An Automated Method for Lumen and Media ... - Semantic Scholar

IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 8, NO. 2, JUNE 2004

131

An Automated Method for Lumen and Media–Adventitia Border Detection in a Sequence of IVUS Frames Marina E. Plissiti, Dimitrios I. Fotiadis, Member, IEEE, Lampros K. Michalis, and George E. Bozios

Abstract—In this paper, we present a method for the automated detection of lumen and media–adventitia border in sequential intravascular ultrasound (IVUS) frames. The method is based on the use of deformable models. The energy function is appropriately modified and minimized using a Hopfield neural network. Proper modifications in the definition of the bias of the neurons have been introduced to incorporate image characteristics. A simulated annealing scheme is included to ensure convergence at a global minimum. The method overcomes distortions in the expected image pattern, due to the presence of calcium, employing a specialized structure of the neural network and boundary correction schemas which are based on a priori knowledge about the vessel geometry. The proposed method is evaluated using sequences of IVUS frames from 18 arterial segments, some of them indicating calcified regions. The obtained results demonstrate that our method is statistically accurate, reproducible, and capable to identify the regions of interest in sequences of IVUS frames. Index Terms—Deformable models, image segmentation, intravascular ultrasound (IVUS).

I. INTRODUCTION THEROSCLEROSIS causes partial or total obstruction of human arteries. Early diagnosis and accurate assessment of plaque position and volume are essential for the selection of the appropriate treatment. Several imaging techniques exist for the estimation of the severity of the disease in vivo. Intravascular ultrasound (IVUS) is a commonly used diagnostic tool, which provides real-time visualization of plaque morphology, detection of typical plaque components, such as calcium, and quantification of plaque eccentricity and wall thickness. IVUS image sequences consist of cross-sectional images of the arterial segment and are acquired with the insertion of a

Manuscript received May 28, 2003; revised January 14, 2004. This work was supported in part by the Greek General Secretariat for Research and Technology (PENED 318—An Intelligent System for the Early Diagnosis of Coronary Artery Disease). M. E. Plissiti is with the Department of Computer Science, Unit of Medical Technology and Intelligent Information Systems, University of Ioannina, GR 45110 Ioannina, Greece (e-mail: [email protected]). D. I. Fotiadis is with the Department of Computer Science, Unit of Medical Technology and Intelligent Information Systems, University of Ioannina and Michailideion Cardiology Center, GR 45110 Ioannina, Greece, and is also with the Biomedical Research Institute, FORTH, GR 45110 Ioannina, Greece (e-mail: [email protected]). L. K. Michalis is with the Department of Cardiology, Medical School, University of Ioannina, GR 45110 Ioannina, Greece, and is also with the Michailideion Cardiology Center, GR 45110 Ioannina, Greece (e-mail: [email protected]). G. E. Bozios is with Medical Physics Laboratory, Medical School, University of Ioannina, GR 45110 Ioannina, Greece (e-mail: [email protected]). Digital Object Identifier 10.1109/TITB.2004.828889

catheter in the vessel. A small transducer which is placed at the tip of the catheter transmits a high-frequency (20–40 MHz) ultrasound signal. The reflection of the ultrasound beam as it passes through the different layers and the scattering of the material, in combination with the constant speed of the catheter’s pullback, produce an image sequence, which is recorded in video format. The video is analyzed by the cardiologists to estimate the severity of the disease and to decide about the patient treatment. Manual processing of IVUS images is a tedious and timeconsuming procedure. A lot of effort has been made in order to develop an accurate automated method for the detection of the regions of interest in IVUS images. The proposed methodologies usually take advantage of the characteristic appearance of the arterial anatomy in two-dimensional IVUS images and the connectivity of frames in the entire IVUS sequence. Several segmentation methods have been proposed. Some of the earlier work on segmentation of IVUS images was based on heuristic graph searching algorithms using a cost function, in which a priori information of the expected pattern in IVUS frames was incorporated [1]–[3]. Segmentation methods based on probabilistic approaches have also been proposed [4], [5]. A class of methods, based on the expected similarity of the regions of interest in adjacent IVUS frames, takes into account that the sequence of frames constitute a three-dimensional object. Under this perspective, active contour principles [6], [7] can be used to extract the desired lumen and media–adventitia borders. Other methods are based on a combination of transversal and longitudinal contour detection techniques [8]–[10]. Many restrictions in automated segmentation of IVUS images derive from the quality of the image, such as the lack of homogeneity of regions of interest and shadowed regions, which are produced by the presence of calcium. The complicated structure of human vessels and the different components each part consists of, result in an image with high intensity variation, even in regions corresponding to the same tissue. In addition, calcified hard plaque regions are typically identified by high-amplitude echo signals with complete distal shadowing. As a result, the morphology of the outer layers of the arterial segment is not possible to be identified. However, an estimation of the media–adventitia border is feasible, when information from the previous frame is used. In this work, we focus on the development and validation of an automated method based on deformable models for accurate IVUS image segmentation. By exploiting the similarity of sequential frames, minimum user interaction is required only for

