- Email: [email protected]
Online artery diameter measurement in ultrasound images using artificial neural networks
Ultrasound in Med. & Biol., Vol. 28, No. 2, pp. 209 –216, 2002 Copyright © 2002 World Federation for Ultrasound in Medicine & Biology Printed in the USA. All rights reserved 0301-5629/02/$–see front matter
PII: S0301-5629(01)00505-1
● Original Contribution ONLINE ARTERY DIAMETER MEASUREMENT IN ULTRASOUND IMAGES USING ARTIFICIAL NEURAL NETWORKS V. R. NEWEY and D. K. NASSIRI Department of Medical Physics and Bioengineering, St. George’s Hospital, Blackshaw Road, Tooting, London, UK (Received 31 July 2001; in final form 20 November 2001)
Abstract—An automated online technique is described for measurement of artery diameter in flow-mediated dilation (FMD) ultrasound (US) images, using artificial neural networks to identify and track artery walls. This allows FMD results to be calculated without the inherent delay of current retrospective methods. Two networks were trained to identify artery anterior and posterior walls using over 3200 examples from carotid artery images. Both networks correctly classified approximately 97% of the randomly selected test samples. The technique was verified using a physical model with absolute measurement error of ⴚ1.16% ⴞ 1.04% (mean ⴞ SD) over the diameter range 2 to 8 mm. Advantages of the technique include: online analysis; wall tracking optimisation before the study proper; measurement of diameter changes over the cardiac cycle; low FMD measurement variance; minimal image degradation; and no unwieldy image store. Measurement of artery diameter changes over the cardiac cycle was explored using simulated image sequences generated with a virtual US scanner. (E-mail: [email protected]) © 2002 World Federation for Ultrasound in Medicine & Biology. Key Words: Online artery diameter measurement, Flow-mediated dilation, Artificial neural network, Ultrasound image, Virtual ultrasound scanner.
influence performance of the wall detection algorithm, but the effect will not be apparent until the study is completed. Additional operator effort and time is required for analysis. VCR recording involves some loss of image quality during record and playback, and computer storage may require a low acquisition rate or short acquisition period to avoid unwieldy archives. For example, storing a reduced image section of 256 ⫻ 256 pixels at 25 images s⫺1 requires almost 1 Gbyte of disk space for a 10-min study. It is not possible to recover the artery diameter changes as a function of the cardiac cycle if the image acquisition rate is low. Recovery of these changes allows diastole to be established without the need for image synchronisation and is potentially useful in elasticity studies. These problems can be resolved if the data are analysed online. Several groups have reported retrospective methods of automatically identifying vessel walls (Detmer et al. 1990; Gustavsson et al. 1994, 1997; Liang et al. 1998; Selzer et al. 1994). However, the reported techniques do not lend themselves easily to online analysis. For example, dynamic programming (Gustavsson et al. 1997) requires the images to be intensity-normalised, based on the average intensity over all images, and smoothed horizontally to remove vessel wall drop-outs. Smoothing
INTRODUCTION Flow-mediated dilation (FMD) is an established noninvasive technique for assessing endothelium dysfunction using ultrasound (US) images (Celermajer et al. 1992; Sorensen et al. 1995). The brachial artery is imaged both at rest and after ischaemia of the proximal forearm, and the percentage change in vessel diameter is an indicator of dysfunction. The scanner transducer is usually fixed spatially for the study with fine positional control, and a pneumatic tourniquet or cuff is used to induce ischaemia. The study typically lasts for 10 to 15 min and images are usually updated at diastole and recorded for retrospective analysis either on a video cassette recorder (VCR) or by computer storage. Artery diameter is calculated by detecting the vessel walls within a user-defined image region, using an automated computer algorithm. However, retrospective analysis can give rise to several problems. At commencement of the study, the images are optimised by adjusting the scanner controls to suit the visual preferences of the operator. This may adversely Address correspondence to: Mr. V. R. Newey, St. George’s Hospital, Department of Medical Physics and Bioengineering, Blackshaw Rd., Tooting, London SW17 OQT UK. E-mail: val@stgmic. demon.co.uk 209
210
Ultrasound in Medicine and Biology
may compromise accuracy due to blurring on all but straight horizontal vessel walls, and some rotational normalisation is, therefore, desirable. To recover diameter changes over the cardiac cycle, it is necessary to measure vessel diameter at twice the Nyquist frequency of the cardiac cycle to avoid aliasing. Output from most scanners is limited to 25 images s⫺1, giving 40 ms per image for all processing. The graph-search dynamic programming method of Liang et al. (1998) is computationally intensive, requiring 1 s per image for processing. Some current retrospective systems also require significant manual intervention in setting up and/or in monitoring performance. This paper describes a PC-based system that automatically locates the correct vessel walls in a user-defined region-of-interest (ROI) and measures vessel diameter online throughout a FMD study. The system was designed to minimise operator workload; the only user input normally required is to drag an orthogonal ROI over the required vessel section and to press a start button. METHODS Two simple artificial neural networks were trained to detect artery anterior and posterior walls. These networks are simplified models of biologic neural networks, with several characteristics that are useful in complex pattern recognition or classification problems. No rule definitions are required in the design process: a network simply learns from training examples. A correctly trained network can “generalize” from training examples when presented with novel cases. Classification performance can degrade progressively in the presence of increasing noise or incomplete data. The statistical rigour sometimes required in the classical design process can be less exacting (Tu 1996) and classes with complex separation boundaries and overlapping distributions and variances can be distinguished (Gevins et al. 1988). A simple network can be computationally efficient once trained, making online operation feasible. The artificial neurons or nodes are arranged in hierarchical layers from input to output, with one or more “hidden” layers (Fig. 1). The input layer scales data to a suitable range for the network. Nodes are joined through variable strength connections that are usually initialised before training to be randomly close to disconnection. A node output is a summation of its connection-weighted inputs multiplied by some nonlinear function, such as the sigmoid f(x) ⫽ (1 ⫹ e⫺ x)⫺ 1. The training method is to present the network with a set of input examples, each of which is from a known class. In this application, the input examples are short sections of vertical image lines from two output classes, vessel “wall” and “nonwall.”
