- Email: [email protected]
Integrating image processing in PACS
European Journal of Radiology 78 (2011) 210–224
Contents lists available at ScienceDirect
European Journal of Radiology journal homepage: www.elsevier.com/locate/ejrad
Integrating image processing in PACS Lorenzo Faggioni, Emanuele Neri ∗ , Francesca Cerri, Francesca Turini, Carlo Bartolozzi Diagnostic and Interventional Radiology, University of Pisa, Via Paradisa, 2, 56100 Pisa, Italy
a r t i c l e
i n f o
Article history: Received 21 June 2009 Accepted 22 June 2009 Keywords: Radiology information system Picture archiving and communication system Image processing
a b s t r a c t Integration of RIS and PACS services into a single solution has become a widespread reality in daily radiological practice, allowing substantial acceleration of workflow with greater ease of work compared with older generation film-based radiological activity. In particular, the fast and spectacular recent evolution of digital radiology (with special reference to cross-sectional imaging modalities, such as CT and MRI) has been paralleled by the development of integrated RIS–PACS systems with advanced image processing tools (either two- and/or three-dimensional) that were an exclusive task of costly dedicated workstations until a few years ago. This new scenario is likely to further improve productivity in the radiology department with reduction of the time needed for image interpretation and reporting, as well as to cut costs for the purchase of dedicated standalone image processing workstations. In this paper, a general description of typical integrated RIS–PACS architecture with image processing capabilities will be provided, and the main available image processing tools will be illustrated. © 2009 Elsevier Ireland Ltd. All rights reserved.
1. Introduction The increasing availability of integrated RIS–PACS solutions for image reading and reporting of medical imaging examinations has fuelled the development of integrated tools for digital image processing on RIS–PACS systems. This evolution represents a further step towards total integration of all instruments needed for interpretation and reporting of diagnostic imaging examinations in a health service environment, and technical problems have been faced, especially in early times, due to the relatively stringent hardware and software requirements for accomplishing such task. In the following sections, a general description of integrated RIS–PACS systems with image processing capabilities will be provided, and the main available image processing tools will be illustrated. Examples of the various image processing options were obtained by means of the open-source software OsiriX (version 3.5.1, http://www.osirix-viewer.com) integrated with our Institution PACS network (Fig. 1), and are illustrated in the following dedicated sections. 1.1. From dedicated workstations to server-based RIS–PACS solutions As known, a strong impulse toward evolution and diffusion of image processing tools has been represented, on one side, by the transition from analogue to digital image acquisition. In partic-
∗ Corresponding author. Tel.: +39 050997313, fax: +39 050997313. E-mail address: [email protected] (E. Neri). 0720-048X/$ – see front matter © 2009 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.ejrad.2009.06.022
ular, the evolution of digital imaging (for instance, conventional radiography) into a filmless direction has introduced the need for image visualisation software in order for images to be evaluated [1,2]. In addition, the possibility for the technician and the reporting radiologist to modify image settings in the post-processing phase as to emphasise peculiar image features (that might be underestimated or missed under standard display settings, thereby potentially compromising diagnosis), has made image conditioning functions on visualisation stations a must-have. This is the case for elementary tools for adjusting window levels (as previously available on consoles of computed tomography [CT] or magnetic resonance imaging [MRI] scanners) or applying filters, such as high frequency filters for magnification of bone details or low frequency filters for noise reduction. Such scenario has been decisive for the evolution of PACS systems and their integration with RIS stations. Another important point is that data transmitted to the integrated RIS–PACS station must contain the whole image information in order to allow accurate image processing, and are to be shared over a universal digital platform. This has lead to adoption of the DICOM format as a universal standard for medical image data exchange; however, the necessity to preserve data integrity of medical images with a high degree of accuracy makes DICOM files relatively large, thus prompting development of faster internet and intranet connections for reasonably quick data flow [3,4]. The problems related to the high amount of generated data have been boosted by the concurrent evolution of cross-sectional imaging techniques, such as CT and MRI. In particular, the advent of multidetector CT (MDCT) scanners with an increasing number of detector rows has urged the search for new technical solutions
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
211
Fig. 1. OsiriX integrated RIS–PACS interface: query list for retrieval from PACS (a) and image list for a given PACS-retrieved examination (b).
enabling efficient RIS–PACS integration, both in terms of image reading and reporting times [5,6]. In fact, modern MDCT scanners with 64 detector rows and beyond can produce thousands of images, each of them having submillimetric thickness in the longitudinal axis and a relatively high noise index in order to keep radiation dose down to acceptable levels. This would make analytic evaluation of every single image (as traditionally done with single slice CT examinations) extremely time consuming – incompatibly with the routine workflow of a diagnostic imaging centre – and error-prone due to the high image noise and increasing reader’s fatigue. In addition, the advancements of MDCT technology have gained to this latter application fields once pertaining to other imaging modalities: this is the case, for example, of CT colonography, CT angiography (with particular reference to CT angiography of peripheral vessels and CT coronography), and trauma evaluation. In this respect, if such transition to MDCT has, on one hand, brought several diagnostic advantages, on the other hand it has implied a
further multiplication of the amount of generated MDCT images. The key point is twofold: (1) many images need a large amount of space to be loaded, both in terms of hard disk and random access memory (RAM) space; (2) post-processing of such large datasets – both in 2D, and especially, in 3D mode – requires vast calculating power, both from CPU and graphics card. Putting things into perspective, for large MDCT datasets to be processed, a relatively enormous amount of power would be needed, far superior to that of common RIS–PACS systems, and comparable to that afforded by dedicated image processing workstations. Indeed, the former are usually cheap, medium-to-low power computers designed for word processing (the RIS part) and visualisation of digital images retrieved online from the local area network (the PACS part), while high-end image processing consoles are much more expensive and usually not designed for RIS integration.
212
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
Fig. 2. By clicking onto one of the buttons encircled in red (a), it is possible to instantly create orthogonal MPR views of a stack of cross-sectional images (b), in this example, MDCT images of a cyst of the mesenteric root (yellow asterisk) are reformatted on the three orthogonal planes. (c) Buttons are provided (red circle) that allow export of a specified set of MPR images in DICOM format as a separate image series.