1089-7771/04$20.00 © 2004 IEEE

132


the first frame of the sequence, where the user must provide an initial estimation for the lumen and the media–adventitia borders. The method can be applied automatically on a sequence of IVUS frames for the extraction of the borders of the regions of interest. The modification of the energy function of the deformable model makes the use of a Hopfield neural network feasible for the minimization of the energy function. The structure of the neural network incorporates knowledge of the image pattern. A simulated annealing (SA) scheme ensures that the minimization procedure reaches the global minimum. The introduction of a new expression for the image energy makes the method robust, resulting in accurate boundaries for all images, regardless of the presence of noise or weak edges, both common in IVUS images. The proposed method is validated using sequences of IVUS frames from 18 patients. II. AUTOMATIC SEGMENTATION OF SEQUENCES OF IVUS IMAGES Each IVUS frame in a sequence of frames obtained from the same arterial segment can be considered quite similar to the previous one. The detection of lumen and media–adventitia borders is an optimization problem and an initial estimation of the desired borders is needed only for the first frame. The minimization of a deformable model energy function must be applied twice for each frame (once for each border) and the detected contours in the current frame are used as the initial estimation for the next frame. Each frame is preprocessed and the border detection follows. A. Preprocessing IVUS frames contain noise and the actual boundaries of regions of interest are difficult to be identified in many cases. Image preprocessing removes speckles and artifacts that can interfere with the detection of desired boundaries. Furthermore, the detection of regions of interest is restricted by the existence of weak edges in IVUS images and image enhancement is required. We use a 3 3 median filter to reduce the effect of speckles, without blurring the edges. A bandpass linear filter is then applied permitting pixel intensities in the range 20–230 to pass. A transfer function is then used for linear mapping of the pixels intensity in 255 levels, which results in the sharpening of the image edges. B. Borders Detection 1) The Deformable Model: Since our approach is based on deformable models, an initial contour for the lumen and media–adventitia border must be provided for each IVUS frame. For the first frame, this is given interactively by the user and a closed curve is provided. This is the only point where the user interacts with the developed system and constitutes the initialization process of our method. For the subsequent frames, the borders extracted in the previous frame are used as initial estimation. The initial estimation forms an active curve (snake), which deforms in order to obtain the final shape of the border. A snake deforms under the influence of internal and external forces [11]. The position of the snake can be represented by the curve

, where is the arc length and are the cartesian coordinates of each point of the curve. The energy of the snake is given as (1) where represents the internal energy of the snake due to bending and is derived from image data. The internal energy can be expressed as (2) where and are the first and second order derivatives, respectively. is the term that forces the snake to be attracted to image features, and is defined as (3) where is the gradient of image . The factors , in (2) and (3) are weights which regulate the contribuand tion of each term. The pixels of the image that minimize this energy function and are close to the region of interest (i.e., they lie in a specific area where the snake can deform) define the boundaries of the desired region. To solve this minimization problem, we employ a Hopfield neural network [12]. A typical Hopfield neural network [13] consists of a single layer of neurons, where each neuron has one of the two outputs, or (firing or not firing). A Hopfield network is fully interconnected with no specific input or output layer. Each node has a bias and is connected with every other node. The connections are bidirectional and symmetric and a specific is assigned to each connection. The state of the weight th neuron depends on the input it receives from other neurons, and is given as (4)

where is the number of neurons. The energy function, which is minimized by the network, is (5) The network converges when the energy function reaches a local minimum. The nodes of the network correspond to image pixels and the pixels that minimize the total energy of the network form the line segments consisted desired boundary. In our approach, points, that are perpendicular to the initial contour, at 15 of intervals, determine the area where the snake can deform. The neurons (Fig. 1), and network consists of one layer of each neuron represents a candidate point of the final boundary. The output of the neuron is zero, for a point that is not included in the set of the boundary points, and one, for a point that belongs to the boundary set.