Volume 28, Number 2, 2002
Each example propagates forward through the layers from input to output, where the difference between the desired and predicted classes can be calculated. The training objective is to minimise the errors between the predicted and desired outputs over the training set by gradually adjusting the connection strengths. The contribution to the output error from each node is found by propagating the output error backward to the input layer; hence, these are called “back-propagation” networks (Rumelhart et al. 1986). By training in this way, the network can “learn” any associations that may exist between input data and output classes and, after being correctly trained, can categorise original input data. After training, the connections can be viewed as long-term memory, and an important consideration is the ratio of training examples to network connections. A small ratio can result in a network that has “memorized” spurious, rather than general features of examples, giving poor classification performance when presented with examples not experienced in training (i.e., a test set). The suggested range for this ratio varies from ⬎ 3:1 to ⬎ 10:1 for networks using various numbers of hidden layers (Ahmed and Tesauro 1988; Fu 1994; Mehrotra et al. 1991). For the networks used here, the ratio was approximately 35:1. Network design Two simple back-propagation networks, as shown in Fig. 1, were trained to identify the artery anterior and posterior walls, respectively, by recognising the edge of the brightest echo at the vessel wall. The networks consisted of a vertical image line section of 21 pixels for input, one node in a hidden layer with sigmoid function and a single output node labeled 1 ⫽ “wall,” 0 ⫽ “nonwall.” The network output level represents the probability of the input data being a vessel wall. Training examples in approximately equal “wall” and “nonwall” categories were generated from a series of carotid US images selected to give a representative range of wall and tissue quality. Over 1600 examples per network, selected by an expert, were split into equal size partitions with random selection, and used to train and test a series of networks. When selecting the “wall” examples, the expert pointed to the edge of the bright echo at the blood-wall interface using a mouse cursor. The 21-pixel vertical sections were arranged to span the cursor locations symmetrically. For the “nonwall” category, tissue and blood examples within the images were selected randomly. The definitive networks correctly classified 96% to 97% of their respective test sets. The wall-blood interface echo is relatively independent of vessel diameter, and the network input line section length of 21 pixels was adequate for wall detection over a wide range of vessel size (2 to 8 mm) and image resolution (9 to 32
Online artery diameter ● V. R. NEWEY and D. K. NASSIRI
211
Fig. 1. Architecture of artificial neural networks used by the vessel wall detectors. Both networks consisted of one hidden node, one output node and sigmoid transfer function. The network output is the probability of the input being a vessel wall. For clarity, some input connections are not shown.
pixels per mm). The trained networks were then used to identify the spatial positions in an image ROI at which the points of highest wall probability occurred, and these positions were translated into vessel diameter. Wall-tracking scheme The tracking system was designed to identify the vessel walls in the first image by searching the entire ROI pixel-by-pixel with each network in turn, with the input line section centred on the current search location. This search generated matrices containing the probability of each ROI point being a vessel wall. The matrices were examined to select the most likely vessel wall starting points using a range of criteria. These included the wall probability level, wall order (i.e., anterior before posterior), wall separation distance and the mean intensity levels above and below the starting points. Walls were then propagated laterally from the starting points to create a vessel wall pair. Lateral propagation consisted of a small sideways step followed by a vertical search with the appropriate network to find the point at the new step with the highest probability of being a wall. This was repeated from the highest probability point at each step until the edges of the ROI were reached. The detected walls were then overlain on the image and then vessel diameter was calculated normal to each point along the vessel axis. The highest probability point on each wall was then used as the focus for wall propagation in the following image. Optimisation of the number of network inputs, vertical search length and horizontal step size minimised the search area in subsequent images. To
enhance tracking performance in poor-quality images, a range of factors was explored, including: refinement of the network output level representing the “wall” category; management of poor-quality wall points by interpolation; constraining the vertical separation between adjacent wall points and the wall movement between consecutive images; increasing network complexity by adding more nodes, inputs or outputs; and contextual reinforcement (i.e., adding inputs to the networks from pixels in the vicinity of the input pixel line). Examples of tracking performance with poor and good quality Acuson XP/10 images are shown in Figs. 2 and 3, respectively. Online graphical displays of vessel diameter and wall probability levels were provided for quality assurance. The tracking process was entirely automatic, except for selection of the ROI. Controls were provided for minor ROI adjustments, and wall tracking could be manually overridden online by pointing and clicking. A FMD graph for a study is shown in Fig. 4. In practice, displaying data at 25 images s⫺1 gave cluttered graphic visualisation due to the large volume of information. The graph of Fig. 4 has, therefore, been smoothed using a 2-s running average filter to give an approximately mean value over the cardiac cycle. This was adjusted, after recovery of the cardiac cycle diametric changes, to show diastolic diameter. Recovery of diameter changes over the cardiac cycle In FMD studies, it is customary to measure dilation at diastole. However, with current retrospective walltracking techniques, this requires synchronisation of the
212
Ultrasound in Medicine and Biology
Volume 28, Number 2, 2002
Fig. 4. Graph of artery diameter variation in a clinical FMD study for a normal subject. Dilation for this subject when comparing data sections before “cuff inflate” and between the “peak start” and “peak end” cursors was 7.34% and 7.54% on two separate occasions on different days. Variation in measured diameter due to respiration is visible as a low-amplitude oscillation. Fig. 2. Wall-tracking performance with a low-quality image. The walls have been automatically located within the ROIs of Figs. 2 and 3. The operator was only required to drag the ROI box and click a start button. Arteries are shown both with and without wall markers, for comparison purposes.
scanner image to one point in the ECG. The scanner image is, therefore, only updated once in every heartbeat and information that is potentially useful (e.g., in vessel elasticity studies) is lost. By continuously updating the scanner image and measuring the vessel diameter at 25 images s⫺1, it is possible to recover vessel diameter changes over the cardiac cycle as the vessel expands and contracts with blood pressure changes. The changes involved can be very small, approaching the spatial resolution of the US images, and it is necessary to average a series of cardiac events temporally to quantify changes. Although changes of less than one pixel in magnitude that do not cross pixel boundaries will not be detected, in
Fig. 3. Wall tracking in a good-quality image.
practice, involuntary movement (e.g., respiration) (see Fig. 4) and inclined vessel walls cause some pixel boundary changes. The situation can be improved by zooming the image (zoom levels of 10 to 30 pixels/mm are usual in FMD studies) to give higher spatial resolution. Cardiac events were detected using a cross-correlation search technique that locates the most highly correlated events. A section of the study data (the correlation window) of random width, but within the range 0.3 to 2 s, and random location was cross-correlated with the study data to obtain the locations of correlated events. The detected events were averaged after normalisation to heart rate, and the search was directed to maximise the systolic-to-diastolic difference by fine-tuning the correlation window width and position. The diameter changes over the cardiac cycle can be recovered reliably by this method, as shown in Fig. 5. However, poor image quality may cause distortion through repetitious errors, such as consistent tracking failure at the same point in the cardiac cycle (e.g., diastole). To demonstrate accurate recovery of diameter changes over the cardiac cycle using the technique, it was necessary to simulate the very small wall displacement involved. It was not possible, in practice, to generate images precisely with the required “worst case” wall movement using a conventional scanner and phantom vessel. For example, it would be necessary to generate physical movement of potentially ⬍ 1 image pixel dynamically over the cardiac cycle in a vessel phantom, while keeping the vessel horizontal to within one pixel. Consideration of scanner noise introduces a further level of uncertainty. To provide a more precise and repeatable evaluation, the correlation technique was, therefore, validated using a simulated US scanner and images (i.e., a virtual US scanner) (Newey and Nassiri 2000). It was possible, using this method, to modulate vessel diameter in a virtual cardiac time series.
Online artery diameter ● V. R. NEWEY and D. K. NASSIRI
Fig. 5. Variation in vessel diameter over the cardiac cycle recovered from a simulated image sequence (r ⫽ 0.99). The cardiac event (– – – –) used for this test was recovered from 1.5 min of the baseline clinical data (precuff inflation) in Fig. 4 (minimum diameter was adjusted to 4 mm for this test).
