventricle (LV). The images acquired before the contrast uptake by cardiac tissue form the leftmost column. Each row shows a different slice.

in the quantitative analysis is introduced by the acquisition and patient related image artifacts. To reduce the manual interaction, operator bias, and the effect of artifacts (e.g., motion artifacts), several semi- and fully- automated image processing methods have been developed. These methods have been shown to overcome the problems associated with image artifacts and to significantly reduce both the measurement variability and the analysis time. Rapid developments in the field of CMRP image processing have resulted in an extensive literature over the last two decades. The objective of this paper is to provide a survey of such a literature. Our methodology involved the screening of all the publications related to CMRP image processing in last 20 years (until November 2010) with following keywords: mr perfusion, cardiac mr perfusion and first-pass myocardial perfusion. All major international journals and conference proceedings were screened for the papers that were within the scope of this survey. This ensures a comprehensive coverage of all the peer-reviewed literature that is available online. However, in those cases wherein more than one article was published by the same author(s) on the same method (e.g., a conference paper extended into a journal paper), we selected the latest of the published papers, preferably a journal publication. The organization of this article is as follows. In Section 2, we briefly introduce CMRP image acquisition components, perfusion quantification, and the major artifacts related to CMRP images. Sections 3–6 form the core of this article wherein we categorize the existing literature, and discuss the main ideas of surveyed methods within their respective categories. In Section 7, we provide a summary of the current and future trends in CMRP image processing, and conclude with the future perspectives on their clinical acceptance. 2. Background 2.1. CMRP image acquisition A typical acquisition procedure for CMRP images involves a combination of three major components: a contrast agent, a pre-pulse

scheme, and an imaging sequence. The use of an external contrast enhancing substance, such as gadolinium, was first introduced in the late 1980s (Brown and Higgins, 1988; Miller et al., 1989). This was soon followed by the development of a technique by Atkinson et al. (1990) in 1990, which allowed the dynamic evaluation of perfusion images using first-pass of the contrast agent. Since then, MR has been employed extensively to study myocardial perfusion both in laboratory animals and in humans (Pijls et al., 1990; Burstein et al., 1991; Manning et al., 1991; Schaefer et al., 1991; Klein et al., 1993; Minoves et al., 1993; Eichenberger et al., 1994; Tsekos et al., 1995; Larsson et al., 1996; Hiller et al., 1998; Bauer, 1998; Vallée et al., 1998; Cullen et al., 1999; Frank and Globits, 1999; Penzkoffer et al., 1999; Zavaljevski et al., 1999; Bardonova et al., 2000; Al-Saadi et al., 2000a). Contrast in CMRP images can be induced either with the injection of exogenous (originating from outside the body) contrast agents or endogenously using a physiological phenomenon. One of the most widely used exogenous contrast agents is gadolinium which, due to its extreme toxicity in free form, is always chelated. Contrast agents, such as Gd-DTPA (Gadolinium-Diethylene Triamine Pentaacetic Acid), alter the local magnetic field, and thereby enhance the relaxation rate of water protons in close proximity to the contrast agent. This results in an enhanced signal from the tissues that are perfused (Weinmann et al., 1984; Wesbey et al., 1984; Runge et al., 1985). Nevertheless, an analysis based on the assumption that signal intensity enhancement within the myocardium depends on pure myocardial perfusion can result in an inaccurate diagnosis, as about half of the contrast agent leaks into the interstitial space during the first pass (Tong et al., 1993a; Tong et al., 1993b). It must therefore be taken into account during quantitative image analysis that the myocardial signal intensity does not depend only on tissue blood volume and perfusion, but also on diffusion. In case a gadolinium based contrast agent is employed, patients with chronic renal insufficiency should be identified, and prophylactic measures should be taken to reduce the risk of nephrotoxicity (Sam et al., 2003; Thomsen, 2004; Bellin and Molen, 2008). Endogenous contrast based approaches use blood flow and oxygenated related endogenous contrast instead of the contrast induced by an exogenous substance. Most prominent among them use spin labeling (Fidler et al., 2004), blood oxygen level dependent (BOLD) contrast (Fieno et al., 2004; Klocke and Li, 2003) and magnetization transfer contrast (Wolff and Balaban, 1989; Wolff et al., 1991). Table 1 summarizes widely used pre-pulse schemes and imaging sequence protocols along with their advantages and disadvantages. Pre-pulse or magnetization preparation schemes are used to modify T1 contrast or suppress signal from specific tissues or regions (Plein et al., 2010). As shown in the table, inversion recovery and saturation recovery are the two main types of pre-pulse schemes. In the earlier days of CMRP imaging (1990–1995), the inversion recovery pre-pulse was frequently used due to its ability to produce a higher contrast between hypo- and hyper-perfused tissues. However, despite its higher contrast producing abilities, the inversion recovery scheme suffered from longer acquisition times, limited temporal resolution and lesser reliability in case of arrhythmias. With the introduction of saturation recovery in 1995 (Tsekos et al., 1995), the disadvantages of inversion recovery were overcome at the expense of reduced signal contrast. Moreover, due to its insensitivity to arrhythmias and shorter acquisition time, it became possible to incorporate multi-slice coverage, thereby improving temporal resolution. Usually, 3–6 slices are acquired with a temporal resolution of 150 milliseconds/slice (varying with number of slices and the heart rate) to enable the detection of large-vessel CAD. Images at each slice location are generally acquired over 40–60 cardiac cycles, which results in the corresponding number of phases for each slice. The z-resolution is limited by

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

769

Table 1 Advantages and disadvantages of the commonly used image acquisition parameters. Type

Advantages

Disadvantages

High signal contrast between hypo- and normally perfused tissues

Less reliable in arrhythmias Long acquisition time Low temporal resolution

Shorter acquisition window High temporal resolution Insensitive to arrhythmias

Less signal contrast between hypo- and normally perfused tissues

Shorter repetition times

Lower SNR and CNR due to lower flip angles Long acquisition time per image limits maximum heart rate for the acquisition

Balanced steady state free precision (bSSFP)

Highest CNR

Most susceptible to artifacts, especially dark banding artifacts

Hybrid echo planar imaging (EPI)

Higher CNR due to higher flip angles Fast imaging due to shorter acquisition time per image

Low SNR Ghosting effects

Pre-pulse Inversion recovery

Saturation recovery

Imaging sequence Spoiled gradient-recalled echo (GRE)

Fig. 2. Typical CMRP image acquisition process: the cardiac contraction is frozen by synchronizing the image acquisition to the heart cycle using the R-wave on the ECG as a trigger signal for the scanner.

the number of slices that can be acquired within one cardiac cycle. The spatial resolution of CMRP images depends on various factors such as spatial coverage, heart rate, field of view, imaging sequence, etc., which are varied to achieve an adequate spatial resolution for the detection of sub-endocardial defects (<3 mm in-plane) and the assessment of their transmurality (Kellman and Arai, 2007); a typical image therefore has an in-plane spatial resolution of 2–3 mm. Another important component of CMRP image acquisition is the image sequence protocol as shown in Fig. 2 and Table 1. The most commonly used sequences for CMRP imaging are: spoiled gradient-recalled echo (GRE), balanced Steady state free precision (bSSFP), and Hybrid echo-planar imaging (EPI) (Plein et al., 2010). In addition, faster image acquisition and higher spatial resolution can be achieved by using parallel imaging (e.g., with spoiled GRE) as it exploits the spatial distribution of array coil elements in reducing the amount of phase encoding (Plein et al., 2010) steps (e.g., k-space sampling) required to reconstruct the image. This reconstruction is performed either in the k-space with simultaneous acquisition of spatial harmonics (SMASH) (Sodickson and Manning, 1997), or in image domain using SENSE (Pruessmann

et al., 1999). Presently, CMRP image acquisition is most commonly performed at 1.5 Tesla, but acquisitions at 3.0 Tesla are becoming increasingly popular as a higher field strength has been shown to improve both, contrast to noise (CNR) and signal to noise ratios (SNR) (Cheng et al., 2007). Recent technical advancements notwithstanding, the image acquisition components are also associated with specific image artifacts (Section 2.3). The contribution of these artifact mechanisms is generally influenced by the choice of imaging sequence as well as the other acquisition components, such as pre-pulse scheme and contrast agent. Moreover, with technical developments, such as high field strengths (e.g., 3 T), potential of artifacts in the images may also increase significantly. 2.2. Perfusion quantification Perfusion quantification refers to the assessment of myocardial blood flow with quantitative, observer-independent, and reproducible measures, which can provide new insights into the etiology of CADs. These measures rely on the rate at which the contrast agent arrives in myocardial tissue and the signal enhancement

770

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

induced by this arrival. Typically, the myocardial blood flow is quoted as arterial input rate (mL/min) per gram of tissue, which gives mL/min per g of tissue as the unit of myocardial blood flow (Jerosch-Herold, 2010). Two kinds of approaches exist for perfusion quantification: (a) semi-quantitative analysis, and (b) absolute quantification. 2.2.1. Semi-quantitative analysis Semi-quantitative analysis of myocardial perfusion has been shown to improve the sensitivity and specificity of CMRP image analysis in detecting significant stenoses (Kivelitz et al., 1997; Cullen et al., 1999; Nagel et al., 1998; Wilke and Jerosch-Herold, 1998). Such an analysis can be performed with the perfusion parameters that are obtained by analyzing the time-intensity curves (TIC). A TIC (Fig. 3) represents the sampled myocardial intensity in a CMRP image sequence for all the time points within a region of interest (ROI). The initial constant segment of this TIC is called the baseline, which represents the time points before the arrival of contrast agent into the ventricles. The most commonly used perfusion parameters are provided in the following list; among these, the up-slope and myocardial perfusion reserve index have been considered as the most relevant semi-quantitative parameters for the detection of a perfusion abnormality. a. Peak signal intensity (SIpeak): The peak value of the TIC attained during the first pass of the contrast agent, relative to the baseline before contrast enhancement (Keijer et al., 1995). b. Up-slope: The first derivative of the TIC during the initial ascent of the first pass. The up-slope of the myocardium is usually normalized by the up-slope of LV blood pool to correct for differences of the speed and compactness of the contrast agent bolus (Al-Saadi et al., 2000a; Schwitter et al., 2001; Nagel et al., 2003). c. Time to peak (Tpeak): The time from the onset of contrast enhancement (called as the ‘foot’ to the SIpeak). d. Mean transit time (Tmean): The average time required for a contrast agent to pass through the ROI (Keijer et al., 1995). e. Area under the TIC (ATIC): The area under the TIC from the beginning of its ascent to a user-defined point. f. Myocardial perfusion reserve index (MPRI): The MPRI is defined as the ratio of hyperemic and basal myocardial blood flow. Generally, MPRI is estimated as the ratio of the normalized up-slopes measured for rest and stress images (Al-Saadi et al., 2000b; Jerosch-Herold et al., 2004). A recurring problem with the computation of these parameters is the disruptive changes in the TIC due to noise and the artifacts

Fig. 3. A time-intensity curve (TIC) and semi-quantitative perfusion parameters.

such as Gibbs ringing (Jerosch-Herold, 2010). As a result, the distortion of the TIC may also introduce ambiguities in selecting the points for the computation of the baseline, the foot, and the upslope, which subsequently leads to incorrect estimation of other semi-quantitative parameters (e.g., MPRI). It is therefore important that the user must be aware of these pitfalls during the semi-quantitative analysis of CMRP images. 2.2.2. Absolute perfusion quantification Absolute quantification is often performed to gain more insights into the coronary physiology as the global reduction in patients with multi-vessel disease makes the semi-quantitative analysis inadequate (Attili et al., 2010). Moreover, the absolute quantification was found to be more closely correlated with the gold standard for regional tissue perfusion, i.e., microsphere blood flow, than the semi-quantitative perfusion parameters (Jerosch-Herold et al., 2002; Mühling et al., 2003; Christian et al., 2008; Christian et al., 2004). Absolute quantification approaches can be either: (a) model-based, or (b) model independent; the model refers to spaces in myocardial tissue structure and their behavior defined by a set of mathematical functions (Lee and Johnson, 2009; Jerosch-Herold, 2010). Whereas model-based approaches are often based on compartmental analysis, model-independent approaches are based on central volume principle (Zierler, 1962; Zierler, 1965; Zierler, 2000). For more details on quantification approaches to assess myocardial blood flow, we encourage the reader to refer to the comprehensive review articles by Jerosch-Herold (2010), Attili et al. (2010), and Christiansen et al. (2010). 2.3. CMRP image artifacts CMRP images are often associated with one or more artifacts that can distort the quantitative image analysis results. While some of those artifacts are attributed to the acquisition system, others are related to the patient itself (e.g., motion artifacts). Acquisition related artifacts (e.g., surface coil inhomogeneity, dark rim artifacts) may alter the time intensity curves, and hence, prevent accurate quantitative analysis. Following are the major artifacts associated with CMRP images and the corresponding acquisition based solutions: 2.3.1. Surface coil inhomogeneity Surface coils are used to improve the SNR in MR images. However, as the sensitivity of the coil falls rapidly with distance from the coil, intensity inhomogeneities are introduced in the images. These inhomogeneities can have a significant effect on the quality of acquired images (Axel et al., 1987; Belaroussi et al., 2006). The surface coil induced intensity inhomogeneity can partially be compensated by using two surface coils on either side of the body. A commonly used solution, specific to myocardial perfusion imaging, involves the use of adiabatic (B1-insensitive), RF inversion or saturation pulses, or composite pulses (Jerosch-Herold, 2010). Postacquisition, the CMRP images can be corrected for intrinsic spatial signal intensity variations within the myocardium by baseline correction, which involves the normalization of myocardial signal by its pre-contrast value (Jerosch-Herold, 2010). More details on the intensity inhomogeneity correction methods for MR images can be found in the review articles by Belaroussi et al. (2006),Kellman and Arai (2007) and Jerosch-Herold (2010). 2.3.2. Dark rim artifacts Dark rim artifacts (DRA) appear as a dark sub-endocardial rim that may be confused with actual hypo-intense regions as a result of reduced blood flow. There have been several studies on the causes of DRA (Di Bella et al., 2005; Schreiber et al., 2002; Arai,