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
A widely adopted solution to this problem, permitting to perform heavy-duty image processing on RIS–PACS terminals without upgrading hardware, has been represented by accessing via network post-processing tools that are physically installed on remote servers. More precisely, post-processing algorithms are run on powerful remote servers connected through the PACS to every RIS–PACS station, which therefore acts as a client computer. This allows to leave the whole image processing workload onto the server, which receives input commands from RIS–PACS computers, while in turn, the server sends the output in real time to RIS–PACS clients. The output can be saved either in DICOM or other image format (such as JPEG, TIFF, PNG, etc.) on RIS–PACS clients or to an external user device if allowed by the administrator settings of the client system, or even sent to the PACS in DICOM format for permanent online storage. 2. Image processing tools on integrated RIS–PACS systems: an overview 2.1. 2D post-processing tools 2.1.1. Multiplanar reformat (MPR) As known, the MPR technique allows to create a bidimensional image on an arbitrary plane from a dataset of complanar images. More precisely, the information about position and signal intensity of voxels belonging to a stack of images acquired on parallel planes with a given spatial orientation (for example, a series of axial images) is used to produce an output image where voxels are geometrically projected on a user-defined plane with a different inclination (e.g. a coronal, sagittal or oblique plane), as a function of position and signal intensity of voxels from source images [7,8]. In order to achieve good quality MPR reconstructions, it is essential that image spatial resolution is high enough to approach voxel isotropy [9], so that no substantial loss of information occurs when image data are displayed on planes other than the native one. In this sense, the widespread availability of MDCT scanners and the progressive diffusion of high-field MRI equipment
213
(especially for vascular and neuroimaging applications) has been very beneficial to the generation of excellent MPR reconstructions with submillimetric voxel size. The MPR algorithm has the advantage of preserving the whole contrast resolution of native images, thus allowing reliable measurement of tissue density and contrast enhancement; indeed, on most integrated RIS–PACS systems it is possible to take density measurements on regions of interest (ROI) from MPR images, as well as geometric measurements, such as distances, areas, or angles. Moreover, almost all current integrated RIS–PACS systems with image processing options allow instant MPR of CT and MRI datasets; in addition, the majority of them allow to automatically perform MPR on standard planes (axial, coronal, sagittal: orthogonal MPR) by generating stacks of reformatted images spanning from a given initial to a given final limit of the source dataset, that can be saved as independent image series in DICOM format or even as a movie (Fig. 2). It is also possible to view current and previous PACS-retrieved studies of the same patient in MPR mode with image synchronisation, allowing for more accurate cross-study image comparison. More generally, the user is free to define oblique reconstruction planes arbitrarily by orienting and moving an electronic caliper on the three standard orthogonal planes, on which structures (e.g. blood vessels) are identified based on their anatomical course. In this way it is possible, for example, to create images on planes perpendicular to the longitudinal axis of a vessel in order to measure its diameter or section area, thereby avoiding the measurement error due to the geometric distortion that occurs whenever such measures are taken on native images from vessels coursing non-perpendicularly to the acquisition plane. A variant of the MPR algorithm is CPR (Curved Planar Reformation), that allows to project native images from a succession of planes lying onto a user-defined polygonal trajectory [8]. CPR is particularly useful, for example, to display vessels with a tortuous course in the three spatial directions, such as coronary arteries. This operation can be carried out by selecting a series of points aligned along the centre of the lumen (centreline) on one or more MPR projections, thus obtaining a continuous rectified representation of the vessel course (Fig. 3). With CPR it may also be easier to perform
Fig. 3. CPR of renal arteries (lower image) can be performed by manually tracing the vessel centreline (green dotted line) on MIP-reformatted magnetic resonance angiography source images (upper image).
214
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
Fig. 4. Usage of slice averaging allows to reduce image noise on MPR images (a, source images; b, orthogonal MPR with slice averaging).
Fig. 5. MIP of the left renal artery (right image) can be performed through the MPR interface by selecting slab thickness and orientation (blue parallel lines) on MPR views (upper and lower left images).
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
215
Fig. 6. minIP can be effective as to enhance visualisation of bronchial fistula (a, source MDCT axial images; b, MPR views).
measurements of vessel size, as the selected vessel is displayed continuously without distortions of its lumen contour. Options are sometimes offered that allow automatic or semi-automatic detection of the centreline, as described more thoroughly in Section 3, thus simplifying and/or quickening the radiologist’s work. Alongside ‘pure’ MPR viewing, it is possible to modify visualisation settings of MPR views by applying slice averaging tools [8] (Fig. 4), which are helpful for reducing image noise (therefore enabling radiation dose reduction in MDCT data acquisition), as well as effective image number, thus accelerating the whole image reading process. Of course, the degree of averaging must be set at
a level that allows relatively expedite image evaluation without missing findings of diagnostic importance due to partial volume effect; as a rule of thumb, at our Institution an average reconstruction thickness of 2–3 mm is considered to be a good compromise, given a native slice thickness of 0.625 mm for source MDCT axial images. 2.1.2. Maximum Intensity Projection (MIP) and Minimum Intensity Projection (minIP) Besides slice averaging, MPR-processed data can be displayed by using the Maximum Intensity Projection (MIP) or the Minimum
216
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
Fig. 7. By tracing the vessel centreline on sagittal MIP image (left), it is possible to obtain MIP view of CPR of the superior mesenteric artery (right): see accurate depiction of the artery and of its branches without substantial overlap of vessels from other vascular territories.
Intensity Projection (minIP) visualisation algorithms. As known, with the MIP technique a slab with a defined thickness and a given spatial orientation is selected, and for each set of voxels aligned along the chosen direction, only that with the highest intensity is represented on the resulting reconstructed image (Fig. 5). Similarly, with the minIP algorithm only the voxel with the lowest intensity is shown on the final image among those included in the slab [8] (Fig. 6). As a consequence, the MIP and minIP techniques are of primary importance for displaying anatomical structures with homogenous intensity that are not comprised into their morphological continuity on a single plane, such as bones and contrast-enhanced blood vessels and urinary system with MIP, and air-containing structures (e.g. airways or the large bowel) with minIP. As for MPR with slice averaging, it is possible for the user to select slab width and orientation, in order to adjust the degree of voxel superimposition. Some software packages also allow to perform MIP or minIP onto curved MPR views, which can be beneficial e.g. to get a panoramical view of branch vessels originating from a main trunk with a tortuous course (Fig. 7). Moreover, as unwanted anatomical structures may overlap with thick-slab MIP/minIP, some integrated RIS/PACS solutions offer the option of pre-editing the original dataset as to remove undesired structures, either through intensity threshold- and/or region growing-based segmentation algorithms, or by manual cutting on native and/or MPR images [10,11]. Other applications allow direct editing on MIP/minIP views, while others offer editing options on 3D views and permit MIP/minIP on datasets previously modified in 3D mode (Fig. 8). 2.2. 3D post-processing tools The server-based approach for remote data handling has been especially beneficial as for the possibility to perform 3D image processing on RIS–PACS stations, such as Volume Rendering (VR), Shaded Surface Display (SSD), and Virtual Endoscopy (VE). This is particularly true for VR algorithms, that require a relatively large amount of computational power as their output must retain the whole information content of the input dataset, in contrast with less hardware-demanding 2D tools (for instance, MIP uses approx-
Fig. 8. Coronal thick-slab MIP view generated after removal of the aorta and renal arteries allows optimal depiction of the hepatic arterial vasculature and the superior mesenteric artery system. In this case of a patient candidate to liver transplantation, an accessory artery to the right hepatic lobe stems from the superior mesenteric artery (red arrow).