Virtual ultrasound scanner The objective here was to simulate the artery images from a 6-MHz linear-array transducer, and a Toshiba PowerVision 6000 scanner with L80-6.0 transducer was used as a control when assessing fidelity. Soft tissue image texture, blood, vessel wall, tissue attenuation of US, scanner noise, gain, time-gain compensation (TGC), logarithmic signal compression and zoom were modelled. The virtual scanner produced simulated images, based on the technique of Nassiri et al. (1983), containing parallel walled horizontal arteries of known diameter and depth. This technique operates by convolving (Gonzalez and Wintz 1987) a 2-D simulated US radiofrequency (RF) point spread function with a “structure” image to give an RF image. In the structure image, soft tissue was modelled as a random distribution of point scatterers of appropriate strength for muscle and blood, with parallel lines added for the vessel walls. The brightness of echoes in the RF image was controlled by adjusting the amplitude of points/lines in the structure image and vessel wall drop-out was simulated by modulating the amplitude of individual points along the lines. The absolute value of each vertical line in the RF image was then filtered to match the rectification and smoothing processes used in the formation of a conventional Bmode image. The resulting B-mode image was attenuated by 0.5 dB/cm MHz and a first order noise approximation was added to give echoes of decreasing signalto-noise ratio (SNR) with beam penetration. To duplicate the effects of scanner gain, TGC and compression, the final image was multiplied by user-adjustable gain and TGC controls and compressed logarithmically. The virtual scanner and an example image are shown in Fig. 6.
213
Fig. 6. Virtual US scanner showing a simulated horizontal vessel of 4 mm diameter, calibration markers and scanner gain and TGC controls.
Imaging system A fully integrated application software package VIA (vascular image analysis) was written for the project, together with the virtual US scanner using National Instruments LabVIEW and IMAQ (Austin, Texas, USA). The neural networks were developed using NeuralWorks Professional II/PLUS (Pittsburgh, PA, USA). The system was designed to analyse online or computerstored images for development purposes. The system operated online at 25 images s⫺1 on a 400-MHz PC, with ROI width of typically 2 to 3 cm. An Acuson 128 XP/10 US scanner (Calif., USA) was initially used for clinical imaging, followed by an ATL HDI 3000 scanner (Eindhoven, Netherlands). 7-MHz linear array transducers were used in both cases and the scanner image outputs were connected to a National Instruments PCI-1408 frame-grabber optimised accordingly for the dynamic range. The Acuson 128 scanner was configured for vascular imaging with transmit power of ⫺9 dB, log compression 40 dB, persistence A, preprocessing curve 2 (crisp borders), postprocessing curve 5 (high contrast), overall gain 3 dB. The HDI 3000 scanner was configured for linear map 1, 35 dB gain and persistence off. The tourniquet deflation point in a FMD study represents a period of relatively high workload for the operator. For example, it is necessary to deflate the tourniquet, insert a data-logging marker, monitor image quality and, possibly, refine the transducer position, etc. To minimise this workload, the study timing and marker insertion were managed by the system, with audio/visual operator prompts at relevant points. It was possible, therefore, to handle image-scale alterations automatically when changing to Doppler mode for flow validation immediately following tourniquet deflation.
214
Ultrasound in Medicine and Biology
Volume 28, Number 2, 2002
Table 1. Measurement error as a function of vessel diameter ⫽ ⫺1.16% ⫾ 1.04% (mean ⫾ SD) Expected mm 2.00 3.00 4.00 5.00 6.00 7.00 8.00 Measured mm 1.98 2.97 3.90 5.00 6.01 6.90 7.81 Error % ⫺1.00 ⫺1.00 ⫺2.50 0.00 0.16 ⫺1.43 ⫺2.37
SYSTEM VALIDATION Measurement error The vessel absolute diameter measurement error of the technique was assessed (Table 1) using a physical model. This consisted of two parallel 0.25-mm diameter nylon monofilaments mounted under tension on vernier calipers so that the distance between the filaments was known. The filaments were immersed in a water/glycerol (9% volume) mixture with an US propagation velocity of 1540 m s⫺1, and imaged using an ATL HDI 3000 scanner (Fig. 7) at a zoom level of 18.5 pixels/mm. The distance between the filaments was varied from 2 to 8 mm in 1-mm steps and measured with the wall-tracking software after adjusting the scanner controls to minimise changes in appearance of the filaments caused by spatial nonuniformity of the US beam. Recovery of diameter changes over the cardiac cycle Recovery was evaluated using a simulated cardiac image sequence generated with the virtual US scanner and representing about 10 min of data at 25 images per s. Movement of the vessel walls and the effects of noise can be precisely and repeatably modelled using this method. The simulation was based on a vessel of 4 mm nominal diameter, with parallel horizontal walls. The diameter was modulated by the cardiac cycle shown in Fig. 5 (dotted line), giving a maximum change of 0.63 pixel over the cardiac cycle. This modulation cycle was recovered from approximately 2 min of data before the “cuff inflate” cursor of the FMD study shown in Fig. 4.
Fig. 8. Correlation between original and recovered cardiac cycles as a function of the number of cardiac events averaged. Events from approximately 3 to 10 min of data were required to give reasonable correlation. In clinical practice, 1 to 3 min of baseline data at 25 images/s was generally adequate.
In addition to the cardiac cycle modulation, a random vertical movement of the vessel in the range 0 to 0.2 mm (0 to 2.52 pixels) was added to each image to simulate spatial noise. Wall brightness was also modulated to simulate the slight variation seen over the cardiac cycle in clinical studies. This combination of parallel, perfectly horizontal walls with less than one pixel movement throughout the cardiac cycle represents an extreme case, likely to be experienced rarely in practice. Sequence lengths of approximately 15,000 images were generated and vessel diameter measured throughout each sequence using the neural network wall detectors. Correlation between actual and recovered events as a function of the number of cardiac events (Fig. 8) was explored to assess the clinical requirements for adequate recovery. Calculation of SNR provided an index of the quality of recovered cardiac cycle changes. DISCUSSION
Fig. 7. Example monofilament wire image used for assessing measurement accuracy.