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

2000; Albert et al., 1993; Storey et al., 2002), and the following have been agreed upon as the most probable: a. Gibbs ringing is usually caused by truncation or the non-uniformity of the k-space weighting; the non-uniformity across kspace leads to point spread function (PSF) distortion causing both blurring (loss of spatial resolution) and edge enhancement (Di Bella et al., 2005). The Gibb’s ringing can be partially mitigated at the expense of spatial resolution by windowing or apodization, which involves the k-space filtering with a window function (Kellman and Arai, 2007; Gerber et al., 2008). b. Partial volume effect arises due to the spill-over of signals from neighboring tissues. It may also give rise to dark rim artifacts at the border of the LV blood pool and the endocardium (Di Bella et al., 2005; Kellman and Arai, 2007), thus making the segmentation of endocardial wall more difficult. A solution to this problem was suggested and validated by Köstler et al. (2004) who added a ‘‘spillover’’ term to their myocardial contrast enhancement model. The ‘‘spillover’’ term was calculated by applying variable scaling and timeshift to the arterial input function. c. Band effect occurs at the tissue boundaries when the object moves between the acquisitions of different phase encoding lines (Storey et al., 2002). These effects can be mitigated by shim re-adjustment (Kellman and Arai, 2007) or parallel imaging (Gerber et al., 2008). 2.3.3. Motion artifacts The displacement or deformation of the heart during CMRP image acquisition leads to motion artifacts. These artifacts are introduced mainly due to: (a) cardiac motion during pumping of the blood, and (b) respiratory motion during image acquisition. In general, motion artifacts introduce blurring, ghosting, and mis-registrations (Schultz et al., 1984). a. Cardiac motion artifacts These artifacts are eliminated by means of ECG gating and by acquiring the images during the filling phase of the cardiac cycle when the heart is relatively stationary. This phase begins at the R wave of the ECG after the relatively heart-rate invariant systolic phase (Weissler et al., 1968). Nevertheless, there still remains a possibility of the triggering faults that may result in blurring or deformation of the heart image. b. Respiratory motion artifacts Respiratory motion induced artifacts are the most commonly observed patient related artifacts in CMRP images. Several studies have been conducted to determine the effect of respiration (Holland et al., 1998; Scott et al., 2009) and to quantify the rigid and non-rigid motion of the heart with respiration (McLeish et al., 2002). McLeish et al. (2002) showed that the motion of the heart is predominantly linear with diaphragm in the superior-inferior direction. The extent of this movement, however, varies substantially with respiratory phases (inspiratory or expiratory), diaphragm location, and between patients. Furthermore, as shown by Wang et al. (1995), respiratory motion of the heart can be decomposed into uniform and non-uniform constituents; while the former is roughly approximated as global translation, the latter is the difference between motion at the center and edge of the field of view. For a more comprehensive discussion on motion in cardiovascular imaging, the reader is referred to the detailed review article by Scott et al. (2009). Although it is possible to perform CMRP image acquisition on freebreathing patients, breath-hold based acquisition is more popular as it allows for the compensation of respiratory motion artifacts.

771

Table 2 Abbreviations and notations. Name

Description

AAM ASSERT

Active Appearance Modeling Adaptive Semi-automatic and Self-evaluated Registration Technique Bull’s Eye Plot Coronary Artery Disease Cross-Correlation Cardiac Magnetic Resonance Perfusion Computed Tomography Computed Tomography based Coronary Angiography Delayed Enhancement MR Echo-Planar Imaging A Flexible Appearance Modeling Environment Factor Analysis for Medical Image Segmentation Free-form Deformation Gradient Recalled Echo Gradient Vector Flow Independent Component Analysis Left ventricle Minimum Description Length Mutual Information Myocardial Perfusion Reserve Index Magnetic Resonance Imaging Mean Squared Differences Normalized Gradient Fields Normalized Mutual Information Positron Emission Tomography Proton Density Partial Least Squares Regression Partial Volume Effect Right Ventricle Standard Deviation Sensitivity Encoding for fast MRI Simultaneous Acquisition of Spatial Harmonics Single Photon Emission Computed Tomography Sum of Squared Differences Balanced Steady State Free Precession Stochastic Active Contour Scheme Time Intensity Curve Volumetric Bull’s Eye Plot Whole Heart MR

BEP CAD CC CMRP CT CTCA DE-MR EPI FAME FAMIS FFD GRE GVF ICA LV MDL MI MPRI MRI MSD NGF NMI PET PD PLSR PVE RV SD SENSE SMASH SPECT SSD bSSFP STACS TIC VBEP WH-MR

For a breath-hold based acquisition, the patient is asked to hold his/her breath for 45–50 seconds. In addition, ECG based triggering ensures that the motion of the heart is effectively frozen during the acquisition of images over multiple cardiac cycles. However, a prolonged single breath hold in many patients, especially those with CADs, leads to sudden respiration during the image acquisition, and as a consequence, to displacement and/or deformation of the heart. The limitations of acquisition based strategies in eliminating motion artifacts necessitated the development of post-processing techniques, which apply corrections in either k-space, such as autofocus (Atkinson et al., 1997), motion decoupling (Yang et al., 1998) etc., or in image domain (Sections 3–5). In general, image domain techniques have been shown to perform significantly better as they have inherently less limitations in correcting complex spatial domain deformations than the k-space based techniques (McLeish et al., 2002). The detailed discussion on such image domain based techniques is provided in subsequent sections.

3. Classification of CMRP image processing methods Interpretation of CMRP images requires image processing methods that help in effectively extracting and quantifying the diagnostic information. More specifically, these methods should be able to: (a) ensure the spatial correspondence between the

772

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

images in the CMRP sequence, (b) address the problems associated with the presence of anatomical structures (e.g., lungs, diaphragm, trabeculae, etc.) and low contrast around the myocardium, and (c) distinguish the perfusion defect induced intensity variations from those due to noise and artifacts. To address all or most of these requirements, several methods have been proposed. The majority of these methods involve either a registration or a segmentation component, but a few methods also employ both in succession or in a feedback loop. Registration is required to ensure the spatial correspondence (to eliminate motion artifacts) and segmentation helps in the semi- or fully automatic identification and delineation of the region of interest for diagnostic assessment. Visualization of the extracted information after registration and segmentation is another task that has gained popularity in recent years. The importance of visualization techniques has been enhanced by the requirement of multi-modality information fusion to present multiple parameters (e.g., perfusion parameters from CMRP images and structural information from Computed Tomography based Coronary Angiography (CTCA) images) in an integrated yet intuitive display. To present all the methods in a systematic fashion, we have classified CMRP image processing methods into two broad categories on the basis of the major components i.e., registration and segmentation. In addition, a separate category has been dedicated to the discussion of multi-modality information fusion and visualization approaches. All these categories are followed by a summary of the common traits observed in the reviewed approaches along with their advantages and disadvantages. It must be noted that in the event of one method consisting of both a registration and a segmentation technique, it is included in the category closer to its major component. 4. Methods based on registration Registration involves aligning the different images of the same object into strict spatial congruence by applying a suitable transformation. Fig. 4 shows a typical registration pipeline. Methods to geometrically align CMRP images use either: (a) rigid transformation based registration, or (b) non-rigid/elastic transformation based registration. This spatial alignment is performed with a target image that has to be transformed and a fixed reference image by using an optimizer. The optimizer (e.g., least squares, gradient descent, etc.) controls the iterative registration process to achieve a desired spatial correspondence between the target and the reference images. If the transformation results in points that do not coincide with the regular grid, an interpolation scheme is required to determine the intensity value of those points at non-grid positions. Among the most common interpolation strategies are: bilinear, bicubic, splines, and nearest neighbor interpolation.

Registration schemes can often be computationally expensive and prone to the influence of undesired regions in the target and reference images. Limiting the processing to a region of interest (ROI) helps in eliminating the interference from surrounding structures thereby increasing the efficiency of the registration. The accuracy of spatial correspondence or registration between the transformed target and the reference images is provided by the similarity metric, which is a major constituent of a typical registration pipeline (Fig. 4). In some cases, a similarity function controls both the evaluation of the quality of registration and the optimization procedure. All the registration based methods are enlisted in Tables 3 and 4, along with their constituents discussed earlier. This table also provides a quick overview of the automation extent (semi or full), validation type (qualitative or quantitative), and the number and type (free breathing or breath-hold) of validation datasets. The rigid and non-rigid registration methods can be further classified into feature and intensity based methods. Feature based methods employ features such as points, surface, landmarks, regions, etc. to maximize the spatial correspondence between the reference and target images. On the other hand, intensity based approaches operate directly on image intensities or their occurrences without prior identification and extraction of features. In most cases, a relation is assumed between intensity values in the sequence and the similarity metric is computed only for the overlapping area between the images being registered. 4.1. Rigid transformations based registration methods Rigid transformations preserve all distances and angles. Methods based on rigid transformations rely on the assumption that motion artifacts lead predominantly to in-plane motion of the object being imaged. The effect of through-plane motion is often not considered due to the sparsity of slices (limited by the image acquisition system). Assuming the non-rigid or elastic deformations of the myocardium in CMRP images to be insignificant, rigid transformation offers a simple yet efficient solution to compensate for translation and rotation of the myocardium over time. 4.1.1. Feature based Rigid transformation based alignment of CMRP images was first employed by Yang et al. (1998), who used raw k-space MR data as opposed to the reconstructed data. In their two-step approach, lateral motion correction using the inter frame phase differences of the k-space data was followed by the correction of myocardial boundaries using a cubic B-spline based deformation model; the former step achieved a higher improvement than the latter. Gradient-based approaches have been applied by several authors. In the approach proposed by Delzescaux et al. (2001, 2003), the reference image was obtained by computing the maxima

Fig. 4. A typical registration scheme.

Table 3 Registration based methods. Authors (year)

Validation

Data (type)

– Potential criterion based

Semi Semi

QL, QT QL,QT

20 11 (breath-hold)

Joint conditional entropy

–

Semi

QL

2

Cross-correlation

–

–

Semi

QT

Displacement Cross-correlation

Isobarycentric, Least squares –

– –

Semi Full

QL, QT QT

10 (breath-hold followed by free breathing) 6 35 (breath-hold)

Mean squared difference Normalized cross-correlation

Least squares –

Bicubic Trilinear

Semi Semi

QL, QT QL

8 12 (breath-hold)

Mutual information Mean squared difference Mean squared difference Image gradient Normalized mutual information Mean squared difference Mutual information Mutual information

Gradient descent – M-estimators – – – Levenberg–Marquardt –

Linear – Bicubic – – – – –

Semi Semi Full Semi Semi Semi Semi Full

QL, QL, QL, QL QL, QL, QL, QL,

5 8 (free breathing) 4 15 15 10 (breath-hold) 10 11

– – – –

– – – –

Full Semi Full Full

QL, QT QL QL, QT QL, QT

12 (free breathing) 9 10 (free breathing) 10 (free breathing)

Tautz et al. (2010) Wollny et al. (2010)

Intensity based statistical measure – Minimum description length Minimum description length, Normalized mutual information Phase-difference Normalized gradient field/sum of squared differences

– Variants of gradient-descent and Levenberg– Marquardt

– B-splines

Full Full

QL, QT QL, QT

8 6 (free breathing)

Intensity based Ablitt et al. (2002) Rougon et al. (2005a) Li and Sun (2009) Xue et al. (2008) Xue et al. (2009)

Localized cross-correlation Regional correlation and f-information Custom energy functional Normalized mutual information Localized cross-correlation

Quasi-Newton Gradient-descent Expectation-minimization Simple downhill –

B-splines – – B-splines B-splines

Semi Full Semi Semi Semi

QL QL QL, QT QL, QT QL, QT

8 (breath-hold) 15 9 76 40

Comte et al. (2004) Milles et al. (2008) Intensity based Bidaut and Vallée (2001) Breeuwer (2001), Breeuwer et al. (2001) Bracoud et al. (2003) Dornier et al. (2003) Kawakami et al. (2004) Sun et al. (2004) Wong et al. (2005) Adluru et al. (2006a) Kocher et al. (2007) Schöllhuber (2008) Non-rigid/elastic Feature based Gallippi et al. (2002) Gao et al. (2002) Stegmann et al. (2005) Ólafsdóttir et al. (2006)

Optimization

Interpolation

Inter-frame phase difference Spatial gradient norm

Least squares –

Mutual information

QT QT QT QT QT QT QT

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

Auto

Rigid Feature based Yang et al. (1998) Delzescaux et al. (2001), Delzescaux et al. (2003) Bansal and Funka-Lea (2002) Gupta et al. (2003)

Method components Similarity metric

QL: Qualitative, QT: Quantitative; see Table 2 for the abbreviations.