PLISSITI et al.: AUTOMATED METHOD FOR LUMEN AND MEDIA–ADVENTITIA BORDER DETECTION

Fig. 1.

133

Neural network architecture.

The energy of the snake, which is minimized by the Hopfield neural network, can be expressed as [12]

(7) where if otherwise.

(6) are the and coordinates of the th point of the where th line segment and is the output of the neuron. It can be proven [12] that the interconnective strengths are given as

(8)

The bias for each neuron can be determined using a priori knowledge about the pattern of IVUS images and the information provided by the image. Nodes having high probability to be a boundary point are set to have high bias value while nodes with low probability are set to have low bias. In our method, we seek for points that have high image gradient. If the gradient of the image at a pixel is greater than a specific threshold then the pixel is considered to be a candidate boundary point. Neurons of the neural network that correspond to pixels with low gradient have low probability to be members of the desired boundary, and for this reason they have low bias value. If there is no neuron in a row of the neural network with high image gradient, then the middle neuron is selected as the winner. Those rows of the neural network are further analyzed, as it is explained later, because they indicate calcified regions when we are looking for the media–adventitia border. We attribute low bias value to nodes lying in specific areas of the image, because the expected boundary is not possible to lie in those areas. For setting the bias value of each neuron, we consider the position of the neuron with respect to its distance from the ultrasound catheter. The dimension of the catheter, as well as its position in each frame, is known. If a neuron lies in the area of the catheter, or its distance from the catheter is small (the point corresponds to blood area), we bias this neuron in order not to be chosen as the winner. The same procedure is used

134


for the neurons that lie in the area out of the media–adventitia border, when we search for the lumen border. We represent the set of pixels belonging to these areas with . We incorporate the above knowledge in the bias of each neuron if AND OR otherwise where is the image gradient, is an empirical threshold, is the number of neurons in a row of the neural network, and is a low value for the bias, in order to exclude that neuron (we and ). have used , in each iteration, is updated The state of the neuron according to (9) The output of the neuron lution function

is determined by a maximum evo(10)

where if otherwise. (11) Using (4)–(11), we can initialize the strengths for each connection and the state, output, and bias for each neuron, and the total energy of the network. In the first iteration, the output of all neurons is zero, except those neurons that correspond to points of the initial estimation contour. To prevent entrapment of the energy function in local minima, an SA scheme is employed. After iteration steps, the network changes its state from energy to energy with probability if if

(12)

where is a control parameter called temperature. The algorithm starts with a high temperature, where most changes are accepted, and progressively decreases so that changes corresponding to a positive variation of the energy become gradually rare. The value of initial temperis estimated such as at the beginning of SA, the value ature is greater than 0.9. Setting the of acceptance probability initial value as (where is the initial energy of . The cooling schedule is the network), this is true if . The search is termidescribed by the relation nated after a predefined number of iterations (in our case, five iterations) at a temperature no improvement of the energy func. The effect of the SA scheme in the tion occurs border detection process is illustrated in Fig. 2. The iterative procedure for the minimization of the energy function is summarized as follows: 1) Update the state of each neuron synchronously. 2) If the new state of the neuron reduces the total energy of the network, then the new state is acceptable and the states of all neurons and the total energy of the network

Fig. 2. Effect of the SA scheme in media–adventitia border detection. (a) The borders extracted with and (b) without the use of SA scheme.