This technique has several advantages over established practice. FMD results are available immediately. Data for the complete study are available, rather than just a few salient events, and changes in vessel diameter over the cardiac cycle are recovered. Vessel wall tracking quality can be optimised by adjustment of the scanner controls before the study proper. This largely normalises the intensity range of the images to the range presented to the neural networks in the training regimen, thereby optimising the images to suit the networks. With retrospective analysis, however, images are subjectively optimised for visual, rather than wall detection properties and the detection algorithm must, therefore, handle a wider range of image quality. Online analysis provides
Online artery diameter ● V. R. NEWEY and D. K. NASSIRI
continuous graphical presentations of vessel diameter, detected walls and wall probability levels as tracking quality indicators. These features, some of which are only available retrospectively with off-line techniques, make minor changes in image and tracking quality and vessel alignment readily apparent at a time when they can be rectified. The method also avoids the image quality loss associated with VCR recording, and no processing of the image is necessary. Retrospective systems, however, often require processing steps, such as intensity or rotational normalisation or smoothing which, although improving the tracking qualities, may cause some spatial degradation of the image. There is no unwieldy image archive; file size for a complete study is typically approximately 0.3 Mb and the study is completed much faster than retrospective methods, contributing to lower study costs. Some of these factors probably contribute to the low measurement variance of 0.9% ⫾0.6% (mean ⫾ SD) and coefficient of variance 0.66% found in a clinical trial (Sidhu et al. 2001). The variance compares favourably with the findings of Preik et al. (2000) of 1.3% ⫾ 0.9% and 0.78%, respectively, when applying the retrospective graph-search dynamic programming system of Liang et al. (1998). It should be noted that these figures include biologic dilation variability, as well as measurement variation due to the technique. Although online analysis demands vigilance at critical study points, such as tourniquet deflation, this is balanced by the automated study timing. Further, this allows the system automatically to select baseline and dilation data segments for final analysis. After case completion, the operator is generally only required to select the data file and to press a button to calculate dilation, a significant reduction in analytical workload. The measurement error of ⫺1.16% ⫾ 1.04% (mean ⫾ SD) in Table 1 is probably attributable to the US echo axial size (Li et al. 1993). This will cause an underestimation of the distance between filaments in the physical model. In practice, a secondary source of error is the network training method of pointing to the edge of the brightest echo at the vessel wall/blood interface when selecting the “wall” training category. Although conferring more robust tracking, these points correspond to the media interface. The intima interface was ignored and the technique, therefore, overestimates vessel diameter because the media-intima distance represents the wall thickness. However, these error sources should not be significant in FMD studies, where relative rather than absolute change in diameter is of importance. As with all FMD studies, good image quality is mandatory for both accurate measurements and optimum wall tracking. The technique can also be used to detect nonhorizontal vessel interfaces by searching the ROI sequentially with network input line sections at more than one orientation
215
(i.e., vertical and horizontal) and combining the outputs to obtain the probability matrix. The simulated cardiac sequence required 7 to 8 min of data to recover diameter changes over the cardiac cycle with reasonable quality (Fig. 5). In clinical practice, however, the cardiac cycle is generally recovered with good quality from 1 to 3 min of data. Recovery of diameter changes over the cardiac cycle may be useful in other studies, such as investigation of the elastic properties of the vessel wall, where wall displacement combined with blood pressure can be used to calculate elasticity indices. The virtual US scanner technique can be applied in other areas of US image analysis (e.g., in B-mode image quality assurance) (Newey and Nassiri 2000). The wall-tracking technique has proven to be reliable under a range of operating conditions. The training, simulation and clinical studies all used images from different types of scanners with a range of zoom settings (9 to 32 pixels/mm). In addition, the training images were of the carotid artery, and clinical FMD studies have used the brachial artery. To date, over 120 clinical cases in continuing brachial FMD and carotid elasticity studies (to be reported) have been analysed at St. George’s Hospital Medical School using an ATL HDI 3000 scanner. In all cases, the correct vessel walls were automatically located in the ROI. Manual intervention was required rarely when image quality deteriorated to a point where walls disappeared and reappeared at a substantially different location within or outside the ROI. The design aim was the development of fast and, therefore, simple neural networks to attain online performance. As faster PCs become available, more complex neural networks trained, for example, with additional contextual information, could be used to further enhance the robustness of the method. Acknowledgements—The authors gratefully thank J. S. Sidhu and V. Mwansa of St. George’s Hospital Medical School, Department of Cardiological Sciences, for suggesting operational refinements to the system and clinical support. They also thank N. Parchure and G. Rosso for their clinical contributions.
REFERENCES Ahmad S, Tesauro G. Scaling and generalization in neural networks: A case study. Proc Neural Info Proc Syst Conf (NIPS) 1988;1:160 – 176. Celermajer DS, Sorensen KE, Gooch V.M., et al. Non-invasive detection of endothelial dysfunction in children and adults at risk of atherosclerosis. Lancet 1992;340:1111–1115. Detmer P, Bashein G, Martin R. Matched filter identification of leftventricle endocardial borders in transesophageal echocardiograms. IEEE Trans Med Imaging 1990;9:396 – 404. Fu LM. Neural networks in computer intelligence. New York: McGraw-Hill, 1994. Gevins AS, Stone R., Ragsdale SD. Differentiating the effects of three benzodiazapines on non-REM sleep EEG spectra: A neural network pattern classification analysis. Neuropsychobiol 1988;19(2):108 – 115.
216
Ultrasound in Medicine and Biology
Gonzalez RC, Wintz P. Digital image processing. Addison Wesley, 1987. Gustavsson T, Gharbieh A, Hamarneh G, Liang Q. Implementation and comparison of four different boundary detection algorithms for quantitative ultrasonic measurements of the human carotid artery. Comput Cardiol 1997;24:69 –72. Gustavsson T, Liang Q, Wendelhag I, Wikstrand J. A dynamic programming procedure for automated ultrasonic measurement of the carotid artery. Comput Cardiol 1994;297–300. Li S, McDicken WN, Hoskin PR. Blood vessel diameter measurement by ultrasound. Physiol Meas 1993;14:291–297. Liang W, Browning R, Lauer R, Sonka M. Automated analysis of brachial ultrasound time series. Multi-dimensional images. SPIE 1998;3337:108 –118. Mehrotra KG, Chilukuri KM, Ranka S. Bounds on the number of samples needed for neural learning. IEEE Trans Neural Networks 1991;2:548 –558. Nassiri DK, Nicholas D, Hill CR. B-scan texture classification: A study using physical and theoretical models. In: Lerski RA, Morley P, eds. Third meeting of the World Federation for Ultrasound in Medicine & Biology. Oxford: Pergamon Press, 1983:133–139. Newey VR, Nassiri DK. Analysis of variability in visual B-mode image
Volume 28, Number 2, 2002 evaluation; Application of a virtual scanner and phantom. Eur J Ultrasound 2000;13:1S. Preik M, Lauer T, Hei S, et al. Automated ultrasonic measurement of human arteries for the determination of endothelial function. Ultraschall Med 2000;21:195–198. Rumelhart DE, Hinton GE, Williams RJ. Learning internal representations by error propagation. In: Rumelhart DE, McClelland JL, eds. Parallel distributed processing: Exploration in the microstructure of cognition. Vol. 1. Cambridge: MIT Press, 1986. Selzer RH, Hodis H, Kwong-Fu H, et al. Evaluation of computerized edge tracking for quantifying intima-media thickness of the common carotid artery from B-mode ultrasound images. Atherosclerosis 1994;111:1–11. Sidhu JS, Mwansa V, Newey VR, Nassiri DK, Kaski JC. The development of an on-line, automated technique to determine endothelial function. Circulation 2001;104(17):supplement II, abstract 1161. Sorensen KE, Celermajer DS, Spiegelhalter DJ, et al. Non-invasive measurement of human endothelium dependent arterial responses: Accuracy and reproducibility. Br Heart J 1995;74:247–253. Tu JV. Advantages and disadvantages of using artificial neural networks versus logistic regression for predicting medical outcomes. J Clin Epidemiol 1996;49:1225–1231.