773

774

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

Table 4 Segmentation based methods. Authors (year)

Methods (salient features)

Auto

Validation

Data (type)

Model independent segmentation Behloul et al. (1997) Bella and Sitek (2001) Spreeuwers et al. (2001)

Fuzzy logic, Fuzzy inference system Factor analysis, k-means clustering Region growing, Polar transformation, Snakes

Semi Semi Full

QL QL, QT QL

– 2 14

Model based segmentation Boudraa et al. (2001)

Parametric modeling

Semi

QT

Janier et al. (2002) Santarelli (2003) Discher et al. (2005) Pluempitiwiriyawej and Sotthivirat (2005) Adluru et al. (2006b) Lorenzo et al. (2006) Baka et al. (2008)

Factor analysis in medical image segmentation Gradient vector flow snakes, Anisotropic filtering Bi-ventricular heart template deformation, Level sets Stochastic active contour scheme, Thresholding, Median filtering, Morphological operations (closing) Level sets Shape modeling, Level sets Active appearance motion model

Semi Full Full Full

QL QL, QT QL, QT QL

22 (breathhold) 13 9 15 1

Semi Semi Semi

QL, QT QL QL, QT

Gupta et al. (2010)

Independent component analysis, Active appearance modeling

Full

QL, QT

8 16 19 (breathhold) 18 (breathhold)

QL: Qualitative, QT: Quantitative.

of the norm of the spatial gradient of images in a CMRP image sequence. On the reference image thus selected, RV, LV and pericardial contours were traced (Delzescaux et al., 2001) before the application of an adaptive semi-automatic and self-evaluated registration technique (ASSERT) (Delzescaux et al., 2001) to reduce in-plane rigid displacements while rejecting the images with nonrigid deformations. These contours were used to build seven registration models and the transformation corresponding to each model was estimated based on potential criterion optimization. For each image, the optimal registration model was selected on the basis of a superimposition score. Following registration, factor images representing myocardial perfusion were obtained by analyzing the registered images with a method called factor analysis for medical image segmentation (FAMIS) (Frouin et al., 1992; Janier et al., 2002). More details on FAMIS are provided in Section 5.2. Also employing a gradient based strategy, Bansal and Funka-Lea (2002) used an information theoretic registration framework that integrated two channels of information: the pixel intensities and the local gradient information. An ROI defined by myocardial contours on one image was correlated with other images in the sequence. This correlation was performed by estimating edge probabilities in the target images and modeling with 2D Markov random fields (MRF) (Geman and Geman, 1993; Li, 1994). As discussed earlier, an ROI is often used to localize the LV, either manually or automatically. To avoid the tedious manual approach, Gupta et al. (2003) localized the LV by utilizing its center of mass and used correlation as the similarity metric. However, as iterations involved in correlation computation may lead to time inefficiency, the images were first converted into binary maps with a threshold-based normalization technique. The constrained search around the center of mass and binary-map thresholding led to a higher robustness in comparison to global search and the subsequent simplification of correlation calculations, respectively. In contrast to the approach of Gupta et al. (2003), the LV was localized manually with an ROI enclosing epi- and endocardial contours in the registration method of Comte et al. (2004). The ROI thus defined was used to generate the reference TICs for the validation. The registration was based on the precise location and tracking of the interface between the left lung and the heart, which served as an anatomical landmark. While the detection of the left lung was performed with region growing and morphological operations, the contour of the heart–lung interface was detected using a Laplacian filter. Each image in the sequence was subsequently aligned to reference image by matching it with reference interface

and shifting it on the basis of the displacement calculated using two methods: (i) isobarycentric method, (ii) least squares method; the former was rapid but did not consider rotation. This method was limited by the poor heart–lung interface detection. Milles et al. (2008) incorporated time-intensity variation awareness into their method by applying independent component analysis (ICA) (Comon, 1994; Hyvärinen and Oja, 2000) to extract physiologically relevant features together with their time-intensity behavior. The three feature images, one each for the RV, the LV and the baseline, were used to compute a time-varying reference image mimicking intensity changes in the CMRP image sequence. With cross-correlation (CC) as the similarity metric, rigid registration was applied for each slice in a multi-pass and multi-resolution scheme. 4.1.2. Intensity based Bidaut and Vallée (2001) selected one of the images corresponding to the baseline of TIC as the reference image, with which the rest of the images in the sequence were matched using pixel-based mean squared differences (MSD) as the similarity metric. In this automatic method, MSD was minimized using a least squares iterative scheme, and a multi-resolution (coarse-to-fine) approach was employed to accelerate the registration process. Until the convergence, the registration process was iterated with an average of all registered images from the last iteration as the new reference image. Rather than using a reference image, Breeuwer (2001) and Breeuwer et al. (2001) applied the automatic registration on pairs of two successive images in the image sequence considering the information only within an ROI, which was also detected automatically using the method proposed by Spreeuwers et al. (2001). The myocardial boundaries were manually drawn on either one arbitrary image or on the maximum intensity projection (MIP) image from the sequence, and were then propagated to the remaining images in the sequence. As an alternative, a completely automatic myocardial boundary detection method (Spreeuwers et al., 2001) was also applied by using constraints on the shape of the myocardium. In contrast to the methods that include only the LV within the ROI, the rectangular ROIs used by Bracoud et al. (2003) and Wong et al. (2005) included both the ventricles. While Bracoud et al. (2003) used a reference image (with constant intensity) in their registration method, Wong et al. (2005) applied pair-wise image registration. In both the cases, mutual information (MI) served as the similarity metric; Wong et al. (2005) used its normalized form i.e., normalized mutual information (NMI). The rectangular ROI,

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

encompassing both the ventricles (Netsch and Muiswinkel, 2002) was also employed for comparison with a conventional ROI (including only the LV cavity) by Dornier et al. (2003); the latter was called an optimal mask. The rigid registration in their method was performed with a continuous model based on B-splines of degree N (Thevenaz et al., 2002). Similar to the approach used by Gupta et al. (2003), position of the LV was assessed by computing its center of mass within the optimal mask. In addition to the residual motion correction of the LV with a registration algorithm, a compartmental model was also used to accurately evaluate the tissue perfusion inside the LV. A compartmental model was also employed by Adluru et al. (2006a), but for generating a model image corresponding to each image in the unregistered sequence. The rigid registration was performed twice: first, between the reference image (with optimal contrast between the myocardium and the LV blood pool) and the other images of the sequence within a rectangular mask encompassing only the heart, and later, between the corresponding registered and model images. The model images were obtained after fitting the registered image data (from the first registration) to a two compartment model. For both the instances of registration, MSD was used as the similarity metric. Gradient-based non-rigid registration strategies have been proposed by Kawakami et al. (2004) and Sun et al. (2004). Kawakami et al. (2004) approached the problem of motion correction with a gradient-based image registration method – developed by Nestares and Heeger (2000) for automatic alignment of MRI volumes – that allowed a robust estimation of displacement parameters. An M-estimator was used along with the alignment algorithm proposed by Bidaut and Vallée (2001) to increase the robustness of this estimation (Nestares and Heeger, 2000). Similarly, Sun et al. (2004) introduced an image gradient based similarity metric to reduce the effects of high contrast variation in both cardiac and renal MR perfusion image sequences. Here, the ROI – a bounding box – and the reference image were selected manually before identifying the large translations by tracking the ROI with integer pixel shifts. The rigid registration was assisted by shape priors that were obtained using segmentation with the coupled propagation of epi- and endocardial contours (Paragios, 2003). Shape priors were also used by Schöllhuber (2008) in a novel hierarchical pattern matching approach for a robust detection of the heart in CMRP images. The reliance on time-intensity variations was circumvented by locating the heart in distinct time steps. Prior knowledge of the anatomy (diameter and shape of the myocardium) and contrast agent induced intensity variations (minimal gray level variances) were exploited in a hierarchical pattern to locate the best candidate as the ROI with minimal computational costs. The frame at which the heart was detected served as the reference image, and similar to Bracoud et al. (2003) and Wong et al. (2005), MI was used as the similarity metric. A particularly vexing problem in the analysis of CMRP images is the presence of PVE (see Section 2.3.2), which can lead to unsharp myocardial boundaries – especially epicardial borders- due to dark rim artifacts. To this end, Kocher et al. (2007) focused on correcting for PVEs in mixed voxels containing both the myocardial tissue and blood. Similar to several other approaches, the first step in their proposed process involved 2D rigid registration (similarity metric: MI) to remove motion artifacts. From the registered images, the arterial input function (AIF) was estimated by applying a spatial filter within an ROI. This was followed by the investigation of two 1-compartment models for the reduction of PVE. 4.2. Non-rigid/elastic transformations based registration Several authors argue that, in addition to in-plane motion induced rigid transformations (translations and rotations), non-uniform

775

cardiac motion and respiration induced through-plane motion may also lead to significant non-rigid/elastic changes (e.g., shape deformation) in the left ventricular geometry (McLeish et al., 2002; Ablitt et al., 2002; Gallippi et al., 2002; Scott et al., 2009; Tautz et al., 2010). To correct for these deformations, non-rigid transformation based registration techniques are required. 4.2.1. Feature based methods Landmarks were used as features by Gallippi et al. (2002), Stegmann et al. (2005) and Ólafsdóttir et al. (2006). In the method proposed by Gallippi et al. (2002), the translations and geometric distortions were corrected using an automatic image registration method that involved the edge-based (3  3 Sobel filter) landmark detection of a few landmarks in the neighborhood of high spatial frequencies. This detection was performed on a template or reference image (center image of the perfusion sequence). Subsequently, each image in the sequence was warped to the template image using a score based on local intensity variation extents, means, variances, and edge directions. For each pixel in the registered perfusion sequence, seven parameter maps were generated using the corresponding TICs. Stegmann et al. (2005) utilized variance and clustering properties in an annotated training set to develop an unsupervised registration strategy, called cluster aware active appearance model. Here, the time-intensity variations in CMRP images were modeled with a slice-coupled active appearance model (AAM) (Cootes et al., 1998), which was augmented with an minimum description length (MDL) framework (Davies et al., 2002) for the optimization of landmark correspondences. AAMs provide the advantage of shape and texture based learning along with fast computations, as indicated in the study conducted by Ólafsdóttir et al. (2006) to compare AAM based approaches with free-form deformation (FFD) based non-rigid registration algorithms. Both the approaches showed similar results in terms of point-to-point errors, but latter performed better when compared on the basis of point-to-curve errors. The advantage of a training procedure was evident from the higher precision values obtained with AAMs in comparison to FFD-based approaches. Not relying on landmarks, Gao et al. (2002) introduced a deformation model based on partial least square regression (PSLR) (Wold, 1975) to extract the intrinsic correlations between the latent factors of both the input (intensity distributions around the chest wall and diaphragm) and the observed output (motion vectors of myocardial deformation). This allowed the extraction of intra-frame tissue deformation, thereby facilitating free-form registration with a dense displacement vector field within the image plane.Tautz et al. (2010) incorporated a multi-scale approach in their non-rigid registration method based on phase-based displacement estimation (Fleet and Jepson, 1990; Mellor and Brady, 2005; Mellor and Brady, 2004; Knutsson and Andersson, 2005) and the Demons algorithm (Thirion, 1998). To estimate non-rigid deformations, the displacement estimation was implemented in a coarse-to-fine dyadic scale space and the registration was iterated three times for successive images in the temporal sequence on each scale to refine the estimation, with the central frame of the sequence serving as the reference image. As discussed earlier (Section 2.3.3), CMRP images can also be acquired without a breath-hold constraint. In contrast to a breathhold CMRP image sequence, a free-breathing sequence exhibits quasi-periodicity (Wollny et al., 2010). This quasi-periodicity was exploited by Wollny et al. (2010) for decoupling breathing motion from intensity variation. Their automatic method involved two steps: (i) identification and registration of a subset of images corresponding to the same phase of the breathing cycle, and (ii) registration of the remaining images to synthetic references, created by linearly combining the registered images from the first step.

776

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

4.2.2. Intensity based methods The approach of Ablitt et al. (2002) to non-rigid registration involved a rapid multi-resolution free-form registration method (Veeser et al., 2001) using localized cross-correlation (LCC) as the similarity metric, and the last image in the perfusion sequence as the reference image. Prior to non-rigid registration, the prospective through-plane motion of the LV was corrected with tissue tagging. The remaining in-plane distortion of the heart was removed by recovering the intrinsic relationship between respiration and cardiac deformation using PLSR. Subsequently, perfusion analysis of the registered images was performed with factor analysis by extracting tracer characteristics of different myocardial regions (Benali et al., 1993; Buvat et al., 1998; Martel et al., 2001). Rougon et al. (2005a) proposed a region-driven approach based on regionalized statistical similarity measures: correlation ratio and f – information (Vajda, 1989). A spatial prior was included by means of a reference image associated labeled scene model. The computation of similarity criteria involved regularization (Rougon et al., 2005b) and regional probability densities estimation by using specific Parzen kernels defined over each component of the scene. By iteratively adapting the image data to the scene model, joint non-rigid registration and segmentation of the target image were obtained. Instead of relying on a single reference image, Li and Sun (2009) generated a sequence of pseudo ground images and performed non-rigid registration between the corresponding images in the pseudo ground truth and original sequences. The pseudo ground truth estimation and the demons based non-rigid registration (Wang et al., 2005) were integrated into an energy functional that was iteratively minimized in an expectation-minimization fashion. The lack of comprehensive validation on large clinical datasets including different CMRP imaging pulse sequences and slice positions motivated Xue et al. (2008) to conduct a multi-center study, using a method including both the rigid and non-rigid registration. Multiple levels of resolution and optimization levels were used for rigid registration before the application of a non-rigid registration strategy, involving multi-level B-spline based FFD transformation. With NMI as the similarity metric, registration was performed consecutively between temporally adjacent images, starting from the template (peak intensity image). A different approach to the reference frame detection, called consecutive motion compensation, was adopted by Xue et al. (2009) in the extension of their previous work (Xue et al., 2008). Here, the reference frame was detected on the basis of maximum cross-correlation ratio between every frame in the perfusion series and the standard deviation (SD) image. The SD image consisted of the regions with highest standard deviation along the time dimension. Subsequently, the reference frame was first non-rigidly registered (Chef d’Hotel et al., 2002) with its immediate neighbors before multiple non-rigid registrations were performed between temporally adjacent slices. Furthermore, additional steps were included for: (a) surface coil inhomogeneity correction using proton density images, and (b) the generation of a robust parameter map.

4.3. Summary Generally, rigid registration provides more consistent results despite its inability to align the images accurately in case of deformations. Its consistency can be attributed to its ability to align the images using a precise nearest rigid transformation. Although the alignment with non-rigid registration is better, the precision is lower compared to rigid registration. Whereas local deformation values have to be calculated from a subset of image pixels in non-rigid registration, all the pixels in the ROI are employed in rigid registration to calculate the lower dimensional transformation.

As a result, rigid registration offers higher robustness to noise and better precision (less than a pixel) than non-rigid registration. Within rigid and non-rigid registration, feature based strategies achieve faster convergence (due to less iterations of optimization) and sub-pixel accuracies. Nevertheless, they require feature extraction and pre-processing before the actual registration. They are also more susceptible to noise and outliers than the intensity based methods. Compared to feature based methods, intensity based approaches are more flexible, as the information used for the alignment is not restricted to specific features. These methods are often generic, and hence, can be applied to any dataset without pre-processing. In addition, they are robust against noise and insensitive to outliers, as only the intensities are used for registration. However, as the intensity based methods use pixel intensities directly instead of features, they often suffer from difficulties in achieving sub-pixel accuracy and faster convergence. In addition to the accuracy and robustness, time efficiency of the registration process is a major factor determining its acceptance in the clinics. Several authors have therefore incorporated multi-scale and multi-resolution schemes into their registration approaches. These schemes increase the robustness of registration by initially emphasizing on global, large-scale image structures and then proceeding to further refinement based on small image details.