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
217
Fig. 9. By choosing different transfer curves for VR, the user can emphasise different features of the image dataset: in this case of Leriche syndrome, a logarithmic inverse curve with a relatively low intensity threshold may be used to display collateral vessels (a), while one with a higher threshold allows better depiction of the aorto-iliac axes and splanchnic arteries (b). A linear curve with a very low threshold would obscure vessels but enable visualisation of soft tissues, such as the skin (c). In addition, the possibility for the user to create customised transfer curves is included (d).
218
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
Fig. 9. (Continued ).
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
219
Fig. 10. In this case of sacral metastasis from thyroid tumour vascularised from branches of the right internal iliac artery (a: MIP semi-axial image), it is possible to use different CLUT as to obtain optimal VR depiction of the relationship between the lesion and the host bone, together with its feeding vessels (b and c).
220
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
Fig. 11. VR with manual sculpting of patient table (a) and abdominal VR MDCT angiogram obtained after automatic bone removal (b); in this latter case, vessels of the arcade of Riolan were manually cropped for better image clarity.
imately 10% of the original information, because only voxels with highest intensity are represented on the output view). This way, a complete set of 2D and 3D options integrated on a RIS–PACS station is available to the user, virtually eliminating the need for him to move onto high-end standalone workstations for image processing. 2.2.1. Volume Rendering (VR) As known, while MIP is based on the extraction of maximum intensity voxels inside a user-defined slab among those crossed by
an array of parallel lines, VR algorithms harness the entire spatial and intensity information of the dataset as to represent on the reconstructed image a weighted average of the intensity of all voxels included in the slab. Each voxel is assigned a value of opacity, transparency, and colour depending on its intensity, position, and the prospective direction in which the image volume is observed according to a predetermined transfer function, that allows to generate the final image by weighing the intensity of the various voxels encountered by each ray. VR plug-ins let the
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
221
Fig. 12. (a) SSD view of lower limb arteries in patient with obstruction of right superficial femoral artery and aneurysm of the left superficial femoral artery; the dialogue box for colour and threshold settings is visible on the left half of the screen. (b) MIP of lower limb arteries (left) and magnified VR of popliteal and infrapopliteal arteries (right) are shown as a reference.
user define his own transfer function by choosing a colour and a transparency value and setting at least two voxel intensity thresholds, corresponding to the minimum and maximum luminance of visualisable points on the reconstructed image. So, it is possible to define linear transfer functions (implying that the ratio between the intensity of two voxels on native images and those of the corresponding pixels on the VR image is constant), as well as nonlinear (e.g. logarithmic or logarithmic inverse) ones, or even more complex functions with multiple thresholds, as to enhance or attenuate the contribution to the VR image of voxels within a given intensity range [7,8,12,13] (Fig. 9). It is also possible to choose among different colour look-up tables (CLUT) in order to enhance depiction of complex anatomical structures; preset CLUT are available on most RIS–PACS software, and can be manually customised by the user for
his convenience (Fig. 10). Furthermore, VR plug-ins often provide either automatic or manual image editing tools: automatic editing allows to remove unwanted voxels (usually bone or patient table) from the final image either by automatic software recognition or by mouse-clicking on the structures to be sculpted away, while manual editing can be done by the user by mouse-tracing the parts of the VR image to be preserved or deleted (Fig. 11). This enables accurate selection of the anatomy to be represented, thereby reducing or eliminating superimposition of surrounding structures that might hamper the diagnostic quality of the reconstructed VR capture. It is also possible to on-the-fly change the rendering mode for edited data (for example, from VR to MIP) as to optimise depiction of certain image features with the most appropriate post-processing technique [13]. VR images can be exported either in DICOM or non-
222
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
DICOM exchange format, or dynamically assembled in a movie rotating around a user-defined axis for offline file export. To this latter purpose, several file formats (such as AVI, MOV, or MPEG) are available, along with the possibility to use standard codecs for video encoding. 2.2.2. Shaded surface display (SSD) Some integrated RIS–PACS solutions also include SSD capabilities. Briefly, with this algorithm the user fixes a threshold defining the minimum intensity of the voxels to be imaged, while voxels with lower intensity will be discarded, and the reconstructed image will be composed by surfaces corresponding to the interface between voxels with intensity above and below the threshold (i.e. void voxels), obtained by simulating the illumination of surfaces from an external light source in a given user-defined direction [8,11,14]. In the past, given the availability of relatively lowpowered workstations, SSD was usually advantageous over VR due to its relatively modest computational requirements, because representation of dataset voxels on the output image is binary, allowing only a fraction of them to be processed. However, the binary nature of the SSD algorithm is also its main weakness, due to the difficulty to find an appropriate threshold for generation of nondistorted 3D images, causing hole-like artefacts to appear on output views in case of bad tissue separation. For this reason, and because heavy data processing is no longer a problem (as it is performed on remote servers), SSD has been superseded by the more robust and powerful VR algorithms, and few applications still support it (Fig. 12). 2.2.3. Virtual endoscopy (VE) By applying SSD or VR algorithms in order to explore hollow structures from an intraluminal, rather than extraluminal light source, it is possible to achieve endoscopy-like views. SSD-based VE techniques require the definition of an intensity threshold and generate surfaces formed by the interfaces between lower and higher intensity voxels. VR-based VE techniques are more robust and require the definition of a central intensity value (window level) and an intensity range (window width), corresponding to the signal intensity spectrum of voxels to be displayed with different opacities on the VE view. Most common integrated RIS–PACS image processing tools include VE capabilities, allowing to define window levels, opacity, and colour settings, together with a set of customisable CLUT, resembling the user interface of VR tools. On