PII: S0301-5629(01)00505-1
● Original Contribution ONLINE ARTERY DIAMETER MEASUREMENT IN ULTRASOUND IMAGES USING ARTIFICIAL NEURAL NETWORKS V. R. NEWEY and D. K. NASSIRI Department of Medical Physics and Bioengineering, St. George’s Hospital, Blackshaw Road, Tooting, London, UK (Received 31 July 2001; in final form 20 November 2001)
Abstract—An automated online technique is described for measurement of artery diameter in flow-mediated dilation (FMD) ultrasound (US) images, using artificial neural networks to identify and track artery walls. This allows FMD results to be calculated without the inherent delay of current retrospective methods. Two networks were trained to identify artery anterior and posterior walls using over 3200 examples from carotid artery images. Both networks correctly classified approximately 97% of the randomly selected test samples. The technique was verified using a physical model with absolute measurement error of ⴚ1.16% ⴞ 1.04% (mean ⴞ SD) over the diameter range 2 to 8 mm. Advantages of the technique include: online analysis; wall tracking optimisation before the study proper; measurement of diameter changes over the cardiac cycle; low FMD measurement variance; minimal image degradation; and no unwieldy image store. Measurement of artery diameter changes over the cardiac cycle was explored using simulated image sequences generated with a virtual US scanner. (E-mail: [email protected]) © 2002 World Federation for Ultrasound in Medicine & Biology. Key Words: Online artery diameter measurement, Flow-mediated dilation, Artificial neural network, Ultrasound image, Virtual ultrasound scanner.
influence performance of the wall detection algorithm, but the effect will not be apparent until the study is completed. Additional operator effort and time is required for analysis. VCR recording involves some loss of image quality during record and playback, and computer storage may require a low acquisition rate or short acquisition period to avoid unwieldy archives. For example, storing a reduced image section of 256 ⫻ 256 pixels at 25 images s⫺1 requires almost 1 Gbyte of disk space for a 10-min study. It is not possible to recover the artery diameter changes as a function of the cardiac cycle if the image acquisition rate is low. Recovery of these changes allows diastole to be established without the need for image synchronisation and is potentially useful in elasticity studies. These problems can be resolved if the data are analysed online. Several groups have reported retrospective methods of automatically identifying vessel walls (Detmer et al. 1990; Gustavsson et al. 1994, 1997; Liang et al. 1998; Selzer et al. 1994). However, the reported techniques do not lend themselves easily to online analysis. For example, dynamic programming (Gustavsson et al. 1997) requires the images to be intensity-normalised, based on the average intensity over all images, and smoothed horizontally to remove vessel wall drop-outs. Smoothing
INTRODUCTION Flow-mediated dilation (FMD) is an established noninvasive technique for assessing endothelium dysfunction using ultrasound (US) images (Celermajer et al. 1992; Sorensen et al. 1995). The brachial artery is imaged both at rest and after ischaemia of the proximal forearm, and the percentage change in vessel diameter is an indicator of dysfunction. The scanner transducer is usually fixed spatially for the study with fine positional control, and a pneumatic tourniquet or cuff is used to induce ischaemia. The study typically lasts for 10 to 15 min and images are usually updated at diastole and recorded for retrospective analysis either on a video cassette recorder (VCR) or by computer storage. Artery diameter is calculated by detecting the vessel walls within a user-defined image region, using an automated computer algorithm. However, retrospective analysis can give rise to several problems. At commencement of the study, the images are optimised by adjusting the scanner controls to suit the visual preferences of the operator. This may adversely Address correspondence to: Mr. V. R. Newey, St. George’s Hospital, Department of Medical Physics and Bioengineering, Blackshaw Rd., Tooting, London SW17 OQT UK. E-mail: val@stgmic. demon.co.uk 209
210
Ultrasound in Medicine and Biology
may compromise accuracy due to blurring on all but straight horizontal vessel walls, and some rotational normalisation is, therefore, desirable. To recover diameter changes over the cardiac cycle, it is necessary to measure vessel diameter at twice the Nyquist frequency of the cardiac cycle to avoid aliasing. Output from most scanners is limited to 25 images s⫺1, giving 40 ms per image for all processing. The graph-search dynamic programming method of Liang et al. (1998) is computationally intensive, requiring 1 s per image for processing. Some current retrospective systems also require significant manual intervention in setting up and/or in monitoring performance. This paper describes a PC-based system that automatically locates the correct vessel walls in a user-defined region-of-interest (ROI) and measures vessel diameter online throughout a FMD study. The system was designed to minimise operator workload; the only user input normally required is to drag an orthogonal ROI over the required vessel section and to press a start button. METHODS Two simple artificial neural networks were trained to detect artery anterior and posterior walls. These networks are simplified models of biologic neural networks, with several characteristics that are useful in complex pattern recognition or classification problems. No rule definitions are required in the design process: a network simply learns from training examples. A correctly trained network can “generalize” from training examples when presented with novel cases. Classification performance can degrade progressively in the presence of increasing noise or incomplete data. The statistical rigour sometimes required in the classical design process can be less exacting (Tu 1996) and classes with complex separation boundaries and overlapping distributions and variances can be distinguished (Gevins et al. 1988). A simple network can be computationally efficient once trained, making online operation feasible. The artificial neurons or nodes are arranged in hierarchical layers from input to output, with one or more “hidden” layers (Fig. 1). The input layer scales data to a suitable range for the network. Nodes are joined through variable strength connections that are usually initialised before training to be randomly close to disconnection. A node output is a summation of its connection-weighted inputs multiplied by some nonlinear function, such as the sigmoid f(x) ⫽ (1 ⫹ e⫺ x)⫺ 1. The training method is to present the network with a set of input examples, each of which is from a known class. In this application, the input examples are short sections of vertical image lines from two output classes, vessel “wall” and “nonwall.”