5. Methods based on segmentation The objective of segmentation is to provide reliable, fast, and effective object recognition and delineation either by clinicians or automatically using robust techniques. Although, segmentation algorithms are often applied on the images that have already been aligned spatially using a registration scheme, some methods use them to aid registration process as well. The segmentation strategies can be classified into model independent and model based approaches (Table 3).

5.1. Model independent Model independent segmentation approaches for CMRP image processing often rely on thresholding and clustering based techniques. Techniques based on thresholding assume a relationship between the objects of interest and distinctly quantifiable features such as image intensity and gradient. In clustering methods, the features of the interesting structures are extracted from the already classified points and are updated through iterations. Spreeuwers et al. (2001) proposed a thresholding based method to coarsely segment the LV and RV by searching for the local maxima in space and time. Region growing was applied with the seeds at the RV and LV centers, and their boundaries were extracted by tracing the outlines of the found regions using a polar transformation. Behloul et al. (1997) used clustering based segmentation to perform myocardial boundary detection. A slice from cine-MR images was segmented using fuzzy clustering and fuzzy inference systems (FIS). The segmented cine-MR images were then warped to obtain the desired myocardial boundaries in CMRP images. Here, the ROI was defined manually as a rectangular region on each image of the perfusion sequence. In another clustering based approach (Bella and Sitek, 2001), automatic grouping and fitting of temporally alike tissue time-signal voxels was performed on the basis of TIC analysis. Two types of TIC analysis techniques, i.e., factor analysis (Benali et al., 1993) and k-means clustering (Theiler and Gisler, 1997), were explored to extract the blood input function that enabled the automatic fitting of the data to a two-compartment model.

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

5.2. Model based Due to high variability in contrast around the epi- and endocardium, numerous segmentation algorithms use models, such as parametric and deformable models, that exploit prior knowledge on myocardial shape, texture, size, or location, to identify and delineate the myocardial boundaries. Several methods used parametric models. Boudraa et al. (2001) mapped the first-pass transit of contrast bolus through the heart to obtain a similarity map, or parametric image, representing regions with different temporal dynamics. The similarity map was generated by first eliminating the motion artifacts with manual translations, and then estimating the pixel by pixel correspondence to a myocardial reference using a temporal covariance measure. Although the similarity map resembled commonly used bull’s eye representation, it offered a better spatial resolution preservation. Janier et al. (2002) used FAMIS (Di Paola et al., 1982; Frouin et al., 1992; Benali et al., 1993) to obtain accurate myocardial tissue factor (FAMISt) images for the interpretation of perfusion in a selected group of patients with well-documented CAD. FAMIS is based on a statistical model which assumes that a time-varying image sequence can be studied by resolving it into a few spatial distributions, each with homogeneous signal intensity variations. These spatial distributions are estimated by factor images and the functions describing the signal variations are estimated by factors (Benali et al., 1993). Prior to the application of FAMIS, manual translation was applied when necessary on an image-by-image basis, and subsequently, a single FAMISt image was obtained with FAMIS for each raw CMRP image sequence. The efficacy of FAMIS was then evaluated with a multi-case, multi-reader receiver operating characteristic analysis. Deformable models have been used more often than the parametric models. Adluru et al. (2006b) employed level sets (Osher and Sethian, 1988) to utilize temporal and spatial information in the perfusion images for the automatic segmentation of the myocardium. Their previous approach (Lorenzo et al., 2006) was extended here by removing the manual selection of both the myocardial seed points and the shape model. Before the segmentation, CMRP images were rigidly registered to a reference image, which was chosen from the temporal center of the sequence. To perform this registration, spatially weighted MSDs were minimized between the reference frame and other images in the sequence. The segmentation was performed with a level set based framework that incorporated temporal information by using a spectral speed function (Cates et al., 2004). Considering the importance of myocardial contour identification and delineation in CMRP image analysis, the popularity of statistical contour based schemes is not surprising. Santarelli (2003) applied non-linear and anisotropic smoothing filter to evaluate image gradients, and then the myocardial contours were delineated through a deformable model based on gradient-vector-flow (GVF) snakes (Xu and Prince, 2002). However, this approach required manual tracing of a coarse polygon in the LV for myocardial segmentation. Precluding any manual interaction, model based unsupervised method proposed by Discher et al. (2005) involved statistical contour based automatic detection of a bi-ventricular patient-specific heart template on a reference image with maximum myocardial contrast. The template consisted of four regions: the thorax, the RV, the LV, and the left myocardium and the curve flow in this segmentation was performed with classical level sets. To ensure the inclusion of only cardiac region, a mathematical morphology (Soille, 1999) based spatio-temporal approach preceded the template detection step. Finally, an intensity based non-rigid registration technique (Rougon et al., 2005a) was used to eliminate the motion artifacts. In another contour based approach (Pluempitiwiriyawej and Sotthivirat, 2005), the problems

777

associated with the presence of papillary muscles around the LV were overcome using the stochastic active contours scheme (STACS) (Pluempitiwiriyawej et al., 2005). Information about common variations in the myocardial shape and texture can also be included in the model using statistical shape models, such as AAM. Baka et al. (2008) employed multibanded AAMs (Stegmann and Larsen, 2003) with the integration of spatial and temporal characteristics of the perfusion sequence into their segmentation method. Here, the multi-bands comprised different combinations of temporal (related to temporal perfusion profile) and spatial feature images (gradient orientation maps, Scott et al., 2003). Four temporal features (calculated pixel-wise) characterizing the TIC, and one spatial feature were examined. The maximum intensity projection feature and gradient orientation map were found to form the best combination, and the addition of more temporal features was deemed redundant. In another approach utilizing AAMs, Gupta et al. (2010) incorporated the prior knowledge on temporal variation in a CMRP sequence by applying ICA to extract physiologically relevant feature images and their corresponding weights, for each time point (Milles et al., 2008). In addition to the feature images, initialization parameters and the optimal contrast image were also obtained for the AAM based myocardial segmentation. Based on an ICA based model, a reference image for each time point was computed, with which the corresponding motion affected image was rigidly registered using CC as the similarity metric. The use of ICA and AAM allowed this approach to be completely independent of any manual interaction. 5.3. Summary As myocardial tissue often has distinctive intensity, choosing threshold-based techniques can be effective and efficient. However, these algorithms are highly sensitive to noise and interference of surrounding structures such as epicardial boundaries. As a result, these algorithms are seldom used alone; instead, they are often used either in combination with other strategies or in segmentation driven registration approaches (Bansal and FunkaLea, 2002). Clustering algorithms are computationally efficient, but the number of classes, the position of initial points and other used parameters should be properly determined beforehand in order to get good results. Model based approaches have been more popular due to their ability to efficiently handle contrast variations in the myocardium. Aided by the inherent a priori information on shape and structure of the object to be segmented, these approaches are more stable against local image artifacts and perturbations than model-independent algorithms. Among the model based approaches, majority of the methods adopted contour based deformable models, as they can be conformed specifically to myocardial shape and provide good results with appropriate modeling. In addition, they offer sub-pixel accuracy, noise insensitivity, and intuitive interaction mechanisms. However, many model based approaches require either a selection of proper parameters or a training step to achieve desired results. Another limitation of deformable models can be their poor convergence to concave boundaries. 6. Methods for multi-modality information fusion and visualization The myocardium can be studied with different imaging modalities, and each of these modalities provide complementary information about its structural and functional integrity. The first signs of myocardial deterioration are often provided by the analysis of myocardial perfusion, using SPECT, PET or MRI. With MRI,

778

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

Fig. 5. Image processing methods for intra- and inter-modality information fusion: the vertices of this triangle represent imaging modalities (or techniques), and an edge connecting any two vertices represents the fusion of the respective modalities.

one can look at CMRP or delayed enhancement (DE) images (Vöhringer et al., 2007; Jackson et al., 2007) to analyze the extent of deterioration, and to predict the myocardial viability. The later stages of myocardial ischemia can be analyzed using cine-MR imaging, which provides an indication of the contractile function of the ventricular wall (Manning and Pennell, 2002). As a result, several methods were proposed to concurrently analyze the information from different modalities either for visualization only, or for more detailed correlation studies. An overview of the combinations from different techniques and modalities is provided in Fig. 5 and Table 5. As discussed earlier, motion correction with rigid or non-rigid registrations is usually the major preprocessing step for the analysis of CMRP images. It is often followed by manual or automatic identification and delineation of the myocardial boundaries. From the region enclosed within these boundaries, the TIC, and subsequently, the perfusion parameters can be obtained. These parameters assist clinicians in detecting as well as locating a defect in the myocardium and assessing its severity, if a defect is present. A bull’s eye plot (BEP), based on the American Heart Association’s (AHA) 17-segment model (Cerqueira et al., 2002), is often used to display these parameters. It divides the left ventricle into 17 segments according to the correspondence between the myocardial segments and the supplying coronary branch: ramus circumflex (RCX), left anterior descending (LAD), and right coronary artery (RCA). The 17 segment AHA model allows the visualization of various perfusion parameters (SIpeak, Tpeak, Up-slope, Tmean, etc.) through a color-coded BEP as shown in Fig. 6. In one of the first applications of information fusion from different MRI techniques (Dromigny-Badin et al., 1998), cardiac contraction information obtained from cine-MR image sequences was fused with perfusion

information from CMRP images. Myocardial boundaries in cineMR images were detected semi-automatically using the method described in (Baldy et al., 1994). After the identification of systolic and diastolic images by comparing the region enclosed by endocardial contours, a contraction-map image (two B-spline curves) was obtained to represent the local contraction. Similarly, information from CMRP images was represented in the form of a myocardial perfusion map. The pixel values in this map were proportional to the TIC slope in the same pixel (Canet et al., 1995). Due to the complementary information and different representations of contraction and perfusion maps, the two maps could be fused to obtain a new image, in which the contraction and perfusion properties of the myocardium were revealed simultaneously. Oeltze et al. (2006) developed an integrated visualization technique to relate perfusion information from CMRP images to morphological information from CTCA images. From CTCA data, epicardial left ventricle was segmented using live-wire (Schenk et al., 2000) and the coronary tree was extracted using region growing (Hennemuth et al., 2005) and thinning algorithms (Selle et al., 2003). The visualization of the coronaries was implemented using convolution surfaces (Oeltze and Preim, 2005). Motion correction in the CMRP data was accomplished with a combination of rigid and non-rigid registration algorithms, using MI as the similarity metric (Rueckert et al., 1999). The ROI was defined manually and corresponding TICs were generated to calculate the parameters for each voxel of the dataset. A synchronized visualization allowed exploration of reduced perfusion with the threedimensional (3D) rendering of the coronary vessels. Multiple perfusion parameters were displayed simultaneously within a single image using the combination of shapes and colors. For instance, a 3D colored height-field display was used to show two different parameters on a single CMRP image; one parameter was represented by the color and the other by the elevation. Colored height fields, in addition to the link with coronary anatomy, further enhanced the voxel-wise analysis of different perfusion parameters. The integration of statistical methods and interactive feature specification by Oeltze et al. (2007) allowed the detection of trends in the data, thus facilitating the detection of suspicious structures and preprocessing failures. Furthermore, dimensionality reduction performed with correlation analysis and PCA (Müller et al., 2006) provided a better understanding of the inter-parameter relations. In 2008, Oeltze et al. (2008) proposed another visualization framework for the integrated analysis of CMRP, cine-MR and DEMR data. A glyph based method was employed, and myocardial viability and contractility parameters were displayed in a 3D space. Serving as the reference image, an end-diastolic image from the cine-MR data was used to register: (a) slice-wise DE-MR data using rigid transformations and normalized CC as the similarity metric; and (b) the baseline images (from different slices) in CMRP data

Table 5 Methods for multi-modality information fusion and visualization. Authors (year)

Modalities (in addition to CMRP)

Salient features

Data

Dromigny-Badin et al. (1998) Oeltze et al. (2006) Oeltze et al. (2007) Termeer et al. (2007) Oeltze et al. (2008) Hennemuth et al. (2008a) Termeer et al. (2008) Kirisßli et al. (2010b) Stolzmann et al. (2010)

cine-MR CTCA CTCA DE-MR, WH-MR cine-MR, DE-MR DE-MR, WH-MR CTCA, DE-MR CTCA CTCA, DE-MR

Overlay of BEPs Combined BEPs Combined BEPs, PCA VBEP Glyphs Region growing, Watershed, Live wire VBEP Linkage of 2D BEP and 3D model of the heart Combined BEPs

4 – 2 – 15 21 – 10 5

cine-MR: Real time cardiac MR imaging at each stage of the cardiac cycle over several heart beats. CTCA: Non-invasive imaging of the coronary arteries with CT scanner. DE-MR: Cardiac MR imaging after 10–20 min of contrast agent injection. WH-MR: MR scan of the whole heart in 3D exposing mainly anatomical detail. See Table 2 for the abbreviations.

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

(a)

(b)

Fig. 6. The perfusion of (a) left ventricular myocardium is typically shown using (b) a Bull’s Eye Plot (BEP) based on American Heart Association’s (AHA) 17-segment model.

using the approach proposed by (Hennemuth et al., 2007). For CMRP dataset, the myocardial segmentation was performed in one motion corrected image with Live Wire algorithm (Schenk et al., 2000) before propagating the obtained contours to all the frames of respective slices. In cine-MR images, the segmentation was limited to the end-diastolic and end-systolic phases. A histogram analysis based method (Hennemuth et al., 2008b) was used for the segmentation of scar in DE-MR images. Eventually, glyphs were used to facilitate joint visualization of myocardial perfusion, contractility and viability. Similar to Oeltze et al. (2008), Termeer et al. (2007) also used the data only from MRI but acquired using different techniques. A comprehensive framework for the visualization of coronary artery disease (CoViCAD) was proposed to combine information from 3D Whole Heart-MR (WH-MR) and DE-MR datasets; the former exposed anatomical details and the latter provided scar information. A novel addition to the BEP allowed the linkage of scar to specific coronary arteries. The myocardium was unfolded to extend the existing BEP representation while preserving the volumetric nature of the LV wall. The volumetric BEP (VBEP), thus generated, provided transmurality information in association with anatomical information. Additionally, the coronary arteries were also projected on the epicardial surface of the VBEP.