some software packages it is even possible to toggle between VR and VE modes from the same dialogue panel, reflecting the conceptual and mechanistic similarity between those two techniques. The position and orientation of the virtual endoscope, together with the magnification degree of the VE view, can be defined by means of an electronic cursor onto synchronised orthogonal or oblique MPR projections (Fig. 13). It is also possible to create a centreline (‘flythrough’) along which VE views are taken, either automatically by software-assisted image segmentation, or manually by user tracing. This is particularly useful for exploration of anatomical cavities of homogenous intensity, such as airways, the colon, or contrastenhanced vessels or cardiac cavities. In addition, some programs allow to make geometric measurements from VE views, that can be useful e.g. for the assessment of vascular ostia from inside blood vessels or cardiac cavities [15]. Finally, software implementations exist that enable export of VE images either as standalone captures or in conjunction with MPR views, thus indicating the exact position from which VE is performed. As with VR reconstructions, VE screenshots can be saved either in DICOM mode for PACS storage, or in non-DICOM image format for offline file export, or even as a movie of consecutive VE images. This latter option is especially useful in case of fly-through VE, having the purpose of showing the internal structure of a hollow viscus or cavity through its entire length. 3. Dedicated image analysis plug-ins With the growing diffusion of cross-sectional imaging and the increasing size of image datasets over time (with particular reference to MDCT datasets), the need for image processing tools with a higher degree of automation has occurred. This holds especially true for cardiovascular imaging, in which the enormous amount of obtainable data has justified the development of automated image processing tools assisting the radiologist into the interpretation of findings from thousands of images in a reliable and reasonably fast fashion, compatibly with the routine workflow of a health service. Dedicated plug-ins for automated and semi-automated image analysis are currently available on most RIS–PACS software packages, ranging from CT colonography applications to solutions for cardiovascular imaging. Such tools integrate standard 2D and 3D post-processing algorithms with filtering and segmentation functions dedicated to one particular organ or anatomical territory, as to enable anatomy-specific image analysis. For example, CT colonography tools perform segmentation of air-containing voxels from the
Fig. 13. VE image of colonic polyp with synchronised orthogonal MPR views for virtual endoscope positioning.
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
223
Fig. 14. Segmentation of the right coronary artery with CPR view (a) and automatic vessel analysis (b) by means of the open-source CMIV CTA Plug-in for OsiriX (CMIV, Linköping University, Sweden).
original dataset (e.g. through region-growing algorithms), allowing both panoramic or targeted 2D and 3D visualisation of the large bowel, together with more organ-tailored functions such as virtual dissection, virtual biopsy, or virtual navigation; this latter can usually be accomplished with optimised window and CLUT settings through fly-through centreline-oriented VE. Another example is represented by vessel and cardiac analysis plug-ins, that allow extraction of voxels from contrast-enhanced vessels (automatic vessel tracking) by prior user definition of vessel start and end points (Fig. 14); subsequently, the software can derive geometrical information (such as vessel lumen diameter and section area, luminal stenosis or dilation degree, or aneurysm volume), leading to potential reduction of reader’s error due to inter- and intra-rater variability of manually taken measurements, as well as to lower overall image interpretation time. Of course, vessel tracking – as well as segmentation algorithms in general – can be controlled and eventually modified by the user in case of unsuccessful or erroneous performance, for example due to artefacts on native images. 4. Conclusions Integration of image processing tools with RIS–PACS systems represents an evolutionary step toward centralisation of all instru-
ments for medical image reading and reporting onto one single machine. This has the potential, on one hand, to save money and space by avoiding purchase of dedicated high-end workstations, and, on the other hand, to reduce overall reading time, thus resulting into increased productivity. We believe that with the increasing availability of adequate RIS–PACS systems with image processing capabilities, usage of standalone workstations should be reserved to selected cases of particular complexity, in which the added value of highly advanced and costly dedicated software might be relevant to diagnosis.
References [1] Carrino JA. Digital imaging overview. Semin Roentgenol 2003;38(3):200–15. [2] Tellis WM, Andriole KP, Jovais CS, Avrin DE. RIS minus PACS equals film. J Digit Imaging 2002;15(Suppl. 1):20–6. [3] Mildenberger P, Eichelberg M, Martin E. Introduction to the DICOM standard. Eur Radiol 2002;12(4):920–7. [4] Khorasani R. Image compression in your PACS: should you do it? What are the issues? J Am Coll Radiol 2004;1(10):780–1. [5] Kotter E, Baumann T, Jäger D, Langer M. Technologies for image distribution in hospitals. Eur Radiol 2006;16(6):1270–9. [6] Lee KH, Lee HJ, Kim JH, et al. Managing the CT data explosion: initial experiences of archiving volumetric datasets in a mini-PACS. J Digit Imaging 2005;18(3):188–95.
224
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
[7] Lell MM, Anders K, Uder M, et al. New techniques in CT angiography. Radiographics 2006;26(Suppl. 1):S45–62. [8] Cody DD. AAPM/RSNA physics tutorial for residents: topics in CT. Image processing in CT. Radiographics 2002;22(5):1255–68. [9] Dalrymple NC, Prasad SR, El-Merhi FM, Chintapalli KN. Price of isotropy in multidetector CT. Radiographics 2007;27(1):49–62. [10] Glockner JF. Three-dimensional gadolinium-enhanced MR angiography: applications for abdominal imaging. Radiographics 2001;21(2): 357–70. [11] Prokop M, Shin HO, Schanz A, Schaefer-Prokop CM. Use of maximum intensity projections in CT angiography: a basic review. Radiographics 1997;17(2):433–51.
[12] Calhoun PS, Kuszyk BS, Heath DG, Carley JC, Fishman EK. Three-dimensional volume rendering of spiral CT data: theory and method. Radiographics 1999;19(3):745–64. [13] Fishman EK, Ney DR, Heath DG, Corl FM, Horton KM, Johnson PT. Volume rendering versus maximum intensity projection in CT angiography: what works best, when, and why. Radiographics 2006;26(3):905–22. [14] Davis CP, Hany TF, Wildermuth S, Schmidt M, Debatin JF. Postprocessing techniques for gadolinium-enhanced three-dimensional MR angiography. Radiographics 1997;17(5):1061–77. [15] Cirillo S, Tosetti I, Gaita F, Bianchi F, Gandini G, Regge D. Magnetic resonance angiography of the pulmonary veins before and after radiofrequency ablation for atrial fibrillation. Radiol Med 2005;109(5–6):488–99.