Volume 28, Number 2, 2002
Each example propagates forward through the layers from input to output, where the difference between the desired and predicted classes can be calculated. The training objective is to minimise the errors between the predicted and desired outputs over the training set by gradually adjusting the connection strengths. The contribution to the output error from each node is found by propagating the output error backward to the input layer; hence, these are called “back-propagation” networks (Rumelhart et al. 1986). By training in this way, the network can “learn” any associations that may exist between input data and output classes and, after being correctly trained, can categorise original input data. After training, the connections can be viewed as long-term memory, and an important consideration is the ratio of training examples to network connections. A small ratio can result in a network that has “memorized” spurious, rather than general features of examples, giving poor classification performance when presented with examples not experienced in training (i.e., a test set). The suggested range for this ratio varies from ⬎ 3:1 to ⬎ 10:1 for networks using various numbers of hidden layers (Ahmed and Tesauro 1988; Fu 1994; Mehrotra et al. 1991). For the networks used here, the ratio was approximately 35:1. Network design Two simple back-propagation networks, as shown in Fig. 1, were trained to identify the artery anterior and posterior walls, respectively, by recognising the edge of the brightest echo at the vessel wall. The networks consisted of a vertical image line section of 21 pixels for input, one node in a hidden layer with sigmoid function and a single output node labeled 1 ⫽ “wall,” 0 ⫽ “nonwall.” The network output level represents the probability of the input data being a vessel wall. Training examples in approximately equal “wall” and “nonwall” categories were generated from a series of carotid US images selected to give a representative range of wall and tissue quality. Over 1600 examples per network, selected by an expert, were split into equal size partitions with random selection, and used to train and test a series of networks. When selecting the “wall” examples, the expert pointed to the edge of the bright echo at the blood-wall interface using a mouse cursor. The 21-pixel vertical sections were arranged to span the cursor locations symmetrically. For the “nonwall” category, tissue and blood examples within the images were selected randomly. The definitive networks correctly classified 96% to 97% of their respective test sets. The wall-blood interface echo is relatively independent of vessel diameter, and the network input line section length of 21 pixels was adequate for wall detection over a wide range of vessel size (2 to 8 mm) and image resolution (9 to 32
Online artery diameter ● V. R. NEWEY and D. K. NASSIRI
211
Fig. 1. Architecture of artificial neural networks used by the vessel wall detectors. Both networks consisted of one hidden node, one output node and sigmoid transfer function. The network output is the probability of the input being a vessel wall. For clarity, some input connections are not shown.
pixels per mm). The trained networks were then used to identify the spatial positions in an image ROI at which the points of highest wall probability occurred, and these positions were translated into vessel diameter. Wall-tracking scheme The tracking system was designed to identify the vessel walls in the first image by searching the entire ROI pixel-by-pixel with each network in turn, with the input line section centred on the current search location. This search generated matrices containing the probability of each ROI point being a vessel wall. The matrices were examined to select the most likely vessel wall starting points using a range of criteria. These included the wall probability level, wall order (i.e., anterior before posterior), wall separation distance and the mean intensity levels above and below the starting points. Walls were then propagated laterally from the starting points to create a vessel wall pair. Lateral propagation consisted of a small sideways step followed by a vertical search with the appropriate network to find the point at the new step with the highest probability of being a wall. This was repeated from the highest probability point at each step until the edges of the ROI were reached. The detected walls were then overlain on the image and then vessel diameter was calculated normal to each point along the vessel axis. The highest probability point on each wall was then used as the focus for wall propagation in the following image. Optimisation of the number of network inputs, vertical search length and horizontal step size minimised the search area in subsequent images. To
enhance tracking performance in poor-quality images, a range of factors was explored, including: refinement of the network output level representing the “wall” category; management of poor-quality wall points by interpolation; constraining the vertical separation between adjacent wall points and the wall movement between consecutive images; increasing network complexity by adding more nodes, inputs or outputs; and contextual reinforcement (i.e., adding inputs to the networks from pixels in the vicinity of the input pixel line). Examples of tracking performance with poor and good quality Acuson XP/10 images are shown in Figs. 2 and 3, respectively. Online graphical displays of vessel diameter and wall probability levels were provided for quality assurance. The tracking process was entirely automatic, except for selection of the ROI. Controls were provided for minor ROI adjustments, and wall tracking could be manually overridden online by pointing and clicking. A FMD graph for a study is shown in Fig. 4. In practice, displaying data at 25 images s⫺1 gave cluttered graphic visualisation due to the large volume of information. The graph of Fig. 4 has, therefore, been smoothed using a 2-s running average filter to give an approximately mean value over the cardiac cycle. This was adjusted, after recovery of the cardiac cycle diametric changes, to show diastolic diameter. Recovery of diameter changes over the cardiac cycle In FMD studies, it is customary to measure dilation at diastole. However, with current retrospective walltracking techniques, this requires synchronisation of the
212
Ultrasound in Medicine and Biology
Volume 28, Number 2, 2002
Fig. 4. Graph of artery diameter variation in a clinical FMD study for a normal subject. Dilation for this subject when comparing data sections before “cuff inflate” and between the “peak start” and “peak end” cursors was 7.34% and 7.54% on two separate occasions on different days. Variation in measured diameter due to respiration is visible as a low-amplitude oscillation. Fig. 2. Wall-tracking performance with a low-quality image. The walls have been automatically located within the ROIs of Figs. 2 and 3. The operator was only required to drag the ROI box and click a start button. Arteries are shown both with and without wall markers, for comparison purposes.
scanner image to one point in the ECG. The scanner image is, therefore, only updated once in every heartbeat and information that is potentially useful (e.g., in vessel elasticity studies) is lost. By continuously updating the scanner image and measuring the vessel diameter at 25 images s⫺1, it is possible to recover vessel diameter changes over the cardiac cycle as the vessel expands and contracts with blood pressure changes. The changes involved can be very small, approaching the spatial resolution of the US images, and it is necessary to average a series of cardiac events temporally to quantify changes. Although changes of less than one pixel in magnitude that do not cross pixel boundaries will not be detected, in
Fig. 3. Wall tracking in a good-quality image.
practice, involuntary movement (e.g., respiration) (see Fig. 4) and inclined vessel walls cause some pixel boundary changes. The situation can be improved by zooming the image (zoom levels of 10 to 30 pixels/mm are usual in FMD studies) to give higher spatial resolution. Cardiac events were detected using a cross-correlation search technique that locates the most highly correlated events. A section of the study data (the correlation window) of random width, but within the range 0.3 to 2 s, and random location was cross-correlated with the study data to obtain the locations of correlated events. The detected events were averaged after normalisation to heart rate, and the search was directed to maximise the systolic-to-diastolic difference by fine-tuning the correlation window width and position. The diameter changes over the cardiac cycle can be recovered reliably by this method, as shown in Fig. 5. However, poor image quality may cause distortion through repetitious errors, such as consistent tracking failure at the same point in the cardiac cycle (e.g., diastole). To demonstrate accurate recovery of diameter changes over the cardiac cycle using the technique, it was necessary to simulate the very small wall displacement involved. It was not possible, in practice, to generate images precisely with the required “worst case” wall movement using a conventional scanner and phantom vessel. For example, it would be necessary to generate physical movement of potentially ⬍ 1 image pixel dynamically over the cardiac cycle in a vessel phantom, while keeping the vessel horizontal to within one pixel. Consideration of scanner noise introduces a further level of uncertainty. To provide a more precise and repeatable evaluation, the correlation technique was, therefore, validated using a simulated US scanner and images (i.e., a virtual US scanner) (Newey and Nassiri 2000). It was possible, using this method, to modulate vessel diameter in a virtual cardiac time series.
Online artery diameter ● V. R. NEWEY and D. K. NASSIRI
Fig. 5. Variation in vessel diameter over the cardiac cycle recovered from a simulated image sequence (r ⫽ 0.99). The cardiac event (– – – –) used for this test was recovered from 1.5 min of the baseline clinical data (precuff inflation) in Fig. 4 (minimum diameter was adjusted to 4 mm for this test).