779

A more advanced mapping between the myocardium and the coronary arteries was established in the extension of CoViCAD framework by Termeer et al. (2008). A computational simulation was added to indirectly visualize the coronary anatomy, and it was demonstrated that this simulation could provide a more comprehensive visualization of the information in comparison to rendering the anatomical data directly. Coronary blood flow was also modeled by using CTCA images to obtain detailed coronary artery tree segmentation. Hennemuth et al. (2008a) proposed an approach for the fusion of myocardial viability and perfusion in an anatomical context. In addition to CMRP image alignment using an approach similar to Oeltze et al. (2008)’s, CMRP and DE-MR images were registered to WH-MR datset, with WH-MR volume as the reference, and NMI and normalized CC as the similarity metrics. For the coronary tree extraction, a region growing approach was used (Boskamp et al., 2004; Hennemuth et al., 2005). The DE-MR images were analyzed with methods based on histogram analysis, constrained watershed segmentation (Hennemuth et al., 2008b) and the Live Wire algorithm (Schenk et al., 2000). Stolzmann et al. (2010) argued that the side by side evaluation of anatomic and functional information from CTCA and CMRP images, respectively, often led to incorrect allocation of vascular territories to myocardial segments (Setser et al., 2005) due to the mismatch between the standard myocardial distribution territories (Cerqueira et al., 2002) and actual coronary anatomy (Setser et al., 2005; Schindler et al., 1999). The fusion of images, instead of side by side information visualization, was proposed as a solution. Three characteristic landmark points were defined in CTCA and DE-MR images for rigid registration based alignment. To extract the coronary tree from CTCA images, a region-growing algorithm (Bock et al., 2008) was used. The extraction of perfusion parameters from CMRP images involved B-spline registration for motion correction and a mixture model for segmentation of the hypoperfused myocardial regions. From the DE-MR images, the extent and location of scar tissue was determined by adopting a strategy similar to (Hennemuth et al., 2008a). An integrated visualization framework was developed to view registered surfaces and volume renderings of the coronary arteries and LV overlaid

Fig. 7. Common types of information obtained from different cardiac imaging modalities and techniques.

Fig. 8. An example of multi-parameter visualization: (a) presents a 3D model of the heart, with its coronary artery tree color-coded with the degree of stenosis. On the LV, the patient-specific perfusion territory of the LAD is shown in red, of the LCX in blue, of the Obtuse Marginal branch (MO) in yellow and of the RCA in green. (b) Shows how the perfusion information can be integrated into a 3D model of the heart. (c) Presents a 2D perfusion bulls eye plot (BEP) with the coronary artery tree projected on top; the more transparent the artery, the further it is from the epicardium. (d) Shows a 2D perfusion BEP with both its coronary artery tree and its coronary perfusion territories projected on top; arteries are color coded with the degree of stenosis. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

780

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

with parameters from CMRP (perfusion deficits) or DE-MR (scar tissue). In the integrated visualization and analysis approach of Kirisßli et al. (2010b), information on coronary artery tree was obtained from CTCA images using a semi-automatic method based on (Schaap et al., 2009; Metz et al., 2009), and myocardial perfusion parameters were obtained using QMass MR (MEDIS, 2005). To combine information extracted from CTCA and CMRP images, spatial alignment was performed with an iterative closest point approach. The salient features of the proposed visualization were as follows: (a) a 2D projection of the coronary tree on the BEP, (b) the addition of a 3D coronary tree model, and (c) a 3D model of the heart, derived automatically from CTCA images using a multi-atlas based segmentation approach (Kirisßli et al., 2010a). These components were linked to each other to allow an interactive visualization of the perfusion and coronary anatomy, simultaneously. 6.1. Summary The information fusion approaches in the existing literature focus on combining the information from: (a) different modalities, or (b) different techniques, but common modality (Fig. 7). An example of multi-parameter visualization is shown in Fig. 8. A common theme among all these methods is the use of BEP, either in 2D or 3D (Termeer et al., 2008). Several interactive visualization techniques incorporated both new and established registration and segmentation methods. These techniques employed CTCA images for extracting anatomical details of coronaries, CMRP images to obtain TICs and perfusion parameters, and DE-MR images to identify and locate the scar tissue. In addition, cine-MR images were also used by a few approaches to draw the initial contours for myocardial segmentation in CMRP images. For a detailed description of perfusion visualization approaches, articles by Behrens et al. (1996) and Preim et al. (2009) are recommended. 7. Conclusion We have presented a comprehensive survey of CMRP image processing methods in this article. A broad categorization of all such methods was proposed on the basis of two major image processing components: registration and segmentation. The methods for multi-modality information fusion and visualization were discussed within a distinct category. 7.1. Current approaches and future perspective CMRP image processing is a complex task due to surface coil related inhomogeneities, image artifacts (PVEs, motion artifacts, etc.), low contrast between the heart and other thorax structures, limited accurate anatomical landmarks, and lower image resolution. Surface coil inhomogeneity is often corrected by using baseline correction but proton density images have also been used (Xue et al., 2009). PVE correction in image domain has been tried by Kocher et al. (2007) but it is still limited by image resolution. Motion artifacts are most prevalent, and are generally eliminated by either rigid or non-rigid registration techniques. While rigid registration corrects for translations and rotations, elastic or non-rigid registration corrects for the deformation of the heart. Although, both rigid and non-rigid registration approaches have shown promising results with respect to accuracy, robustness and automation, there is no consensus regarding the optimal approach. Rigid registration is computationally more efficient, robust to noise, and provides better consistency along with higher precision. However, the rigid alignment does not take into account the through-plane motion and spatial deformation of the heart

caused by non-uniform cardiac motion and unintended respiratory motion. Non-rigid registration provides better alignment but with lower precision, and is more susceptible to noise. A promising future avenue in registration of CMRP images would be to consider the motion of the entire heart (e.g., by making a patient specific heart model) taking into account both its non-rigid deformation as well as rigid motion. Segmentation of the left ventricular myocardium along endoand epicardial borders is still a very challenging task; in particular, endocardial segmentation is difficult due to trabeculae and papillary muscles, and epicardial segmentation due to image intensity variation in the surrounding structures such as lungs, diaphragm, and epicardial fat. Robust endocardial boundary detection and PVE reduction in CMRP images may be achieved by a technique similar to the one proposed by Codella et al. (2008) and Codella et al. (2010) for cine-MR images. Here, small trabeculations and papillary muscles were taken into account by assuming that the pixel intensities within the LV endocardial boundary partially contribute to the blood volume and partially to the myocardial mass. For the segmentation of more challenging epicardial contours, deformable model based segmentation approaches, especially the contour based approaches, seem to be more promising. In general, future research in the segmentation of CMRP images is expected to focus more on the deformable model based approaches with further improvements in the complex topology handling capabilities, training and representation of the models. Multi-modality information fusion and visualization techniques allow visual assessment of the perfusion deficits from CMRP images as well as the display of complementary information from different modalities. At present, these techniques show: (a) different parameters in one image; (b) combined information from the images acquired with different imaging modalities; and (c) several features derived from the image analysis. Several visualization tools have also been used to assist interactive CMRP image analysis: color coding, geometrical shapes (e.g., glyphs), annotations, overlays, magnifiers, and statistical representations (e.g., plots, graphs). While the proposed approaches have been effective in analyzing information from a single patient at a time, true challenges in the future lie in visualizing the information from multiple patients, at multiple time points (e.g., to track the progression of pathology before and after the therapy), and multiple modalities. Furthermore, with the continuous growth in size and complexity of the image data, new techniques will be required to reduce either the dimensionality or the number of data points. This reduction, however, would also necessitate a spatial context to link different representations of the reduced data effectively and across multiple modalities. 7.2. Perfusion quantification Assuming that the motion artifacts are already eliminated by registration techniques and images are segmented manually or automatically using image processing methods, semi- or absolute quantitative perfusion analysis can be performed. The quantification of perfusion allows the measurement of absolute myocardial blood flow. However, the accuracy of this measurement is limited by the inability of perfusion quantification approaches to measure the arterial input accurately and relatively closer to the myocardial region of interest. Considering the rapid innovations in perfusion quantification field, the standardization is difficult and consensus on optimal approaches may therefore take more time. 7.3. Clinical acceptance of CMRP imaging processing methods Considerable success has been achieved by the reviewed methods with respect to the technical developments and extent of

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785 Table 6 Review and survey articles for further reading. Authors (year)

Topic

Mäkelä et al. (2002) Kellman and Arai (2007) Gerber et al. (2008) Scott et al. (2009) Attili et al. (2010) Beller (2010) Jerosch-Herold (2010) Preim et al. (2009) Beller (2010)

Cardiac image registration methods CMRP imaging sequences

Sourbron (2010)

CMRP imaging: history and state of the art Motion in cardiovascular imaging Quantification in cardiac MRI Multimodality cardiac imaging Quantification of myocardial perfusion with MRI Visual exploration and analysis of perfusion data Current and future trends in multi-modality cardiac imaging Technical aspects of MR perfusion

automation. Although several artifacts influencing the quantitative analysis in CMRP images have been overcome with the use of these methods, their acceptance in clinical routine is still limited. Two major drawbacks seem to be responsible for this limited acceptance. First, it involves a steep learning curve for the clinicians to use these methods without their integration into user friendly frameworks. Second, their acceptance is significantly hampered by the absence of algorithmic validation with multi-vendor and multi-center datasets. To address these issues, it is, therefore, essential to (a) integrate the robust post-processing techniques into user-friendly software packages (e.g., Heiberg et al., 2010; MEDIS, 2005), and (b) create a publicly available database of a few multi-vendor and multi-center datasets so that an unbiased evaluation of these methods could be performed in a systemic way similar to the MICCAI challenges (MICCAI Challenges, 2007–2010). As it stands now, it is difficult to comment on the clinical acceptance of surveyed methods due to the scarcity of published data on the integration of these methods in commercial softwares. In conclusion, the challenges posed by the rapid developments in CMRP imaging are often difficult to meet with the advancements in image acquisition systems and protocols alone. For the CMRP imaging to become a routinely used modality in the non-invasive assessment of CAD, it is therefore essential that it is augmented with image processing methods, which must continue to evolve, both with respect to functionality, and with respect to evaluation. Appendix A Several aspects of CMRP image acquisition, perfusion quantification, image artifacts and general image processing methods could not be discussed in the main body of this paper due to space limitations. To guide the reader to the relevant literature, a few of the latest review/survey articles have been enlisted in Table 6. References Ablitt, N., Gao, J., Gatehouse, P., Yang, G., 2002. Motion decoupling and registration for 3d magnetic resonance myocardial perfusion imaging. In: Proceedings of the International Conference on Computational Science – Part III. Springer-Verlag, pp. 285–294. Adluru, G., DiBella, E., Schabel, M., 2006a. Model-based registration for dynamic cardiac perfusion MRI. Journal of Magnetic Resonance Imaging 24, 1062–1070. Adluru, G., DiBella, E., Whitaker, R., 2006b. Automatic segmentation of cardiac short axis slices in perfusion. In: Proceedings of International Symposium on Biomedical Imaging (ISBI). IEEE, pp. 133–136. Al-Saadi, N., Nagel, E., Gross, M., Bornstedt, A., Schnackenburg, B., Klein, C., Klimek, W., Oswald, H., Fleck, E., 2000a. Noninvasive detection of myocardial ischemia from perfusion reserve based on cardiovascular magnetic resonance. Circulation 101, 1379. Al-Saadi, N., Nagel, E., Gross, M., Schnackenburg, B., Paetsch, I., Klein, C., Fleck, E., 2000b. Improvement of myocardial perfusion reserve early after coronary intervention: assessment with cardiac magnetic resonance imaging. Journal of the American College of Cardiology 36, 1557–1564.

