Session SA2.3

A New Tool for Automated Border Detection and Characterization of Atherosclerotic Plaque Composition in IVUS Images

A Taki*, A Roodaki, A Bigdelou, G Unal,
SK Setarehdan, N Navab

Technical University of Munich
Munich, Germany

Intravascular ultrasound (IVUS) provides high-resolution tomography visualization of the coronary arteries. In this paper, a new tool capable of automatic IVUS image analysis is proposed. This tool can be performed on all the IVUS frames of the heart cycle and thus permits to outperform the longitudinal resolution of Virtual Histology(VH). It also gives the possibility of extracting histological information from the large archive of the IVUS images obtained in the past for which their VH are not available. There are two main sections in this study. One is the border detection to extract plaque area. Shape and intensity priors are introduced to constrain the solution space in a variation framework. Inner and outer borders are obtained by updating their weight vectors that represent the contours in a reduced dimensional shape space with a Gaussian probability distribution. The second step is plaque characterization in which the pixels in the plaque area are classified to different plaque types. Run-length feature extraction method is used for recognizing three plaques i.e. dense calcium (DC), fibro-fatty(FF), and necrotic core(NC). Here, images are first transformed into polar coordinates and then swept by a 11*11 window. Totally 22 features are extracted from each window in horizontal and vertical directions and assigned to the center pixel. After feature extraction step, feature vectors are classified into the defined plaque types using a Support Vector Machine (SVM) classifier. To obtain a standalone executable application, the program with a graphical user interface (GUI) was compiled by C++ compiler in Microsoft Visual Studio 2005.IVUS images with size of 400*400 pixels were acquired using a 30-MHz transducer at a 0.5 mm/s pullback speed. The proposed methods were then applied on the set of 1275 frames from IVUS pullbacks of 18 different patients. The border detection algorithm achieved a 96.99% correct classification for the lumen contour, and a 92.74% for the media-adventitia contour. The characterized IVUS images were validated by their corresponding VH images statistical parameters were calculated for each technique. Using the run-length features, the overall accuracy of 73% was achieved. 73% (93%), 85% (79%), and 42% (84%) sensitivity (specificity) were achieved for DC, FF, and NC respectively.

(Abstract Control Number: 218)