are updated, otherwise, the states of all neurons change accordingly with probability . 3) Check the total energy of the network. If the energy does for more than five not change anymore iterations, the network has reached the global minimum. 4) The firing neurons form the new boundary for the region of interest. is included 2) Image Energy: The gradient of the image (3). For the detection of the regions in the expression for of interest, the snake must be attracted to points which correspond to large image gradients. The gradient of the image can be computed using standard operators. However, the applicability of those operators is limited in noisy environments such as IVUS images. In these images, the intensity of the pixels corresponding to specific tissues varies in a narrow range. Pixels with large image gradient may appear due to noise speckles. These pixels do not correspond to the border of the region of interest and they attract the snake toward their position, which may entail wrong identification of the border. To avoid such problems, we compute the gradient at a specific pixel using the mean intensity value of four adjacent square areas (windows) of fixed size (Fig. 3). Thus, for each pixel of the snake, the image gradient is defined as

(13) where , and represent the mean intensity values of the upper, lower, left, and right , respectively. They are computed as windows of the pixel (14)

where is the size of the window and is the pixel at the center of the window. The size of the window affects the performance of the proposed method. Fig. 4 illustrates the effect of the window size in the detection of lumen and media–adventitia borders. The optimal boundaries were obtained with the use of a 9 9 window. This demonstrates that the use of (13) in the computation of the image gradient reduces significantly the effect of noise.


135

Fig. 3. Image gradient computation at a specific pixel of the image.

Fig. 4. (a) Lumen and (b) media–adventitia borders for different window sizes used in the computation of the image gradient.

3) Border Detection in Calcified Regions: Calcium deposits in IVUS images are easily recognized by the presence of a bright echo and acoustic shadowing of the underlying structures. The presence of calcified lesion prohibits the extraction of media–adventitia border and doctors assess the largest arc of the lesion in order to estimate its severity. Extrapolation of the circumference of the vessel area, based on the largest arc of calcium, is usually applied for the estimation of the media–adventitia border [14]. This empirical approach has been incorporated in our method.

In the case that the neural network line segments lie totally in shadowed regions, there is no candidate neuron to be a winner and after convergence of the neural network, the winner neuron at the specific row is the neuron which corresponds to the contour of the previous estimation. These rows of the neural network indicate the presence of calcified regions or, in general, regions with noise artifacts. To overcome this problem, we have modified appropriately our method.

136


Fig. 5. IVUS image with calcified lesion or artifacts. (a) The boundary between media–adventitia is ambiguous in the region between the arrows A and B due to acoustic shadowing or the presence of a guide wire in the image. (b) Detected borders.

1) The center of the contour is calculated using the winner neuron of each row of the neural network and is given by

Fig. 6. Detected lumen and media–adventitia borders.

we look for smooth initial estimations for the media–adventitia border. Thus, the convex hull of the extracted border points is calculated using the plane sweep approach [15]. III. METHOD OF VALIDATION A. Study Group

(15) are the and coordinates of the winner where neuron in the th row of the neural network. 2) For every winner, if it lies in a calcified region, we search for the pair of neurons at the beginning and the end of the shadowed region. 3) Those neurons define an arc of the circle with center the contour center and radius the equivalent radius. The points of this arc are incorporated into the final contour. The first step is necessary in order to find the relative border position in the frame in the presence of catheter shift. In the next steps, calcified lesions are recognized and finally, the arc of the lesion is constructed. The same procedure can be employed for artifacts produced by the ultrasound catheter (presence of a guide wire). The outcome of the proposed approach for calcified region and for the presence of a guide wire in the image is shown in Fig. 5(a) and (b), correspondingly. 4) Final Contour Extraction: The first application of the method in each frame results in the detection of the media–adventitia border, while the second in the detection of the lumen border (Fig. 6). It is found that small distortions in the shape of the final contour may cause abnormalities in the specification of the searching area for the next frame, which might result in false detection of the desired borders. The error introduced by nonconvex initial estimation in the detection of the media–adventitia border propagates along the application of the algorithm in subsequent frames (Fig. 7). For this reason,