Contents lists available at ScienceDirect
European Journal of Radiology journal homepage: www.elsevier.com/locate/ejrad
Integrating image processing in PACS Lorenzo Faggioni, Emanuele Neri ∗ , Francesca Cerri, Francesca Turini, Carlo Bartolozzi Diagnostic and Interventional Radiology, University of Pisa, Via Paradisa, 2, 56100 Pisa, Italy
a r t i c l e
i n f o
Article history: Received 21 June 2009 Accepted 22 June 2009 Keywords: Radiology information system Picture archiving and communication system Image processing
a b s t r a c t Integration of RIS and PACS services into a single solution has become a widespread reality in daily radiological practice, allowing substantial acceleration of workflow with greater ease of work compared with older generation film-based radiological activity. In particular, the fast and spectacular recent evolution of digital radiology (with special reference to cross-sectional imaging modalities, such as CT and MRI) has been paralleled by the development of integrated RIS–PACS systems with advanced image processing tools (either two- and/or three-dimensional) that were an exclusive task of costly dedicated workstations until a few years ago. This new scenario is likely to further improve productivity in the radiology department with reduction of the time needed for image interpretation and reporting, as well as to cut costs for the purchase of dedicated standalone image processing workstations. In this paper, a general description of typical integrated RIS–PACS architecture with image processing capabilities will be provided, and the main available image processing tools will be illustrated. © 2009 Elsevier Ireland Ltd. All rights reserved.
1. Introduction The increasing availability of integrated RIS–PACS solutions for image reading and reporting of medical imaging examinations has fuelled the development of integrated tools for digital image processing on RIS–PACS systems. This evolution represents a further step towards total integration of all instruments needed for interpretation and reporting of diagnostic imaging examinations in a health service environment, and technical problems have been faced, especially in early times, due to the relatively stringent hardware and software requirements for accomplishing such task. In the following sections, a general description of integrated RIS–PACS systems with image processing capabilities will be provided, and the main available image processing tools will be illustrated. Examples of the various image processing options were obtained by means of the open-source software OsiriX (version 3.5.1, http://www.osirix-viewer.com) integrated with our Institution PACS network (Fig. 1), and are illustrated in the following dedicated sections. 1.1. From dedicated workstations to server-based RIS–PACS solutions As known, a strong impulse toward evolution and diffusion of image processing tools has been represented, on one side, by the transition from analogue to digital image acquisition. In partic-
∗ Corresponding author. Tel.: +39 050997313, fax: +39 050997313. E-mail address: [email protected] (E. Neri). 0720-048X/$ – see front matter © 2009 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.ejrad.2009.06.022
ular, the evolution of digital imaging (for instance, conventional radiography) into a filmless direction has introduced the need for image visualisation software in order for images to be evaluated [1,2]. In addition, the possibility for the technician and the reporting radiologist to modify image settings in the post-processing phase as to emphasise peculiar image features (that might be underestimated or missed under standard display settings, thereby potentially compromising diagnosis), has made image conditioning functions on visualisation stations a must-have. This is the case for elementary tools for adjusting window levels (as previously available on consoles of computed tomography [CT] or magnetic resonance imaging [MRI] scanners) or applying filters, such as high frequency filters for magnification of bone details or low frequency filters for noise reduction. Such scenario has been decisive for the evolution of PACS systems and their integration with RIS stations. Another important point is that data transmitted to the integrated RIS–PACS station must contain the whole image information in order to allow accurate image processing, and are to be shared over a universal digital platform. This has lead to adoption of the DICOM format as a universal standard for medical image data exchange; however, the necessity to preserve data integrity of medical images with a high degree of accuracy makes DICOM files relatively large, thus prompting development of faster internet and intranet connections for reasonably quick data flow [3,4]. The problems related to the high amount of generated data have been boosted by the concurrent evolution of cross-sectional imaging techniques, such as CT and MRI. In particular, the advent of multidetector CT (MDCT) scanners with an increasing number of detector rows has urged the search for new technical solutions
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
211
Fig. 1. OsiriX integrated RIS–PACS interface: query list for retrieval from PACS (a) and image list for a given PACS-retrieved examination (b).
enabling efficient RIS–PACS integration, both in terms of image reading and reporting times [5,6]. In fact, modern MDCT scanners with 64 detector rows and beyond can produce thousands of images, each of them having submillimetric thickness in the longitudinal axis and a relatively high noise index in order to keep radiation dose down to acceptable levels. This would make analytic evaluation of every single image (as traditionally done with single slice CT examinations) extremely time consuming – incompatibly with the routine workflow of a diagnostic imaging centre – and error-prone due to the high image noise and increasing reader’s fatigue. In addition, the advancements of MDCT technology have gained to this latter application fields once pertaining to other imaging modalities: this is the case, for example, of CT colonography, CT angiography (with particular reference to CT angiography of peripheral vessels and CT coronography), and trauma evaluation. In this respect, if such transition to MDCT has, on one hand, brought several diagnostic advantages, on the other hand it has implied a
further multiplication of the amount of generated MDCT images. The key point is twofold: (1) many images need a large amount of space to be loaded, both in terms of hard disk and random access memory (RAM) space; (2) post-processing of such large datasets – both in 2D, and especially, in 3D mode – requires vast calculating power, both from CPU and graphics card. Putting things into perspective, for large MDCT datasets to be processed, a relatively enormous amount of power would be needed, far superior to that of common RIS–PACS systems, and comparable to that afforded by dedicated image processing workstations. Indeed, the former are usually cheap, medium-to-low power computers designed for word processing (the RIS part) and visualisation of digital images retrieved online from the local area network (the PACS part), while high-end image processing consoles are much more expensive and usually not designed for RIS integration.
212
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
Fig. 2. By clicking onto one of the buttons encircled in red (a), it is possible to instantly create orthogonal MPR views of a stack of cross-sectional images (b), in this example, MDCT images of a cyst of the mesenteric root (yellow asterisk) are reformatted on the three orthogonal planes. (c) Buttons are provided (red circle) that allow export of a specified set of MPR images in DICOM format as a separate image series.