781

Albert, M., Haung, W., Lee, J., Patlak, C., Springer Jr, C., 1993. Susceptibility changes following bolus injections. Magnetic Resonance in Medicine 29, 700–708. Arai, A., 2000. Magnetic resonance first-pass myocardial perfusion imaging. Topics in Magnetic Resonance Imaging 11, 383. Atkinson, D., Burstein, D., Edelman, R., 1990. First-pass cardiac perfusion: evaluation with ultrafast MR imaging. Radiology 174, 757. Atkinson, D., Hill, D., Stoyle, P., Summers, P., Keevil, S., 1997. Automatic correction of motion artifacts in magnetic resonance images using an entropy focus criterion. IEEE Transactions on Medical Imaging 16, 903–910. Attili, A., Schuster, A., Nagel, E., Reiber, J., van der Geest, R., 2010. Quantification in cardiac MRI: advances in image acquisition and processing. The International Journal of Cardiovascular Imaging 26, 27–40. Axel, L., Costantini, J., Listerud, J., 1987. Intensity correction in surface-coil MR imaging. American Journal of Roentgenology 148, 418–420. Bache, R., Schwartz, J., 1982. Effect of perfusion pressure distal to a coronary stenosis on transmural myocardial blood flow. Circulation 65, 928. Baka, N., Milles, J., Hendriks, E.A., Suinesiaputra, A., Jerosh-Herold, M., Reiber, J.H.C., Lelieveldt, B.P.F., 2008. Segmentation of myocardial perfusion MR sequences with multi-band active appearance models driven by spatial and temporal features. In: Proceedings of SPIE, pp. 691415-1–10. Baldy, C., Douek, P., Croisille, P., Magnin, I., Revel, D., Amiel, M., 1994. Automated myocardial edge detection from breath-hold cine-MR images: evaluation of left ventricular volumes and mass. Magnetic Resonance Imaging 12, 589–598. Bansal, R., Funka-Lea, G., 2002. Integrated image registration for cardiac MR perfusion data. In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). SpringerVerlag, pp. 659–666. Bardonova, J., Provaznik, I., Novakova, M., 2000. Hidden markov model in wavelet analysis of myocardial ischemia in rabbit. In: Proceedings of Computers in Cardiology. IEEE, pp. 419–421. Bauer, W., 1998. Myocardial perfusion and capillary blood volume during left ventricular remodelling. Magnetic Resonance Materials in Biology, Physics, and Medicine 6, 148–149. Behloul, F., Roux, J., Unterreiner, R., 1997. Automatic boundary detection of the left ventricle and measurement of local myocardial perfusion in MRI. In: Proceedings of Computers in Cardiology. IEEE, pp. 145–148. Behrens, U., Teubnerb, J., Evertsz, C., Walzb, M., Jiirgens, H., Peitgena, H., 1996. Computer assisted dynamic evaluation of contrast-enhanced breast-MRI. In: Proceedings of the International Symposium on Computer and Communication Systems for Image Guided Diagnosis and Therapy, Paris, June 1996, pp. 362– 367. Belaroussi, B., Milles, J., Carme, S., Zhu, Y., Benoit-Cattin, H., 2006. Intensity nonuniformity correction in MRI: existing methods and their validation. Medical Image Analysis 10, 234–246. Beller, G., 2010. Recent advances and future trends in multimodality cardiac imaging. Heart, Lung and Circulation 19, 193–209. Bellin, M.F., Molen, A.J.V.D., 2008. Extracellular gadolinium-based contrast media: an overview. European Journal of Radiology 66, 160–167. Benali, H., Buvat, I., Frouin, F., Bazin, J., Paola, R., 1993. A statistical model for the determination of the optimal metric in factor analysis of medical image sequences (FAMIS). Physics in Medicine and Biology 38, 1065. Bidaut, L.M., Vallée, J.P., 2001. Automated registration of dynamic MR images for the quantification of myocardial perfusion. Journal of Magnetic Resonance Imaging 13, 648–655. Bock, S., Kühnel, C., Boskamp, T., Peitgen, H., 2008. Robust vessel segmentation. In: Proceedings of SPIE, p. 691539. Boskamp, T., Rinck, D., Link, F., Kümmerlen, B., Stamm, G., Mildenberger, P., 2004. New vessel analysis tool for morphometric quantification and visualization of vessels in CT and MR imaging data sets. Radiographics 24, 287–297. Boudraa, A.O., Behloul, F., Janier, M., Canet, E., Champier, J., Roux, J.P., Revel, D., 2001. Temporal covariance analysis of first-pass contrast-enhanced myocardial magnetic resonance images. Computers in Biology and Medicine 31, 133–142. Bracoud, L., Vincent, F., Pachai, C., Canet, E., Croisille, P., Revel, D., Canet, E., Croisille, P., Revel, D., 2003. Automatic registration of MR first-pass myocardial perfusion images. In: Proceedings of Functional Imaging and Modeling of the Heart (FIMH). Springer-Verlag, pp. 215–223. Breeuwer, M., 2001. Towards automatic quantitative analysis of cardiac MR perfusion images. In: Proceedings of International Congress Series, pp. 967–973. Breeuwer, M., Spreeuwers, L., Quist, M., 2001. Automatic quantitative analysis of cardiac MR perfusion images. In: Proceedings of SPIE, pp. 733–742. Brown, J.J., Higgins, C.B., 1988. Myocardial paramagnetic contrast agents for MR imaging. American Journal of Roentgenology 151, 865–871. Burstein, D., Taratuta, E., Manning, W., 1991. Factors in myocardial perfusion imaging with ultrafast MRI and Gd-DTPA administration. Magnetic Resonance in Medicine 20, 299–305. Buvat, I., Benali, H., Paola, R., 1998. Statistical distribution of factors and factor images in factor analysis of medical image sequences. Physics in Medicine and Biology 43, 1695–1711. Canet, E., Revel, D., Sebbag, L., de Lorgeril, M., Baldy, C., Delabre, C., Amiel, M., 1995. Noninvasive assessment of no-reflow phenomenon in a canine model of reperfused infarction by contrast-enhanced magnetic resonance imaging. American Heart Journal 130, 949–956. Cates, J., Lefohn, A., Whitaker, R., 2004. GIST: an interactive, GPU-based level set segmentation tool for 3D medical images. Medical Image Analysis 8, 217–231. Cerqueira, M., Weissman, N., Dilsizian, V., Jacobs, A., Kaul, S., Laskey, W., Pennell, D., Rumberger, J., Ryan, T., Verani, M., 2002. Standardized myocardial segmentation

782

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

and nomenclature for tomographic imaging of the heart. Circulation 105, 539– 542. Chef d’Hotel, C., Hermosillo, G., Faugeras, O., 2002. Flows of diffeomorphisms for multimodal image registration. In: Proceedings of International Symposium on Biomedical Imaging (ISBI). IEEE, pp. 753–756. Cheng, A., Pegg, T., Karamitsos, T., Searle, N., Jerosch-Herold, M., Choudhury, R., Banning, A., Neubauer, S., Robson, M., Selvanayagam, J., 2007. Cardiovascular magnetic resonance perfusion imaging at 3-Tesla for the detection of coronary artery disease: a comparison with 1.5-Tesla. Journal of the American College of Cardiology 49, 2440–2449. Christian, T., Aletras, A., Arai, A., 2008. Estimation of absolute myocardial blood flow during first-pass MR perfusion imaging using a dual-bolus injection technique: comparison to single-bolus injection method. Journal of Magnetic Resonance Imaging 27, 1271–1277. Christian, T., Rettmann, D., Aletras, A., Liao, S., Taylor, J., Balaban, R., Arai, A., 2004. Absolute myocardial perfusion in canines measured by using dual-bolus firstpass MR imaging. Radiology 232, 677–684. Christiansen, J., Karamitsos, T., Myerson, S., Francis, J., Neubauer, S., 2010. Stress perfusion imaging using cardiovascular magnetic resonance: a review. Heart, Lung and Circulation 19, 697–705. Codella, N., Cham, M., Wong, R., Chu, C., Min, J., Prince, M., Wang, Y., Weinsaft, J., 2010. Rapid and accurate left ventricular chamber quantification using a novel CMR segmentation algorithm: a clinical validation study. Journal of Magnetic Resonance Imaging 31, 845–853. Codella, N., Weinsaft, J., Cham, M., Janik, M., Prince, M., Wang, Y., 2008. Left ventricle: automated segmentation by using myocardial effusion threshold reduction and intravoxel computation at MR imaging. Radiology 248, 1004– 1012. Comon, P., 1994. Independent component analysis: a new concept? Signal Processing 36, 287–314. Comte, A., Lalande, A., Aho, S., Walker, P., Brunotte, F., 2004. Realignment of myocardial first-pass MR perfusion images using an automatic detection of the heart-lung interface. Magnetic Resonance Imaging 22, 1001–1009. Cootes, T., Edwards, G., Taylor, C., 1998. Active appearance models. Proceedings of the 5th European Conference on Computer Vision, vol. II. Springer-Verlag, pp. 484–498. Cullen, J.H., Horsfield, M.a., Reek, C.R., Cherryman, G.R., Barnett, D.B., Samani, N.J., 1999. A myocardial perfusion reserve index in humans using first-pass contrast-enhanced magnetic resonance imaging. Journal of the American College of Cardiology 33, 1386–1394. Davies, R., Twining, C., Cootes, T., Waterton, J., Taylor, C., 2002. A minimum description length approach to statistical shape modeling. IEEE Transactions on Medical Imaging 21, 525–537. Delzescaux, T., Frouin, F., de Cesare, A., Philipp-Foliguet, S., Todd-Pokropek, A., Herment, A., Janier, M., 2003. Using an adaptive semiautomated self-evaluated registration technique to analyze MRI data for myocardial perfusion assessment. Journal of Magnetic Resonance Imaging 18, 681–690. Delzescaux, T., Frouin, F., De Cesare, A., Philipp-Foliguet, S., Zeboudj, R., Janier, M., Todd-Pokropek, A., Herment, A., 2001. Adaptive and self-evaluating registration method for myocardial perfusion assessment. Magnetic Resonance Materials in Biology, Physics, and Medicine 13, 28–39. Di Bella, E., Sitek, A., 2001. Time curve analysis techniques for dynamic contrast MRI studies. In: Information Processing in Medical Imaging. Springer, pp. 211–217. Di Bella, E.V.R., Parker, D.L., Sinusas, A.J., 2005. On the dark rim artifact in dynamic contrast-enhanced MRI myocardial perfusion studies. Magnetic Resonance in Medicine 54, 1295–1299. Di Carli, M., Dorbala, S., Meserve, J., El Fakhri, G., Sitek, A., Moore, S., 2007. Clinical myocardial perfusion PET/CT. Journal of Nuclear Medicine 48, 783. Di*Paola, R., Bazin, J., Aubry, F., Aurengo, A., Cavailloles, F., Herry, J., Kahn, E., 1982. Handling of dynamic sequences in nuclear medicine. IEEE Transactions on Nuclear Science 29, 1310–1321. Discher, A., Rougon, N., Preteux, F., 2005. An unsupervised approach for measuring myocardial perfusion in MR image sequences. In: Proceedings of SPIE, pp. 126– 137. Dornier, C., Ivancevic, M., Thevenaz, P., Vallée, J., 2003. Improvement in the quantification of myocardial perfusion using an automatic spline-based registration algorithm. Journal of Magnetic Resonance Imaging 18, 160–168. Dromigny-Badin, A., Zhu, Y., Magnin, I., Revel, D., 1998. Fusion of cine magnetic resonance and contrast-enhanced first-pass magnetic resonance data in patients with coronary artery disease: a feasibility study. Investigative Radiology 33, 12. Eichenberger, A., Schuiki, E., Köhli, V., Amann, F., McKinnon, G., von Schulthess, G., 1994. Ischemic heart disease: assessment with gadolinium-enhanced ultrafast MR imaging and dipyridamole stress. Journal of Magnetic Resonance Imaging 4, 425–431. Fidler, F., Wacker, C., Dueren, C., Weigel, M., Jakob, P., Bauer, W., Haase, A., 2004. Myocardial perfusion measurements by spin-labeling under different vasodynamic states. Journal of Cardiovascular Magnetic Resonance 6, 509–516. Fieno, D., Shea, S., Li, Y., Harris, K., Finn, J., Li, D., 2004. Myocardial perfusion imaging based on the blood oxygen level-dependent effect using T2-prepared steadystate free-precession magnetic resonance imaging. Circulation 110, 1284. Fleet, D., Jepson, A., 1990. Computation of component image velocity from local phase information. International Journal of Computer Vision 5, 77–104. Frank, H., Globits, S., 1999. Magnetic resonance imaging evaluation of myocardial and pericardial disease. Journal of Magnetic Resonance Imaging 10, 617–626.

Frouin, F., Bazin, J., Di Paola, M., Jolivet, O., Di Paola, R., 1992. FAMIS: a software package for functional feature extraction from biomedical multidimensional images. Computerized Medical Imaging and Graphics 16, 81–91. Gallippi, C., Kramer, C., Hu, Y., Vido, D., Reichek, N., Rogers, W., 2002. Fully automated registration and warping of contrast-enhanced first-pass perfusion images. Journal of Cardiovascular Magnetic Resonance 4, 459–469. Gao, J., Ablitt, N., Elkington, A., Yang, G.Z., 2002. In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). Springer-Verlag, pp. 612–619. Geman, S., Geman, D., 1993. Stochastic relaxation, Gibbs distributions and the Bayesian restoration of images. Journal of Applied Statistics 20, 25–62. Gerber, B., Raman, S., Nayak, K., Epstein, F., Ferreira, P., Axel, L., Kraitchman, D., 2008. Myocardial first-pass perfusion cardiovascular magnetic resonance: history, theory, and current state of the art. Journal of Cardiovascular Magnetic Resonance 10, 18. Go, R., Marwick, T., MacIntyre, W., Saha, G., Neumann, D., Underwood, D., Simpfendorfer, C., 1990. A prospective comparison of rubidium-82 PET and thallium-201 SPECT myocardial perfusion imaging utilizing a single dipyridamole stress in the diagnosis of coronary artery disease. Journal of Nuclear Medicine 31, 1899. Gupta, S., Solaiyappan, M., Beache, G., Arai, A., Foo, T., 2003. Fast method for correcting image misregistration due to organ motion in time-series MRI data. Magnetic Resonance in Medicine 49, 506–514. Gupta, V., Hendriks, E., Milles, J., Van Der Geest, R., Jerosch-Herold, M., Reiber, J., Lelieveldt, B., 2010. Fully automatic registration and segmentation of first-pass myocardial perfusion MR image sequences. Academic Radiology 17, 1375– 1385. Hachamovitch, R., Berman, D., Shaw, L., Kiat, H., Cohen, I., Cabico, J., Friedman, J., Diamond, G., 1998. Incremental prognostic value of myocardial perfusion single photon emission computed tomography for the prediction of cardiac death: differential stratification for risk of cardiac death and myocardial infarction. Circulation 97, 535–543. Heiberg, E., Sjögren, J., Ugander, M., Carlsson, M., Engblom, H., Arheden, H., 2010. Design and validation of segment-freely available software for cardiovascular image analysis. BMC Medical Imaging 10, 1. Hennemuth, A., Behrens, S., Kuehnel, C., Oeltze, S., Konrad, O., Peitgen, H., 2007. Novel methods for parameter-based analysis of myocardial tissue in MR images. In: Proceedings of SPIE, p. 54. Hennemuth, A., Boskamp, T., Fritz, D., Kühnel, C., Bock, S., Rinck, D., Scheuering, M., Peitgen, H., 2005. One-click coronary tree segmentation in CT angiographic images. In: Proceedings of International Congress Series. Elsevier, pp. 317–321. Hennemuth, A., Seeger, A., Friman, O., Miller, S., Klumpp, B., Oeltze, S., Peitgen, H., 2008a. A comprehensive approach to the analysis of contrast enhanced cardiac MR images. IEEE Transactions on Medical Imaging 27, 1592–1610. Hennemuth, A., Seeger, A., Friman, O., Miller, S., Peitgen, H., 2008b. Automatic detection and quantification of non-viable myocardium in late enhancement images. In: Proceedings of International Society for Magnetic Resonance in Medicine. Hesse, B., Tägil, K., Cuocolo, A., Anagnostopoulos, C., Bardies, M., Bax, J., Bengel, F., Busemann Sokole, E., Davies, G., Dondi, M., et al., 2005. EANM/ESC procedural guidelines for myocardial perfusion imaging in nuclear cardiology. European Journal of Nuclear Medicine and Molecular Imaging 32, 855–897. Hiller, K., Bock, M., Wacker, C., Schad, L., Waller, C., Haase, A., Van Kaick, G., Ertl, G., Bauer, W., 1998. MR-perfusion measurements: basic methodology and current status. Magnetic Resonance Materials in Physics, Biology and Medicine 6, 98– 99. Holland, A., Goldfarb, J., Edelman, R., 1998. Diaphragmatic and cardiac motion during suspended breathing: preliminary experience and implications for breath-hold MR imaging. Radiology 209, 483. Hyvärinen, A., Oja, E., 2000. Independent component analysis: algorithms and applications. Neural Networks 13, 411–430. Jackson, E., Bellenger, N., Seddon, M., Harden, S., Peebles, C., 2007. Ischaemic and non-ischaemic cardiomyopathies–cardiac MRI appearances with delayed enhancement. Clinical Radiology 62, 395–403. Janier, M., Mazzadi, A., Lionnet, M., Frouin, F., Andre-Fouet, X., Cinotti, L., Revel, D., Croisille, P., 2002. Factor analysis of medical image sequences improves evaluation of first-pass mr imaging acquisitions for myocardial perfusion. Academic Radiology 9, 26–39. Jerosch-Herold, M., 2010. Quantification of myocardial perfusion by cardiovascular magnetic resonance. Journal of Cardiovascular Magnetic Resonance 12, 1–16. Jerosch-Herold, M., Seethamraju, R., Swingen, C., Wilke, N., Stillman, A., 2004. Analysis of myocardial perfusion MRI. Journal of Magnetic Resonance Imaging 19, 758–770. Jerosch-Herold, M., Swingen, C., Seethamraju, R., 2002. Myocardial blood flow quantification with MRI by model-independent deconvolution. Medical Physics 29, 886. Kaandorp, T.A., Lamb, H.J., Bax, J.J., van der Wall, E.E., de Roos, A., 2005. Magnetic resonance imaging of coronary arteries, the ischemic cascade, and myocardial infarction. American Heart Journal 149, 200–208. Kawakami, K., Murase, K., Kumashiro, M., Kawakami, A., Tanki, N., Yamazaki, Y., Tabuchi, T., Nagayama, M., Watanabe, Y., 2004. Automatic motion correction for quantification of myocardial perfusion with dynamic magnetic resonance imaging. Magnetic Resonance in Medical Sciences 3, 105–117. Keijer, J., van Rossum, A., van Eenige, M., Karreman, A., Hofman, M., Valk, J., Visser, C., 1995. Semiquantitation of regional myocardial blood flow in normal human