We analyzed sequences of IVUS frames from 18 arterial segments, with an average length of 39 mm (28–53 mm). IVUS images were acquired using a 30-MHz catheter at a constant pullback speed of 0.5 mm/s (CVIS—Boston Scientific IVUS System). IVUS images were recorded on S-VHS video tapes using fixed image size. Each sequence contains 1408–2650 frames and are digitized at 640 480 pixels (0.03 mm/pixel), using a digitization rate 25 frames/s. B. Methods The accuracy of the proposed method can ideally be determined comparing the borders extracted with the real borders. However, the location of real borders is not known. Therefore, the only way to make the validation is to compare the extracted borders with the borders identified by expert observers. For this reason, two expert observers traced twice (with a month time period difference between the first and second tracing) the borders in manually selected frames corresponding to the R-wave of the simultaneously recorded electrocardiogram (ECG) signal. The number of the frames traced by the experts is 1432. The proposed method, however, has been applied to the entire sequence of IVUS frames for each arterial segment and borders of the regions of interest were obtained automatically for each frame. The results of the method for the frames corresponding to those manually selected were used for the validation. The reliability and reproducibility of manual tracing was assessed by estimating interobserver and intraobserver variability for the area of the regions of interest. The proposed automated method was validated using several approaches.


Fig. 7.

137

Extraction of the media–adventitia border (a) with the computation of the convex hull in sequential frames and (b) without computation of the convex hull.

1) The automated method was considered as an independent observer and the interobserver variability between manual tracing (average) and automated border detection was computed [6]. 2) Linear regression analysis was also used [16]. We have computed correlation coefficient, slope, and y-interception to compare volumes estimated from the proposed method and the average of the two experts. The above have been applied in order to compare computer and manual tracing results for the following: 1) area in IVUS images with shadowed regions; 2) volumes defined by the lumen and media–adventitia borders; 3) area and volume for different initial estimations of the proposed method. The comparison for area defined by the media–adventitia borders in frames with shadowed regions has been performed for 60 frames, having calcified regions or extensive noise artifacts, selected by the experts. The criterion for the selection was its difficulty in the determination of the outer boundary of the vessel. The volumes defined by the lumen and the media–adventitia borders are calculated using Simpson’s rule (16) where is the calculated area in a frame, is the distance between successive frames, and is the number of frames. In our experiments, the distance between successive R-top corresponding frames is approximately 0.5 mm. The reproducibility of the proposed method with respect to the initial estimation was also evaluated. Three expert observers provided their initial estimation twice (totally six different initial estimations for the lumen and media–adventitia border) for the first frame in a sequence of 2050 frames (equivalent to 41-mm

length), which correspond to one arterial segment. For the comparison of the results of the method with expert observers tracings, 82 frames were selected from this sequence, which correspond to the R-top of the ECG signal. We have computed the volume defined by the lumen and the media–adventitia borders and we compared the results with the average of the two experts in the same frames. IV. RESULTS The proposed method is fully automated, even in the presence of calcified regions, and an initial estimation is needed for the borders in the first frame of a sequence of IVUS frames. The method is fast enough since the time for the processing of each frame is 0.6 s and for an entire sequence (1408–2650 frames) the time ranges between 14 and 26.5 min using a Pentium IV 1.4 GHz with 512-MB random access memory (RAM). Intraobserver and interobserver variability of the two experts for media–adventitia area in images with shadowed regions have been computed, resulting in intraobserver vari% % and the second expert ability for the first % % and interobserver variability % %. All 60 images with calcified lesions were processed for the detection of media–adventitia border. Computer-identified media–adventitia areas correlated very well with observer areas resulting in % % mean SD . The results are shown variability graphically in Fig. 8. Intraobserver and interobserver variability of the two experts for lumen and media–adventitia areas have also been computed for the entire set of manually processed IVUS frames. Intraobserver variability for the first expert was for the lumen % % and for media–adventitia % %. % % For the second expert the results were % %, respectively. Results for interobserver and % % for the lumen area and variability were

138

Fig. 8.


Comparison of detected and observer-defined area in images with calcified regions.