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
A widely adopted solution to this problem, permitting to perform heavy-duty image processing on RIS–PACS terminals without upgrading hardware, has been represented by accessing via network post-processing tools that are physically installed on remote servers. More precisely, post-processing algorithms are run on powerful remote servers connected through the PACS to every RIS–PACS station, which therefore acts as a client computer. This allows to leave the whole image processing workload onto the server, which receives input commands from RIS–PACS computers, while in turn, the server sends the output in real time to RIS–PACS clients. The output can be saved either in DICOM or other image format (such as JPEG, TIFF, PNG, etc.) on RIS–PACS clients or to an external user device if allowed by the administrator settings of the client system, or even sent to the PACS in DICOM format for permanent online storage. 2. Image processing tools on integrated RIS–PACS systems: an overview 2.1. 2D post-processing tools 2.1.1. Multiplanar reformat (MPR) As known, the MPR technique allows to create a bidimensional image on an arbitrary plane from a dataset of complanar images. More precisely, the information about position and signal intensity of voxels belonging to a stack of images acquired on parallel planes with a given spatial orientation (for example, a series of axial images) is used to produce an output image where voxels are geometrically projected on a user-defined plane with a different inclination (e.g. a coronal, sagittal or oblique plane), as a function of position and signal intensity of voxels from source images [7,8]. In order to achieve good quality MPR reconstructions, it is essential that image spatial resolution is high enough to approach voxel isotropy [9], so that no substantial loss of information occurs when image data are displayed on planes other than the native one. In this sense, the widespread availability of MDCT scanners and the progressive diffusion of high-field MRI equipment
213
(especially for vascular and neuroimaging applications) has been very beneficial to the generation of excellent MPR reconstructions with submillimetric voxel size. The MPR algorithm has the advantage of preserving the whole contrast resolution of native images, thus allowing reliable measurement of tissue density and contrast enhancement; indeed, on most integrated RIS–PACS systems it is possible to take density measurements on regions of interest (ROI) from MPR images, as well as geometric measurements, such as distances, areas, or angles. Moreover, almost all current integrated RIS–PACS systems with image processing options allow instant MPR of CT and MRI datasets; in addition, the majority of them allow to automatically perform MPR on standard planes (axial, coronal, sagittal: orthogonal MPR) by generating stacks of reformatted images spanning from a given initial to a given final limit of the source dataset, that can be saved as independent image series in DICOM format or even as a movie (Fig. 2). It is also possible to view current and previous PACS-retrieved studies of the same patient in MPR mode with image synchronisation, allowing for more accurate cross-study image comparison. More generally, the user is free to define oblique reconstruction planes arbitrarily by orienting and moving an electronic caliper on the three standard orthogonal planes, on which structures (e.g. blood vessels) are identified based on their anatomical course. In this way it is possible, for example, to create images on planes perpendicular to the longitudinal axis of a vessel in order to measure its diameter or section area, thereby avoiding the measurement error due to the geometric distortion that occurs whenever such measures are taken on native images from vessels coursing non-perpendicularly to the acquisition plane. A variant of the MPR algorithm is CPR (Curved Planar Reformation), that allows to project native images from a succession of planes lying onto a user-defined polygonal trajectory [8]. CPR is particularly useful, for example, to display vessels with a tortuous course in the three spatial directions, such as coronary arteries. This operation can be carried out by selecting a series of points aligned along the centre of the lumen (centreline) on one or more MPR projections, thus obtaining a continuous rectified representation of the vessel course (Fig. 3). With CPR it may also be easier to perform
Fig. 3. CPR of renal arteries (lower image) can be performed by manually tracing the vessel centreline (green dotted line) on MIP-reformatted magnetic resonance angiography source images (upper image).
214
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
Fig. 4. Usage of slice averaging allows to reduce image noise on MPR images (a, source images; b, orthogonal MPR with slice averaging).
Fig. 5. MIP of the left renal artery (right image) can be performed through the MPR interface by selecting slab thickness and orientation (blue parallel lines) on MPR views (upper and lower left images).
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
215
Fig. 6. minIP can be effective as to enhance visualisation of bronchial fistula (a, source MDCT axial images; b, MPR views).
measurements of vessel size, as the selected vessel is displayed continuously without distortions of its lumen contour. Options are sometimes offered that allow automatic or semi-automatic detection of the centreline, as described more thoroughly in Section 3, thus simplifying and/or quickening the radiologist’s work. Alongside ‘pure’ MPR viewing, it is possible to modify visualisation settings of MPR views by applying slice averaging tools [8] (Fig. 4), which are helpful for reducing image noise (therefore enabling radiation dose reduction in MDCT data acquisition), as well as effective image number, thus accelerating the whole image reading process. Of course, the degree of averaging must be set at
a level that allows relatively expedite image evaluation without missing findings of diagnostic importance due to partial volume effect; as a rule of thumb, at our Institution an average reconstruction thickness of 2–3 mm is considered to be a good compromise, given a native slice thickness of 0.625 mm for source MDCT axial images. 2.1.2. Maximum Intensity Projection (MIP) and Minimum Intensity Projection (minIP) Besides slice averaging, MPR-processed data can be displayed by using the Maximum Intensity Projection (MIP) or the Minimum
216
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
Fig. 7. By tracing the vessel centreline on sagittal MIP image (left), it is possible to obtain MIP view of CPR of the superior mesenteric artery (right): see accurate depiction of the artery and of its branches without substantial overlap of vessels from other vascular territories.
Intensity Projection (minIP) visualisation algorithms. As known, with the MIP technique a slab with a defined thickness and a given spatial orientation is selected, and for each set of voxels aligned along the chosen direction, only that with the highest intensity is represented on the resulting reconstructed image (Fig. 5). Similarly, with the minIP algorithm only the voxel with the lowest intensity is shown on the final image among those included in the slab [8] (Fig. 6). As a consequence, the MIP and minIP techniques are of primary importance for displaying anatomical structures with homogenous intensity that are not comprised into their morphological continuity on a single plane, such as bones and contrast-enhanced blood vessels and urinary system with MIP, and air-containing structures (e.g. airways or the large bowel) with minIP. As for MPR with slice averaging, it is possible for the user to select slab width and orientation, in order to adjust the degree of voxel superimposition. Some software packages also allow to perform MIP or minIP onto curved MPR views, which can be beneficial e.g. to get a panoramical view of branch vessels originating from a main trunk with a tortuous course (Fig. 7). Moreover, as unwanted anatomical structures may overlap with thick-slab MIP/minIP, some integrated RIS/PACS solutions offer the option of pre-editing the original dataset as to remove undesired structures, either through intensity threshold- and/or region growing-based segmentation algorithms, or by manual cutting on native and/or MPR images [10,11]. Other applications allow direct editing on MIP/minIP views, while others offer editing options on 3D views and permit MIP/minIP on datasets previously modified in 3D mode (Fig. 8). 2.2. 3D post-processing tools The server-based approach for remote data handling has been especially beneficial as for the possibility to perform 3D image processing on RIS–PACS stations, such as Volume Rendering (VR), Shaded Surface Display (SSD), and Virtual Endoscopy (VE). This is particularly true for VR algorithms, that require a relatively large amount of computational power as their output must retain the whole information content of the input dataset, in contrast with less hardware-demanding 2D tools (for instance, MIP uses approx-
Fig. 8. Coronal thick-slab MIP view generated after removal of the aorta and renal arteries allows optimal depiction of the hepatic arterial vasculature and the superior mesenteric artery system. In this case of a patient candidate to liver transplantation, an accessory artery to the right hepatic lobe stems from the superior mesenteric artery (red arrow).