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785 subjects by first-pass magnetic resonance imaging. American Heart Journal 130, 893–901. Kellman, P., Arai, A., 2007. Imaging sequences for first pass perfusion – a review. Journal of Cardiovascular Magnetic Resonance 9, 525–537. Kirisßli, H., Schaap, M., Klein, S., Papadopoulou, S., Bonardi, M., Chen, C., Weustink, A., Mollet, N., Vonken, E.P.A., van der Geest, R., van Walsum, T., Niessen, W., 2010a. Evaluation of a multi-atlas based method for segmentation of cardiac CTA data: a large-scale, multi-center and multi-vendor study. Medical Physics 37, 6279– 6292. Kirisßli, H.A., Gupta, V., Kirschbaum, S., Neefjes, L., van Geuns, R.J., Mollet, N.R.A., Lelieveldt, B.P.F., Reiber, J.H.C., van Walsum, T., Niessen, W., 2010b. Integrated visualization and analysis of coronary arteries and myocardial perfusion. International Journal of Computer Assisted Radiology and Surgery (CARS) 5 (1), S323. Kivelitz, D., Bis, K., Wilke, N., Jerosch-Herold, M., Juni, J., Hamm, B., 1997. Quantitative MR first pass perfusion imaging demonstrates coronary artery patency post interventions. Radiology 205, 254. Klein, M., Collier, B., Hellman, R., Bamrah, V., 1993. Detection of chronic coronary artery disease: value of pharmacologically stressed, dynamically enhanced turbo-fast low-angle shot MR images. American Journal of Roentgenology 161, 257. Klocke, F., Li, D., 2003. Testing coronary flow reserve without a provocative stress: a bold’’ approach. Journal of the American College of Cardiology 41, 841. Knutsson, H., Andersson, M., 2005. Morphons: segmentation using elastic canvas and paint on priors. In: Proceedings of International Conference on Image Processing. IEEE. Kocher, M., Daire, J., Thevenaz, P., Schindler, T., Keller, P., Didier, D., Vallee, J., 2007. Myocardial perfusion assessment by use of system identification method in a one-compartment model. In: Proceedings of 29th Annual International Conference of the IEEE – Engineering in Medicine and Biology Society. IEEE, pp. 4492–4495. Köstler, H., Ritter, C., Reiss-Zimmermann, M., Beer, M., Hahn, D., Sandstede, J., 2004. Correction for partial volume errors in MR heart perfusion imaging. Magnetic Resonance in Medicine 51, 848–852. Kozàkovà, M., Palombo, C., Pratali, L., Bigalli, G., Marzilli, M., Distante, A., L’Abbate, A., 1997. Assessment of coronary reserve by transoesophageal doppler echocardiography. European Heart Journal 18, 514. Larsson, H., Fritz-Hansen, T., Rostrup, E., et al., 1996. Myocardial perfusion modeling using MRI. Magnetic Resonance in Medicine 35, 716–726. Lee, D.C., Johnson, N.P., 2009. Quantification of absolute myocardial blood flow by magnetic resonance perfusion imaging. JACC: Cardiovascular Imaging 2, 761. Li, C., Sun, Y., 2009. Nonrigid registration of myocardial perfusion MRI using pseudo ground truth. In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pp. 165–72. S., Li, 1994. Markov random field models in computer vision. In: European Conference on Computer Vision. Springer, pp. 361–370. Lorenzo, L., MacLeod, R.S., Whitaker, R.T., Adluru, G., DiBella, E.V.R., 2006. Level sets and shape models for segmentation of cardiac perfusion MRI. In: Proceedings of SPIE, pp. 1320–1331. Mäkelä, T., Clarysse, P., Sipilä, O., Pauna, N., Pham, Q.C., Katila, T., Magnin, I.E., 2002. A review of cardiac image registration methods. IEEE Transactions on Medical Imaging 21, 1011–1021. Manning, W., Atkinson, D., Grossman, W., Paulin, S., Edelman, R., 1991. First-pass nuclear magnetic resonance imaging studies using gadolinium-DTPA in patients with coronary artery disease. Journal of the American College of Cardiology 18, 959–965. Manning, W., Pennell, D., 2002. Cardiovascular Magnetic Resonance, vol. 18. Churchill Livingstone, Philadelphia. Marcus, M., Wilson, R., White, C., 1987. Methods of measurement of myocardial blood flow in patients: a critical review. Circulation 76, 245–253. Martel, A., Moody, A., Allder, S., Delay, G., Morgan, P., 2001. Extracting parametric images from dynamic contrast-enhanced MRI studies of the brain using factor analysis. Medical Image Analysis 5, 29. McLeish, K., Hill, D., Atkinson, D., Blackall, J., Razavi, R., 2002. A study of the motion and deformation of the heart due to respiration. IEEE Transactions on Medical Imaging 21, 1142–1150. MEDIS, 2005. QMass MR. Mellor, M., Brady, M., 2004. Non-rigid multimodal image registration using local phase. In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). Springer, pp. 789– 796. Mellor, M., Brady, M., 2005. Phase mutual information as a similarity measure for registration. Medical Image Analysis 9, 330–343. Metz, C., Schaap, M., Weustink, A., Mollet, N., van Walsum, T., Niessen, W., 2009. Coronary centerline extraction from CT coronary angiography images using a minimum cost path approach. Medical Physics 36, 5568–5579. MICCAI Challenges, 2007–2010. Grand challenges in Medical Image Analysis. . Miller, D., Holmvang, G., Gill, J., Dragotakes, D., Kantor, H., Okada, R., Brady, T., 1989. MRI detection of myocardial perfusion changes by gadolinium-DTPA infusion during dipyridamole hyperemia. Magnetic Resonance in Medicine 10, 246–255. Milles, J., van der Geest, R.J., Jerosch-Herold, M., Reiber, J.H.C., Lelieveldt, B.P.F., 2008. Fully automated motion correction in first-pass myocardial perfusion MR image sequences. IEEE Transactions on Medical Imaging 27, 1611–1621.

783

Minoves, M., Garcia, A., Pavia, J., Herranz, R., 1993. Evaluation of myocardial perfusion defects by means of ’’bull’s eye’’ images. Clinical Cardiology 22, 16– 22. Mühling, O., Wang, Y., Panse, P., Jerosch-Herold, M., Cayton, M., Wann, L., Mirhoseini, M., Wilke, N., 2003. Transmyocardial laser revascularization preserves regional myocardial perfusion: an MRI first pass perfusion study. Cardiovascular Research 57, 63. Müller, W., Nocke, T., Schumann, H., 2006. Enhancing the visualization process with principal component analysis to support the exploration of trends. Proceedings of the 2006 Asia-Pacific Symposium on Information Visualisation. Australian Computer Society, Inc., pp. 121–130. Nagel, E., Klein, C., Paetsch, I., Hettwer, S., Schnackenburg, B., Wegscheider, K., Fleck, E., 2003. Magnetic resonance perfusion measurements for the noninvasive detection of coronary artery disease. Circulation 108, 432–437. Nagel, E., Underwood, R., Pennell, D., Sechtem, U., Neubauers, S., Wickline, S., Hess, O., Schwaiger, M., Fleck, E., 1998. New developments in non-invasive cardiac imaging: critical assessment of the clinical role of cardiac magnetic resonance imaging. European Heart Journal 19, 1286–1293. Nestares, O., Heeger, D., 2000. Robust multiresolution alignment of MRI brain volumes. Magnetic Resonance in Medicine 43, 705–715. Netsch, T., Van Muiswinkel, A., 2002. Automated registration and outlier identification for the improvement of cardiac MR perfusion quantification. In: 19th Annual Meeting of the ESMRMB, Cannes, France, p. 105. Oeltze, S., Doleisch, H., Hauser, H., Muigg, P., Preim, B., 2007. Interactive visual analysis of perfusion data. IEEE Transactions on Visualization and Computer Graphics 13, 1392–1399. Oeltze, S., Grothues, F., Hennemuth, A., Kuß, A., Preim, B., 2006. Integrated visualization of morphologic and perfusion data for the analysis of coronary artery disease. In: Proceedings of Eurographics/IEEE-VGTC Symposium on Visualization, pp. 131–138. Oeltze, S., Hennemuth, A., Glaßer, S., Kühnel, C., Preim, B., 2008. Glyph-based visualization of myocardial perfusion data and enhancement with contractility and viability information. In: Proceedings of Eurographics Workshop on Visual Computing for Biomedicine (2008), Citeseer. Oeltze, S., Preim, B., 2005. Visualization of vasculature with convolution surfaces: method, validation and evaluation. IEEE Transactions on Medical Imaging 24, 540–548. Ólafsdóttir, H., Stegmann, M.B., Ersbøll, B., Larsson, H.B.W., 2006. A comparison of FFD-based nonrigid registration and AAMs applied to myocardial perfusion MRI. In: Proceedings of SPIE, pp. 388–396. Osher, S., Sethian, J., 1988. Fronts propagating with curvature-dependent speed: algorithms based on hamilton-jacobi formulations. Journal of Computational Physics 79, 12–49. Paragios, N., 2003. A level set approach for shape-driven segmentation and tracking of the left ventricle. IEEE Transactions on Medical Imaging 22, 773–776. Penzkoffer, H., Wintersperger, B., Knez, A., Weber, J., Reiser, M., 1999. Assessment of myocardial perfusion using multisection first-pass MRI and color-coded parameter maps: a comparison to 99mTC Sesta MIBI SPECT and systolic myocardial wall. Magnetic Resonance Imaging 17, 161–170. Pijls, N., Uijen, G., Hoevelaken, A., Arts, T., Aengevaeren, W., Bos, H., Fast, J., Van Leeuwen, K., Van der Werf, T., 1990. Mean transit time for the assessment of myocardial perfusion by videodensitometry. Circulation 81, 1331–1340. Plein, S., Greenwood, J., Ridgway, J., 2010. Cardiovascular MR Manual. Springer. Pluempitiwiriyawej, C., Moura, J., Wu, Y., Ho, C., 2005. STACS: new active contour scheme for cardiac MR image segmentation. IEEE Transactions on Medical Imaging 24, 593–603. Pluempitiwiriyawej, C., Sotthivirat, S., 2005. Active contours with automatic initialization for myocardial perfusion analysis. In: Proceedings of 27th Annual International Conference of the IEEE – Engineering in Medicine and Biology Society. IEEE, pp. 3332–3335. Preim, B., Oeltze, S., Mlejnek, M., Gröller, E., Hennemuth, A., Behrens, S., 2009. Survey of the visual exploration and analysis of perfusion data. IEEE Transactions on Visualization and Computer Graphics 15, 205–220. Pruessmann, K., Weiger, M., Scheidegger, M., Boesiger, P., 1999. SENSE: sensitivity encoding for fast MRI. Magnetic Resonance in Medicine 42, 952–962. Ranhosky, A., Kempthorne-Rawson, J., 1990. The safety of intravenous dipyridamole thallium myocardial perfusion imaging: intravenous dipyridamole thallium imaging study group. Circulation 81, 1205–1209. Rougon, N., Discher, A., Preteux, F., 2005a. Region-driven statistical nonrigid registration: application to model-based segmentation and tracking of the heart in perfusion MRI. In: Proceedings of SPIE, pp. 148–159. Rougon, N., Petitjean, C., Prêteux, F., Cluzel, P., Grenier, P., 2005b. A non-rigid registration approach for quantifying myocardial contraction in tagged MRI using generalized information measures. Medical Image Analysis 9, 353–375. Rueckert, D., Sonoda, L., Hayes, C., Hill, D., Leach, M., Hawkes, D., 1999. Nonrigid registration using free-form deformations: application to breast MR images. IEEE Transactions on Medical Imaging 18, 712–721. Runge, V., Schoerner, W., Niendorf, H., Laniado, M., Koehler, D., Claussen, C., Felix, R., James Jr, A., 1985. Initial clinical evaluation of gadolinium DTPA for contrastenhanced magnetic resonance imaging. Magnetic Resonance Imaging 3, 27–35. Sabharwal, N., Lahiri, A., 2003. Role of myocardial perfusion imaging for risk stratification in suspected or known coronary artery disease. Heart 89, 1291. Sakuma, H., Suzawa, N., Ichikawa, Y., Makino, K., Hirano, T., Kitagawa, K., Takeda, K., 2005. Diagnostic accuracy of stress first-pass contrast-enhanced myocardial perfusion MRI compared with stress myocardial perfusion scintigraphy. American Journal of Roentgenology 185, 95.