% % for media–adventitia area. The computed correlation coefficients between the two observers are 0.958 for lumen area and 0.900 for media–adventitia area. Linear regression analysis for volumes of the arterial segments resulted in high correlation coefficients for the media–adventitia [ , Fig. 9(a)] and , Fig. 9(b)], at the lumen [ the same significant level . For the media–adventitia, the mean difference between computed and manual volume was 2.1% and for the lumen 4.9%. We have also carried out a two-tailed t-test [17], demonstrating that a significant correlaexists. tion To evaluate the reproducibility of the method, six different initial estimations for the lumen and media–adventitia border were used for an arterial segment. The results are shown in Table I, where the initial estimations are considered ellipses, which are described by the minor and major axes. The repeatability coefficient used requires that 95% of the differences in the method’s results are less than two standard deviations [16]. V. DISCUSSION We have proposed an automated method, based on deformable models, for the detection of lumen and media–adventitia borders in IVUS images of arterial segments. An estimation of the borders in the first frame is necessary before the application of the method in sequential frames. The method uses as an initial estimation of the borders of the current frame the detected borders of the previous frame (or frames). This accelerates our method and makes it automatic even in the case of the presence of calcified regions, since we have modified it to overcome such problems. This is a consequence of the confined searching space of the neural network and because we seek for convex, ellipsoid boundaries. However, even our

method is superior compared to other methods in this sense; the problem of side branches has not been resolved. We have introduced a modification in the computation of the image energy, since the use of the conventional definition leads to highly distorted detected boundaries caused by the presence of noise in IVUS frames. This results in the selection of pixels that belong to edges separating large and homogenous regions in the image. This is due to the fact that the areas of noisy pixels are small and they do not greatly influence the mean value of the intensity in large areas. The method overcomes the problem in the computation of the media–adventitia border when calcified lesions are present in IVUS frames. At the presence of calcium deposits, segmentation algorithms that are based on the detection of large values of image gradient are likely to fail, due to acoustic shadowing. Our method recognizes calcified regions and incorporates knowledge from previous frames to find the best estimation of the borders at the current frame. In medical practice, the detection of the borders of interest, when it is performed manually, is an approximation of the real borders, and it is based on former images which have clear borders. The computation of the largest arc of the calcified region is acceptable in conventional IVUS analysis by experienced doctors, because in this way, the severity of calcified lesion is estimated. The use of the Hopfield neural network for the minimization of the snake’s energy makes the application of the method faster and computationally more efficient because the uniform separation of the searching domain in similar areas reduces the searching space and takes advantage of the information contained in the entire neighborhood of the region of interest. The calculation of the convex hull produces smooth curves and ensures that the searching lines, which are formed in the next frame, will not coincide or will not be restricted in a small area, resulting in the loss of significant image characteristics.


139

Fig. 9. Comparison of detected and observer-defined volume for (a) media–adventitia and (b) lumen.

Several parameters, which enter the computation of the energy of the snake must be defined. In our approach, we have , addressed the determination of the values of the weights and . Changing the values of the weights the contribution of each term will be altered. The snake must be attracted mainly from the characteristics of the image and have a smooth shape. We have carried out many experiments and the values that yield , and acceptable results for both borders were . The number of neurons of the neural network depends on the resolution of the digitized IVUS frames. This requires some testing before applying the proposed method. In our case, the

, total number of the sampled points at each contour is which is sufficient in order to reach a smooth final contour. The number of points included in each segment is different for the for the lumen and for inner and outer boundary ( the media–adventitia border). As a result, if the initial estimation lies into the searching area, the deformable model equilibrates at the desired contour. We also performed a validation of the proposed method using as golden standard the outcome of the manual drawing of the lumen and media–adventitia border by two expert observers. Those traced twice the borders in selected frames from 18 arterial segments and the obtained contours were compared with

140


TABLE I VARIATION OF THE AVERAGE AREA ENCLOSED BY THE LUMEN AND MEDIA–ADVENTITIA BORDER AND VOLUME OF AN ARTERIAL SEGMENT OBTAINED FOR DIFFERENT INITIAL ESTIMATION. THE OBSERVER-DEFINED AREA AND VOLUME ARE ALSO INCLUDED. RESULTS ARE PRESENTED AS MEAN SD