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
217
Fig. 9. By choosing different transfer curves for VR, the user can emphasise different features of the image dataset: in this case of Leriche syndrome, a logarithmic inverse curve with a relatively low intensity threshold may be used to display collateral vessels (a), while one with a higher threshold allows better depiction of the aorto-iliac axes and splanchnic arteries (b). A linear curve with a very low threshold would obscure vessels but enable visualisation of soft tissues, such as the skin (c). In addition, the possibility for the user to create customised transfer curves is included (d).
218
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
Fig. 9. (Continued ).
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
219
Fig. 10. In this case of sacral metastasis from thyroid tumour vascularised from branches of the right internal iliac artery (a: MIP semi-axial image), it is possible to use different CLUT as to obtain optimal VR depiction of the relationship between the lesion and the host bone, together with its feeding vessels (b and c).
220
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
Fig. 11. VR with manual sculpting of patient table (a) and abdominal VR MDCT angiogram obtained after automatic bone removal (b); in this latter case, vessels of the arcade of Riolan were manually cropped for better image clarity.
imately 10% of the original information, because only voxels with highest intensity are represented on the output view). This way, a complete set of 2D and 3D options integrated on a RIS–PACS station is available to the user, virtually eliminating the need for him to move onto high-end standalone workstations for image processing. 2.2.1. Volume Rendering (VR) As known, while MIP is based on the extraction of maximum intensity voxels inside a user-defined slab among those crossed by
an array of parallel lines, VR algorithms harness the entire spatial and intensity information of the dataset as to represent on the reconstructed image a weighted average of the intensity of all voxels included in the slab. Each voxel is assigned a value of opacity, transparency, and colour depending on its intensity, position, and the prospective direction in which the image volume is observed according to a predetermined transfer function, that allows to generate the final image by weighing the intensity of the various voxels encountered by each ray. VR plug-ins let the
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
221
Fig. 12. (a) SSD view of lower limb arteries in patient with obstruction of right superficial femoral artery and aneurysm of the left superficial femoral artery; the dialogue box for colour and threshold settings is visible on the left half of the screen. (b) MIP of lower limb arteries (left) and magnified VR of popliteal and infrapopliteal arteries (right) are shown as a reference.
user define his own transfer function by choosing a colour and a transparency value and setting at least two voxel intensity thresholds, corresponding to the minimum and maximum luminance of visualisable points on the reconstructed image. So, it is possible to define linear transfer functions (implying that the ratio between the intensity of two voxels on native images and those of the corresponding pixels on the VR image is constant), as well as nonlinear (e.g. logarithmic or logarithmic inverse) ones, or even more complex functions with multiple thresholds, as to enhance or attenuate the contribution to the VR image of voxels within a given intensity range [7,8,12,13] (Fig. 9). It is also possible to choose among different colour look-up tables (CLUT) in order to enhance depiction of complex anatomical structures; preset CLUT are available on most RIS–PACS software, and can be manually customised by the user for
his convenience (Fig. 10). Furthermore, VR plug-ins often provide either automatic or manual image editing tools: automatic editing allows to remove unwanted voxels (usually bone or patient table) from the final image either by automatic software recognition or by mouse-clicking on the structures to be sculpted away, while manual editing can be done by the user by mouse-tracing the parts of the VR image to be preserved or deleted (Fig. 11). This enables accurate selection of the anatomy to be represented, thereby reducing or eliminating superimposition of surrounding structures that might hamper the diagnostic quality of the reconstructed VR capture. It is also possible to on-the-fly change the rendering mode for edited data (for example, from VR to MIP) as to optimise depiction of certain image features with the most appropriate post-processing technique [13]. VR images can be exported either in DICOM or non-
222
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
DICOM exchange format, or dynamically assembled in a movie rotating around a user-defined axis for offline file export. To this latter purpose, several file formats (such as AVI, MOV, or MPEG) are available, along with the possibility to use standard codecs for video encoding. 2.2.2. Shaded surface display (SSD) Some integrated RIS–PACS solutions also include SSD capabilities. Briefly, with this algorithm the user fixes a threshold defining the minimum intensity of the voxels to be imaged, while voxels with lower intensity will be discarded, and the reconstructed image will be composed by surfaces corresponding to the interface between voxels with intensity above and below the threshold (i.e. void voxels), obtained by simulating the illumination of surfaces from an external light source in a given user-defined direction [8,11,14]. In the past, given the availability of relatively lowpowered workstations, SSD was usually advantageous over VR due to its relatively modest computational requirements, because representation of dataset voxels on the output image is binary, allowing only a fraction of them to be processed. However, the binary nature of the SSD algorithm is also its main weakness, due to the difficulty to find an appropriate threshold for generation of nondistorted 3D images, causing hole-like artefacts to appear on output views in case of bad tissue separation. For this reason, and because heavy data processing is no longer a problem (as it is performed on remote servers), SSD has been superseded by the more robust and powerful VR algorithms, and few applications still support it (Fig. 12). 2.2.3. Virtual endoscopy (VE) By applying SSD or VR algorithms in order to explore hollow structures from an intraluminal, rather than extraluminal light source, it is possible to achieve endoscopy-like views. SSD-based VE techniques require the definition of an intensity threshold and generate surfaces formed by the interfaces between lower and higher intensity voxels. VR-based VE techniques are more robust and require the definition of a central intensity value (window level) and an intensity range (window width), corresponding to the signal intensity spectrum of voxels to be displayed with different opacities on the VE view. Most common integrated RIS–PACS image processing tools include VE capabilities, allowing to define window levels, opacity, and colour settings, together with a set of customisable CLUT, resembling the user interface of VR tools. On