Virtual ultrasound scanner The objective here was to simulate the artery images from a 6-MHz linear-array transducer, and a Toshiba PowerVision 6000 scanner with L80-6.0 transducer was used as a control when assessing fidelity. Soft tissue image texture, blood, vessel wall, tissue attenuation of US, scanner noise, gain, time-gain compensation (TGC), logarithmic signal compression and zoom were modelled. The virtual scanner produced simulated images, based on the technique of Nassiri et al. (1983), containing parallel walled horizontal arteries of known diameter and depth. This technique operates by convolving (Gonzalez and Wintz 1987) a 2-D simulated US radiofrequency (RF) point spread function with a “structure” image to give an RF image. In the structure image, soft tissue was modelled as a random distribution of point scatterers of appropriate strength for muscle and blood, with parallel lines added for the vessel walls. The brightness of echoes in the RF image was controlled by adjusting the amplitude of points/lines in the structure image and vessel wall drop-out was simulated by modulating the amplitude of individual points along the lines. The absolute value of each vertical line in the RF image was then filtered to match the rectification and smoothing processes used in the formation of a conventional Bmode image. The resulting B-mode image was attenuated by 0.5 dB/cm MHz and a first order noise approximation was added to give echoes of decreasing signalto-noise ratio (SNR) with beam penetration. To duplicate the effects of scanner gain, TGC and compression, the final image was multiplied by user-adjustable gain and TGC controls and compressed logarithmically. The virtual scanner and an example image are shown in Fig. 6.
213
Fig. 6. Virtual US scanner showing a simulated horizontal vessel of 4 mm diameter, calibration markers and scanner gain and TGC controls.
Imaging system A fully integrated application software package VIA (vascular image analysis) was written for the project, together with the virtual US scanner using National Instruments LabVIEW and IMAQ (Austin, Texas, USA). The neural networks were developed using NeuralWorks Professional II/PLUS (Pittsburgh, PA, USA). The system was designed to analyse online or computerstored images for development purposes. The system operated online at 25 images s⫺1 on a 400-MHz PC, with ROI width of typically 2 to 3 cm. An Acuson 128 XP/10 US scanner (Calif., USA) was initially used for clinical imaging, followed by an ATL HDI 3000 scanner (Eindhoven, Netherlands). 7-MHz linear array transducers were used in both cases and the scanner image outputs were connected to a National Instruments PCI-1408 frame-grabber optimised accordingly for the dynamic range. The Acuson 128 scanner was configured for vascular imaging with transmit power of ⫺9 dB, log compression 40 dB, persistence A, preprocessing curve 2 (crisp borders), postprocessing curve 5 (high contrast), overall gain 3 dB. The HDI 3000 scanner was configured for linear map 1, 35 dB gain and persistence off. The tourniquet deflation point in a FMD study represents a period of relatively high workload for the operator. For example, it is necessary to deflate the tourniquet, insert a data-logging marker, monitor image quality and, possibly, refine the transducer position, etc. To minimise this workload, the study timing and marker insertion were managed by the system, with audio/visual operator prompts at relevant points. It was possible, therefore, to handle image-scale alterations automatically when changing to Doppler mode for flow validation immediately following tourniquet deflation.
214
Ultrasound in Medicine and Biology
Volume 28, Number 2, 2002
Table 1. Measurement error as a function of vessel diameter ⫽ ⫺1.16% ⫾ 1.04% (mean ⫾ SD) Expected mm 2.00 3.00 4.00 5.00 6.00 7.00 8.00 Measured mm 1.98 2.97 3.90 5.00 6.01 6.90 7.81 Error % ⫺1.00 ⫺1.00 ⫺2.50 0.00 0.16 ⫺1.43 ⫺2.37
SYSTEM VALIDATION Measurement error The vessel absolute diameter measurement error of the technique was assessed (Table 1) using a physical model. This consisted of two parallel 0.25-mm diameter nylon monofilaments mounted under tension on vernier calipers so that the distance between the filaments was known. The filaments were immersed in a water/glycerol (9% volume) mixture with an US propagation velocity of 1540 m s⫺1, and imaged using an ATL HDI 3000 scanner (Fig. 7) at a zoom level of 18.5 pixels/mm. The distance between the filaments was varied from 2 to 8 mm in 1-mm steps and measured with the wall-tracking software after adjusting the scanner controls to minimise changes in appearance of the filaments caused by spatial nonuniformity of the US beam. Recovery of diameter changes over the cardiac cycle Recovery was evaluated using a simulated cardiac image sequence generated with the virtual US scanner and representing about 10 min of data at 25 images per s. Movement of the vessel walls and the effects of noise can be precisely and repeatably modelled using this method. The simulation was based on a vessel of 4 mm nominal diameter, with parallel horizontal walls. The diameter was modulated by the cardiac cycle shown in Fig. 5 (dotted line), giving a maximum change of 0.63 pixel over the cardiac cycle. This modulation cycle was recovered from approximately 2 min of data before the “cuff inflate” cursor of the FMD study shown in Fig. 4.
Fig. 8. Correlation between original and recovered cardiac cycles as a function of the number of cardiac events averaged. Events from approximately 3 to 10 min of data were required to give reasonable correlation. In clinical practice, 1 to 3 min of baseline data at 25 images/s was generally adequate.
In addition to the cardiac cycle modulation, a random vertical movement of the vessel in the range 0 to 0.2 mm (0 to 2.52 pixels) was added to each image to simulate spatial noise. Wall brightness was also modulated to simulate the slight variation seen over the cardiac cycle in clinical studies. This combination of parallel, perfectly horizontal walls with less than one pixel movement throughout the cardiac cycle represents an extreme case, likely to be experienced rarely in practice. Sequence lengths of approximately 15,000 images were generated and vessel diameter measured throughout each sequence using the neural network wall detectors. Correlation between actual and recovered events as a function of the number of cardiac events (Fig. 8) was explored to assess the clinical requirements for adequate recovery. Calculation of SNR provided an index of the quality of recovered cardiac cycle changes. DISCUSSION
Fig. 7. Example monofilament wire image used for assessing measurement accuracy.