6

the ones obtained by the application of our method. We have used area and volume to estimate interobserver variability applying direct comparison and linear regression analysis. The reliability of the two observers is assessed using interobserver and intraobserver variability. Intraobserver variability ranges from 0.5% to 5.4% for area and interobserver variability from 0.1% to 5.0%. The above values compare well with others reported in the literature [6], [18]. Linear regression analysis indicates that the method is accurate even in the presence of calcified lesions. The slopes are close to unity, the y-interception confidence interval includes always zero, and the correlation coefficient is higher than 0.985. in the case of The correlation coefficient is lower calcified lesions, but this is also acceptable. VI. CONCLUSION IVUS permits direct visualization of the morphology and the geometry of coronary arterial segments. We have developed an automated method, which can be used for a sequence of IVUS frames to detect lumen and media–adventitia borders. The method is based on the use of deformable models and optimization of the energy function using a Hopfield neural network. We have included an SA scheme to avoid local minima and provide a global minimum. The proposed method is advantageous compared with existing techniques since problems related to calcified regions, image artifacts, and noise have been addressed. In addition, the method is fully automated and an initial estimation of the borders is needed for the first frame only. A correction for the detected border schema has been introduced to provide better initial estimations for subsequent frames in order to improve efficiency and speed. We have proved also that the method is reproducible and not sensitive to the initial estimation. To validate the method, we compared its results with the manual estimation of borders by two expert observers for 18 arterial segments. We have observed small variations between manual and automated detection of borders using several metrics. The task of identifying the regions of interest in IVUS frames is a challenging issue, especially when this must be done in a sequence of frames automatically. Several methods have been developed and several attempts to examine their reliability have been presented. The validation results for our method underline its high performance and prove that it is one of the most efficient and accurate methods for IVUS segmentation. This method must be incorporated into a three-dimensional reconstruction method to finally have a complete visualization of the arterial segment.

ACKNOWLEDGMENT The authors would like to thank Prof. I. E. Lagaris and Prof. A. Likas for their suggestions and valuable comments and S. Psaltis for his help in the statistical analysis of data. REFERENCES [1] M. Sonka, X. Zhang, M. Siebes, M. S. Bissing, S. C. DeJong, S. M. Collins, and C. R. McKay, “Segmentation of intravascular ultrasound images: A knowledge-based approach,” IEEE Trans. Med. Imag., vol. 14, pp. 719–732, Dec. 1995. [2] X. Zhang, C. R. McKay, and M. Sonka, “Tissue characterization in intravascular ultrasound images,” IEEE Trans. Med. Imag., vol. 17, pp. 889–898, Dec. 1998. [3] A. Takagi, K. Hibi, X. Zhang, T. J. Teo, H. N. Bonneau, P. G. Yock, and P. J. Fitzgerald, “Automated contour detection for high-frequency intravascular ultrasound imaging: A technique with blood noise reduction for edge enhancement,” Ultrasound Med. Biol., vol. 26, no. 6, pp. 1033–1041, 2000. [4] D. Gil, P. Radeva, J. Saludes, and J. Mauri, “Automatic segmentation of artery wall in coronary IVUS images: A probabilistic approach,” in Proc. CIC 2000, Cambridge, MA, 2000, pp. 4352–4355. [5] C. Haas, H. Ermert, S. Holt, P. Grewe, A. Machraoui, and J. Barmeyer, “Segmentation of 3D intravascular ultrasonic images based on a random field model,” Ultrasound Med. Biol., vol. 26, no. 2, pp. 297–306, 2000. [6] G. Kovalski, R. Beyar, R. Shofti, and H. Azhari, “Three-dimensional automatic quantitative analysis of intravascular ultrasound images,” Ultrasound Med. Biol., vol. 26, no. 4, pp. 527–537, 2000. [7] R. Shekhar, R. M. Cothren, D. G. Vince, S. Chandra, J. D. Thomas, and J. F. Cornhill, “Three-dimensional segmentation of luminal and adventitial borders in serial intravascular ultrasound images,” Comput. Med. Imaging Graph., vol. 23, pp. 299–309, 1999. [8] T. Hagenaars, E. J. Gussenhoven, J. A. Van Essen, J. Seelen, J. Honkoop, and A. Van Der Lugt, “Reproducibility of volumetric quantification in intravascular ultrasound images,” Ultrasound Med. Biol., vol. 26, no. 3, pp. 367–374, 2000. [9] C. Von Birgelen, C. S. Mintz, A. Nicosia, D. P. Foley, and W. J. Van Der Giessen, “Electrocardiogram-gated intravascular ultrasound image acquisition after coronary stent deployment facilitates on-line three dimensional reconstruction and automated lumen quantification,” J. Amer. Coll. Cardiol., vol. 30, no. 2, pp. 436–43, 1997. [10] J. Dijkstra, G. Koning, J. C. Tuinenburg, P. V. Oemrawsingh, C. von Birgelen, and J. H. C. Reiber, “Automatic border detection in intravascular ultrasound images for quantitative measurements of the vessel, lumen, and stent parameters,” Int. Congress Series, vol. 1230, pp. 916–922, 2001. [11] M. Kass, A. Witkin, and D. Terzopoulos, “Snakes: Active contour models,” Int. J. Comput. Vision, vol. 1, pp. 321–331, 1987. [12] Y. Zhu and H. Yan, “Computerized tumor boundary detection using a Hopfield neural network,” IEEE Trans. Med. Imag., vol. 16, pp. 55–67, Feb. 1997. [13] J. J. Hopfield, “Neurons with graded response have collective computational properties like those of two-state neurons,” in Proc. Nat. Acad. Sci., vol. 81, 1984, pp. 3088–3092. [14] T. Hagenaars, E. J. Gussenhoven, E. Van Der Linden, and N. Bom, “Reproducibility of calcified lesion quantification: A longitudinal intravascular ultrasound study,” Ultrasound Med. Biol., vol. 26, no. 7, pp. 1075–1079, 2000. [15] F. P. Preparata and M. I. Shamos, Computational Geometry, An Introduction. New York: Springer-Verlag, 1998, ch. 3.