some software packages it is even possible to toggle between VR and VE modes from the same dialogue panel, reflecting the conceptual and mechanistic similarity between those two techniques. The position and orientation of the virtual endoscope, together with the magnification degree of the VE view, can be defined by means of an electronic cursor onto synchronised orthogonal or oblique MPR projections (Fig. 13). It is also possible to create a centreline (‘flythrough’) along which VE views are taken, either automatically by software-assisted image segmentation, or manually by user tracing. This is particularly useful for exploration of anatomical cavities of homogenous intensity, such as airways, the colon, or contrastenhanced vessels or cardiac cavities. In addition, some programs allow to make geometric measurements from VE views, that can be useful e.g. for the assessment of vascular ostia from inside blood vessels or cardiac cavities [15]. Finally, software implementations exist that enable export of VE images either as standalone captures or in conjunction with MPR views, thus indicating the exact position from which VE is performed. As with VR reconstructions, VE screenshots can be saved either in DICOM mode for PACS storage, or in non-DICOM image format for offline file export, or even as a movie of consecutive VE images. This latter option is especially useful in case of fly-through VE, having the purpose of showing the internal structure of a hollow viscus or cavity through its entire length. 3. Dedicated image analysis plug-ins With the growing diffusion of cross-sectional imaging and the increasing size of image datasets over time (with particular reference to MDCT datasets), the need for image processing tools with a higher degree of automation has occurred. This holds especially true for cardiovascular imaging, in which the enormous amount of obtainable data has justified the development of automated image processing tools assisting the radiologist into the interpretation of findings from thousands of images in a reliable and reasonably fast fashion, compatibly with the routine workflow of a health service. Dedicated plug-ins for automated and semi-automated image analysis are currently available on most RIS–PACS software packages, ranging from CT colonography applications to solutions for cardiovascular imaging. Such tools integrate standard 2D and 3D post-processing algorithms with filtering and segmentation functions dedicated to one particular organ or anatomical territory, as to enable anatomy-specific image analysis. For example, CT colonography tools perform segmentation of air-containing voxels from the
Fig. 13. VE image of colonic polyp with synchronised orthogonal MPR views for virtual endoscope positioning.
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
223
Fig. 14. Segmentation of the right coronary artery with CPR view (a) and automatic vessel analysis (b) by means of the open-source CMIV CTA Plug-in for OsiriX (CMIV, Linköping University, Sweden).
original dataset (e.g. through region-growing algorithms), allowing both panoramic or targeted 2D and 3D visualisation of the large bowel, together with more organ-tailored functions such as virtual dissection, virtual biopsy, or virtual navigation; this latter can usually be accomplished with optimised window and CLUT settings through fly-through centreline-oriented VE. Another example is represented by vessel and cardiac analysis plug-ins, that allow extraction of voxels from contrast-enhanced vessels (automatic vessel tracking) by prior user definition of vessel start and end points (Fig. 14); subsequently, the software can derive geometrical information (such as vessel lumen diameter and section area, luminal stenosis or dilation degree, or aneurysm volume), leading to potential reduction of reader’s error due to inter- and intra-rater variability of manually taken measurements, as well as to lower overall image interpretation time. Of course, vessel tracking – as well as segmentation algorithms in general – can be controlled and eventually modified by the user in case of unsuccessful or erroneous performance, for example due to artefacts on native images. 4. Conclusions Integration of image processing tools with RIS–PACS systems represents an evolutionary step toward centralisation of all instru-
ments for medical image reading and reporting onto one single machine. This has the potential, on one hand, to save money and space by avoiding purchase of dedicated high-end workstations, and, on the other hand, to reduce overall reading time, thus resulting into increased productivity. We believe that with the increasing availability of adequate RIS–PACS systems with image processing capabilities, usage of standalone workstations should be reserved to selected cases of particular complexity, in which the added value of highly advanced and costly dedicated software might be relevant to diagnosis.
References [1] Carrino JA. Digital imaging overview. Semin Roentgenol 2003;38(3):200–15. [2] Tellis WM, Andriole KP, Jovais CS, Avrin DE. RIS minus PACS equals film. J Digit Imaging 2002;15(Suppl. 1):20–6. [3] Mildenberger P, Eichelberg M, Martin E. Introduction to the DICOM standard. Eur Radiol 2002;12(4):920–7. [4] Khorasani R. Image compression in your PACS: should you do it? What are the issues? J Am Coll Radiol 2004;1(10):780–1. [5] Kotter E, Baumann T, Jäger D, Langer M. Technologies for image distribution in hospitals. Eur Radiol 2006;16(6):1270–9. [6] Lee KH, Lee HJ, Kim JH, et al. Managing the CT data explosion: initial experiences of archiving volumetric datasets in a mini-PACS. J Digit Imaging 2005;18(3):188–95.
224
L. Faggioni et al. / European Journal of Radiology 78 (2011) 210–224
[7] Lell MM, Anders K, Uder M, et al. New techniques in CT angiography. Radiographics 2006;26(Suppl. 1):S45–62. [8] Cody DD. AAPM/RSNA physics tutorial for residents: topics in CT. Image processing in CT. Radiographics 2002;22(5):1255–68. [9] Dalrymple NC, Prasad SR, El-Merhi FM, Chintapalli KN. Price of isotropy in multidetector CT. Radiographics 2007;27(1):49–62. [10] Glockner JF. Three-dimensional gadolinium-enhanced MR angiography: applications for abdominal imaging. Radiographics 2001;21(2): 357–70. [11] Prokop M, Shin HO, Schanz A, Schaefer-Prokop CM. Use of maximum intensity projections in CT angiography: a basic review. Radiographics 1997;17(2):433–51.
[12] Calhoun PS, Kuszyk BS, Heath DG, Carley JC, Fishman EK. Three-dimensional volume rendering of spiral CT data: theory and method. Radiographics 1999;19(3):745–64. [13] Fishman EK, Ney DR, Heath DG, Corl FM, Horton KM, Johnson PT. Volume rendering versus maximum intensity projection in CT angiography: what works best, when, and why. Radiographics 2006;26(3):905–22. [14] Davis CP, Hany TF, Wildermuth S, Schmidt M, Debatin JF. Postprocessing techniques for gadolinium-enhanced three-dimensional MR angiography. Radiographics 1997;17(5):1061–77. [15] Cirillo S, Tosetti I, Gaita F, Bianchi F, Gandini G, Regge D. Magnetic resonance angiography of the pulmonary veins before and after radiofrequency ablation for atrial fibrillation. Radiol Med 2005;109(5–6):488–99.