784

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785

Sam, A., Morasch, M., Collins, J., Song, G., Chen, R., Pereles, F., 2003. Safety of gadolinium contrast angiography in patients with chronic renal insufficiency. Journal of Vascular Surgery 38, 313–318. Santarelli, M., 2003. Automated cardiac MR image segmentation: theory and measurement evaluation. Medical Engineering & Physics 25, 149–159. Schaap, Neefjes, L., Metz, C., van der Giessen, A., Weustink, A., Mollet, N., Wentzel, J., van Walsum, T., Niessen, W., 2009. Coronary lumen segmentation using graph cuts and robust kernel regression. In: Information Processing in Medical Imaging. Springer, pp. 528–539. Schaefer, S., Lange, R., Gutekunst, D., Parkey, R., Willerson, J., Peshock, R., 1991. Contrast-enhanced magnetic resonance imaging of hypoperfused myocardium. Investigative Radiology 26, 551. Schenk, A., Prause, G., Peitgen, H., 2000. Efficient semiautomatic segmentation of 3d objects in medical images. In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). Springer, pp. 71–131. Schindler, T., Magosaki, N., Jeserich, M., Oser, U., Krause, T., Fischer, R., Moser, E., Nitzsche, E., Zehender, M., Just, H., et al., 1999. Fusion imaging: combined visualization of 3D reconstructed coronary artery tree and 3D myocardial scintigraphic image in coronary artery disease. The International Journal of Cardiac Imaging 15, 357–368. Schöllhuber, A., 2008. Fully automatic segmentation of the myocardium in cardiac perfusion MRI. In: Proceedings of Central European Seminar on Computer Graphics. Schreiber, W., Schmitt, M., Kalden, P., Mohrs, O., Kreitner, K., Thelen, M., 2002. Dynamic contrast-enhanced myocardial perfusion imaging using saturationprepared TrueFISP. Journal of Magnetic Resonance Imaging 16, 641–652. Schultz, C., Alfidi, R., Nelson, A., Kopiwoda, S., Clampitt, M., 1984. The effect of motion on two-dimensional fourier transformation magnetic resonance images. Radiology 152, 117. Schwaiger, M., 1994. Myocardial perfusion imaging with PET. Journal of Nuclear Medicine 35, 693. Schwitter, J., Nanz, D., Kneifel, S., Bertschinger, K., Buchi, M., Knusel, P., Marincek, B., Luscher, T., Von Schulthess, G., 2001. Assessment of myocardial perfusion in coronary artery disease by magnetic resonance: a comparison with positron emission tomography and coronary angiography. Circulation 103, 2230. Scott, A.D., Keegan, J., Firmin, D.N., 2009. Motion in cardiovascular MR imaging. Radiology 250, 331. Scott, I., Cootes, T., Taylor, C., 2003. Improving appearance model matching using local image structure. In: Proceedings of Information Processing in Medical Imaging. Springer, pp. 258–269. Selle, D., Preim, B., Schenk, A., Peitgen, H., 2003. Analysis of vasculature for liver surgical planning. IEEE Transactions on Medical Imaging 21, 1344–1357. Setser, R., O’Donnell, T., Smedira, N., Sabik, J., Halliburton, S., Stillman, A., White, R., 2005. Coregistered MR imaging myocardial viability maps and multi-detector row CT coronary angiography displays for surgical revascularization planning: initial experience. Radiology 237, 465. Sodickson, D., Manning, W., 1997. Simultaneous acquisition of spatial harmonics (SMASH): fast imaging with radiofrequency coil arrays. Magnetic Resonance in Medicine 38, 591–603. Soille, P., 1999. Morphological Image Analysis: Principles and Applications. Springer, Berlin. Sourbron, S., 2010. Technical aspects of MR perfusion. European Journal of Radiology 76, 304–313. Spreeuwers, L., Breeuwer, M., Niessen, W., Viergever, M., 2001. Automatic detection of myocardial boundaries in MR cardio perfusion images. Lecture Notes in Computer Science 2208, 1228–1231. Stegmann, M., Larsen, R., 2003. Multi-band modelling of appearance. Image and Vision Computing 21, 61–68. Stegmann, M.B., Ólafsdöttir, H., Larsson, H.B.W., 2005. Unsupervised motioncompensation of multi-slice cardiac perfusion MRI. Medical Image Analysis 9, 394–410. Stolzmann, P., Alkadhi, H., Scheffel, H., Hennemuth, A., Kuehnel, C., Baumueller, S., Kozerke, S., Falk, V., Marincek, B., Donati, O.F., 2010. Image fusion of coronary CT angiography and cardiac perfusion MRI: a pilot study. European Radiology 20, 1174–1179. Storey, P., Chen, Q., Li, W., Edelman, R., Prasad, P., 2002. Band artifacts due to bulk motion. Magnetic Resonance in Medicine 48, 1028–1036. Sun, Y., Jolly, M., Moura, J., 2004. Contrast-invariant registration of cardiac and renal MR perfusion images. In: Proceedings of Medical Image Computing and Computer Assisted Intervention, pp. 903–910. Tautz, L., Hennemuth, A., Andersson, M., Seeger, A., Knutsson, H., 2010. Phase-based non-rigid registration of myocardial perfusion MR image sequences. In: Proceedings of International Symposium on Biomedical Imaging (ISBI), pp. 516–519. Termeer, M., Bescós, J., Breeuwer, M., Vilanova, A., Gerritsen, F., Gröller, M., 2007. CoViCAD: comprehensive visualization of coronary artery disease. IEEE Transactions on Visualization and Computer Graphics 13, 1632–1639. Termeer, M., Oliván Bescós, J., Breeuwer, M., Vilanova, A., Gerritsen, F., Gröller, M.E., Nagel, E., 2008. Visualization of myocardial perfusion derived from coronary anatomy. IEEE Transactions on Visualization and Computer Graphics, 1595– 1602. Theiler, J., Gisler, G., 1997. A contiguity-enhanced k-means clustering algorithm for unsupervised multispectral image egmentation. In: Proceedings of SPIE, pp. 108–118.

Thevenaz, P., Ruttimann, U., Unser, M., 2002. A pyramid approach to subpixel registration based on intensity. IEEE Transactions on Image Processing 7, 27–41. Thirion, J.P., 1998. Image matching as a diffusion process: an analogy with Maxwell’s demons. Medical Image Analysis 2, 243–260. Thomsen, H., 2004. Gadolinium-based contrast media may be nephrotoxic even at approved doses. European Radiology 14, 1654–1656. Tong, C., Prato, F., Wisenberg, G., Lee, T., Carroll, E., Sandler, D., Wills, J., 1993a. Techniques for the measurement of the local myocardial extraction efficiency for inert diffusible contrast agents such as gadopentate dimeglumine. Magnetic Resonance in Medicine 30, 332–336. Tong, C., Prato, F., Wisenberg, G., Lee, T., Carroll, E., Sandler, D., Wills, J., Drost, D., 1993b. Measurement of the extraction efficiency and distribution volume for Gd-DTPA in normal and diseased canine myocardium. Magnetic Resonance in Medicine 30, 337–346. Tsekos, N., Zhang, Y., Merkle, H., Wilke, N., Jerosch-Herold, M., Stillman, A., Ugurbil, K., 1995. Fast anatomical imaging of the heart and assessment of myocardial perfusion with arrhythmia insensitive magnetization preparation. Magnetic Resonance in Medicine 34, 530–536. Vajda, I., 1989. Theory of Statistical Evidence and Information. Kluwer Academic Publishers, Dordrecht. Vallée, J., Sostman, H., Macfall, J., Degrado, T., Zhang, J., Sebbag, L., Cobb, F., Wheeler, T., Hedlund, L., Turkington, T., et al., 1998. Quantification of myocardial perhsion by MRI after coronary occlusion. Magnetic Resonance in Medicine 40, 287–297. Veeser, S., Dunn, M., Yang, G., 2001. Multiresolution image registration for twodimensional gel electrophoresis. Proteomics 1, 856–870. Vöhringer, M., Mahrholdt, H., Yilmaz, A., Sechtem, U., 2007. Significance of late gadolinium enhancement in cardiovascular magnetic resonance imaging (CMR). Herz 32, 129–137. Wagner, A., Mahrholdt, H., Holly, T., Elliott, M., Regenfus, M., Parker, M., Klocke, F., Bonow, R., Kim, R., Judd, R., 2003. Contrast-enhanced MRI and routine single photon emission computed tomography (SPECT) perfusion imaging for detection of subendocardial myocardial infarcts: an imaging study. The Lancet 361, 374–379. Wang, H., Dong, L., O’Daniel, J., Mohan, R., Garden, A., Kian Ang, K., Kuban, D., Bonnen, M., Chang, J., Cheung, R., 2005. Validation of an accelerated ’demons’ algorithm for deformable image registration in radiation therapy. Physics in Medicine and Biology 50, 2887–2905. Wang, Y., Riederer, S., Ehman, R., 1995. Respiratory motion of the heart: kinematics and the implications for the spatial resolution in coronary imaging. Magnetic Resonance in Medicine 33, 713–719. Weinmann, H., Brasch, R., Press, W., Wesbey, G., 1984. Characteristics of gadolinium-DTPA complex: a potential NMR contrast agent. American Journal of Roentgenology 142, 619. Weissler, A., Harris, W., Schoenfeld, C., 1968. Systolic time intervals in heart failure in man. Circulation 37, 149–159. Wesbey, G., Higgins, C., McNamara, M., Engelstad, B., Lipton, M., Sievers, R., Ehman, R., Lovin, J., Brasch, R., 1984. Effect of gadolinium-DTPA on the magnetic relaxation times of normal and infarcted myocardium. Radiology 153, 165. Wilke, N., Jerosch-Herold, M., 1998. Assessing myocardial perfusion in coronary artery disease with magnetic resonance first-pass imaging. Cardiology Clinics 16, 227–246. Wilson, R., Wyche, K., Christensen, B., Zimmer, S., Laxson, D., 1990. Effects of adenosine on human coronary arterial circulation. Circulation 82, 1595– 1606. Wold, H., 1975. Soft modeling by latent variables: the nonlinear iterative partial least squares approach. Perspectives in Probability and Statistics, papers in honour of MS Bartlett, 520–540. Wolff, S., Balaban, R., 1989. Magnetization transfer contrast (MTC) and tissue water proton relaxation in vivo. Magnetic Resonance in Medicine 10, 135–144. Wolff, S., Eng, J., Balaban, R., 1991. Magnetization transfer contrast: method for improving contrast in gradient-recalled-echo images. Radiology 179, 133. Wollny, G., Ledesma-Carbayo, M.J., Kellman, P., Santos, A., 2010. Exploiting quasiperiodicity in motion correction of free-breathing myocardial perfusion MRI. IEEE Transactions on Medical Imaging 29, 1516–1527. Wong, K., Wu, E., Ng, M., Wu, Y., Tse, H., Lau, C., Lo, G., Yang, E., 2005. Image registration in myocardial perfusion MRI. In: Proceedings of IEEE – Engineering in Medicine and Biology Society, pp. 453–454. Xu, C., Prince, J., 2002. Snakes, shapes, and gradient vector flow. IEEE Transactions on Image Processing 7, 359–369. Xue, H., Guehring, J., Srinivasan, L., Zuehlsdorff, S., Saddi, K., Chefdhótel, C., Hajnal, J.V., Rueckert, D., 2008. Evaluation of rigid and non-rigid motion compensation of cardiac perfusion MRI. In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pp. 35–43. Xue, H., Zuehlsdorff, S., Kellman, P., Arai, A., Nielles-Vallespin, S., Chefdhóotel, C., Lorenz, C.H., Guehring, J., 2009. Unsupervised inline analysis of cardiac perfusion MRI. In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pp. 741– 749. Yang, G.Z., Burger, P., Panting, J., Gatehouse, P.D., Rueckert, D., Pennell, D.J., Firmin, D.N., 1998. Motion and deformation tracking for short-axis echo-planar myocardial perfusion imaging. Medical Image Analysis 2, 285–302.

V. Gupta et al. / Medical Image Analysis 16 (2012) 767–785 Zavaljevski, A., Holland, S., Dhawan, A., Dunn, R., Ball, W., 1999. Multilevel computed hemodynamic parameter maps from dynamic perfusion MRI. IEEE Transactions on Instrumentation and Measurement 48, 711–720. Zierler, K., 1962. Theoretical basis of indicator-dilution methods for measuring flow and volume. Circulation Research 10, 393.

785

Zierler, K., 1965. Equations for measuring blood flow by external monitoring of radioisotopes. Circulation Research 16, 309–321. Zierler, K., 2000. Indicator dilution methods for measuring blood flow, volume, and other properties of biological systems: a brief history and memoir. Annals of Biomedical Engineering 28, 836–848.