[16] J. M. Bland and D. G. Altman, “Statistical methods for assessing agreement between two methods of clinical measurements,” Lancet, vol. 1, pp. 307–310, 1986. [17] R. F. Mould, Introductory Medical Statistics, 2nd ed. New York: Adam Hilger, 1989, ch. 11. [18] D. S. Meier, R. M. Cothren, D. G. Vince, and J. F. Cornhill, “Automated morphometry of coronary arteries with digital image analysis of intravascular ultrasound,” Amer. Heart J., vol. 133, pp. 681–690, 1997.

Marina E. Plissiti was born in Ioannina, Greece, in 1977. She received the B.Sc. and the M.Sc. degrees in computer science from the University of Ioannina, Ioannina, Greece, in 1998 and 2001, respectively. She is currently working toward the Ph.D. degree in computer science at the University of Ioannina. Her research interests include medical image processing and artificial intelligence in biomedical applications.

Dimitrios I. Fotiadis (M’01) was born in Ioannina, Greece, in 1961. He received the Diploma degree in chemical engineering from National Technical University of Athens, Greece, and the Ph.D. degree in chemical engineering from the University of Minnesota. Since 1995, he has been with the Department of Computer Science, University of Ioannina, Greece, where he currently is an Associate Professor. He is the director of the Unit of Medical Technology and Intelligent Information Systems. His research interests include biomedical technology, biomechanics, scientific computing, and intelligent information systems.

141

Lampros K. Michalis was born in Arta, Greece, in 1960. He received the M.D. degree with distinction from the Medical School, University of Athens, Greece, in 1984. Since 1995, he has been with the Medical School, University of Ioannina, Greece, where he is currently an Associate Professor of Cardiology. He is in Charge of the Coronary Care Unit and the Catheter Laboratory of the University Hospital of Medical School, Ioannina, Greece. His research interests focus on vibrational angioplasty using guidewires, bioengineering, and interventional cardiology. In 1989, he was awarded with Distinction for his M.D. Thesis from the Athens Medical School, Greece.

George E. Bozios was born in Ioannina, Greece, in 1975. He received the B.Sc. degree in physics from the University of Ioannina, Ioannina, Greece, in 1998. He received the M.Sc. degree in medical physics from the University of Surrey, Guildford, U.K., in 1999. He is currently working toward the Ph.D. degree at the University of Ioannina. His research interests include dosimetry in radiotherapy, dosimetry and application of intravascular brachytharepy, and quality assurance protocols in radiology. He is a Member of Greek Medical Physicists Association (EFIE) and a Member of the European Society of Therapeutic Radiology and Oncology (ESTRO) since 2001.