This technique has several advantages over established practice. FMD results are available immediately. Data for the complete study are available, rather than just a few salient events, and changes in vessel diameter over the cardiac cycle are recovered. Vessel wall tracking quality can be optimised by adjustment of the scanner controls before the study proper. This largely normalises the intensity range of the images to the range presented to the neural networks in the training regimen, thereby optimising the images to suit the networks. With retrospective analysis, however, images are subjectively optimised for visual, rather than wall detection properties and the detection algorithm must, therefore, handle a wider range of image quality. Online analysis provides
Online artery diameter ● V. R. NEWEY and D. K. NASSIRI
continuous graphical presentations of vessel diameter, detected walls and wall probability levels as tracking quality indicators. These features, some of which are only available retrospectively with off-line techniques, make minor changes in image and tracking quality and vessel alignment readily apparent at a time when they can be rectified. The method also avoids the image quality loss associated with VCR recording, and no processing of the image is necessary. Retrospective systems, however, often require processing steps, such as intensity or rotational normalisation or smoothing which, although improving the tracking qualities, may cause some spatial degradation of the image. There is no unwieldy image archive; file size for a complete study is typically approximately 0.3 Mb and the study is completed much faster than retrospective methods, contributing to lower study costs. Some of these factors probably contribute to the low measurement variance of 0.9% ⫾0.6% (mean ⫾ SD) and coefficient of variance 0.66% found in a clinical trial (Sidhu et al. 2001). The variance compares favourably with the findings of Preik et al. (2000) of 1.3% ⫾ 0.9% and 0.78%, respectively, when applying the retrospective graph-search dynamic programming system of Liang et al. (1998). It should be noted that these figures include biologic dilation variability, as well as measurement variation due to the technique. Although online analysis demands vigilance at critical study points, such as tourniquet deflation, this is balanced by the automated study timing. Further, this allows the system automatically to select baseline and dilation data segments for final analysis. After case completion, the operator is generally only required to select the data file and to press a button to calculate dilation, a significant reduction in analytical workload. The measurement error of ⫺1.16% ⫾ 1.04% (mean ⫾ SD) in Table 1 is probably attributable to the US echo axial size (Li et al. 1993). This will cause an underestimation of the distance between filaments in the physical model. In practice, a secondary source of error is the network training method of pointing to the edge of the brightest echo at the vessel wall/blood interface when selecting the “wall” training category. Although conferring more robust tracking, these points correspond to the media interface. The intima interface was ignored and the technique, therefore, overestimates vessel diameter because the media-intima distance represents the wall thickness. However, these error sources should not be significant in FMD studies, where relative rather than absolute change in diameter is of importance. As with all FMD studies, good image quality is mandatory for both accurate measurements and optimum wall tracking. The technique can also be used to detect nonhorizontal vessel interfaces by searching the ROI sequentially with network input line sections at more than one orientation
215
(i.e., vertical and horizontal) and combining the outputs to obtain the probability matrix. The simulated cardiac sequence required 7 to 8 min of data to recover diameter changes over the cardiac cycle with reasonable quality (Fig. 5). In clinical practice, however, the cardiac cycle is generally recovered with good quality from 1 to 3 min of data. Recovery of diameter changes over the cardiac cycle may be useful in other studies, such as investigation of the elastic properties of the vessel wall, where wall displacement combined with blood pressure can be used to calculate elasticity indices. The virtual US scanner technique can be applied in other areas of US image analysis (e.g., in B-mode image quality assurance) (Newey and Nassiri 2000). The wall-tracking technique has proven to be reliable under a range of operating conditions. The training, simulation and clinical studies all used images from different types of scanners with a range of zoom settings (9 to 32 pixels/mm). In addition, the training images were of the carotid artery, and clinical FMD studies have used the brachial artery. To date, over 120 clinical cases in continuing brachial FMD and carotid elasticity studies (to be reported) have been analysed at St. George’s Hospital Medical School using an ATL HDI 3000 scanner. In all cases, the correct vessel walls were automatically located in the ROI. Manual intervention was required rarely when image quality deteriorated to a point where walls disappeared and reappeared at a substantially different location within or outside the ROI. The design aim was the development of fast and, therefore, simple neural networks to attain online performance. As faster PCs become available, more complex neural networks trained, for example, with additional contextual information, could be used to further enhance the robustness of the method. Acknowledgements—The authors gratefully thank J. S. Sidhu and V. Mwansa of St. George’s Hospital Medical School, Department of Cardiological Sciences, for suggesting operational refinements to the system and clinical support. They also thank N. Parchure and G. Rosso for their clinical contributions.
REFERENCES Ahmad S, Tesauro G. Scaling and generalization in neural networks: A case study. Proc Neural Info Proc Syst Conf (NIPS) 1988;1:160 – 176. Celermajer DS, Sorensen KE, Gooch V.M., et al. Non-invasive detection of endothelial dysfunction in children and adults at risk of atherosclerosis. Lancet 1992;340:1111–1115. Detmer P, Bashein G, Martin R. Matched filter identification of leftventricle endocardial borders in transesophageal echocardiograms. IEEE Trans Med Imaging 1990;9:396 – 404. Fu LM. Neural networks in computer intelligence. New York: McGraw-Hill, 1994. Gevins AS, Stone R., Ragsdale SD. Differentiating the effects of three benzodiazapines on non-REM sleep EEG spectra: A neural network pattern classification analysis. Neuropsychobiol 1988;19(2):108 – 115.
216
Ultrasound in Medicine and Biology
Gonzalez RC, Wintz P. Digital image processing. Addison Wesley, 1987. Gustavsson T, Gharbieh A, Hamarneh G, Liang Q. Implementation and comparison of four different boundary detection algorithms for quantitative ultrasonic measurements of the human carotid artery. Comput Cardiol 1997;24:69 –72. Gustavsson T, Liang Q, Wendelhag I, Wikstrand J. A dynamic programming procedure for automated ultrasonic measurement of the carotid artery. Comput Cardiol 1994;297–300. Li S, McDicken WN, Hoskin PR. Blood vessel diameter measurement by ultrasound. Physiol Meas 1993;14:291–297. Liang W, Browning R, Lauer R, Sonka M. Automated analysis of brachial ultrasound time series. Multi-dimensional images. SPIE 1998;3337:108 –118. Mehrotra KG, Chilukuri KM, Ranka S. Bounds on the number of samples needed for neural learning. IEEE Trans Neural Networks 1991;2:548 –558. Nassiri DK, Nicholas D, Hill CR. B-scan texture classification: A study using physical and theoretical models. In: Lerski RA, Morley P, eds. Third meeting of the World Federation for Ultrasound in Medicine & Biology. Oxford: Pergamon Press, 1983:133–139. Newey VR, Nassiri DK. Analysis of variability in visual B-mode image
Volume 28, Number 2, 2002 evaluation; Application of a virtual scanner and phantom. Eur J Ultrasound 2000;13:1S. Preik M, Lauer T, Hei S, et al. Automated ultrasonic measurement of human arteries for the determination of endothelial function. Ultraschall Med 2000;21:195–198. Rumelhart DE, Hinton GE, Williams RJ. Learning internal representations by error propagation. In: Rumelhart DE, McClelland JL, eds. Parallel distributed processing: Exploration in the microstructure of cognition. Vol. 1. Cambridge: MIT Press, 1986. Selzer RH, Hodis H, Kwong-Fu H, et al. Evaluation of computerized edge tracking for quantifying intima-media thickness of the common carotid artery from B-mode ultrasound images. Atherosclerosis 1994;111:1–11. Sidhu JS, Mwansa V, Newey VR, Nassiri DK, Kaski JC. The development of an on-line, automated technique to determine endothelial function. Circulation 2001;104(17):supplement II, abstract 1161. Sorensen KE, Celermajer DS, Spiegelhalter DJ, et al. Non-invasive measurement of human endothelium dependent arterial responses: Accuracy and reproducibility. Br Heart J 1995;74:247–253. Tu JV. Advantages and disadvantages of using artificial neural networks versus logistic regression for predicting medical outcomes. J Clin Epidemiol 1996;49:1225–1231.