Segmentation automatique de la lumiere des arteres sur une sequence d'images Intra Vasculaires a l'Ultrason (IVUS)

Cardiovascular disease (CVD) is the leading cause of death in North America today. It is manifested in different forms such angina pectoris, heart attack, or atherosclerosis. Atherosclerosis is caused by the build-up of fatty substances and calcium deposits inside the walls of the coronary arteries. Eventually, the accumulation grows large enough to harden the plaque and almost completely block the blood flow within the artery. This is also known as artery stenosis. Therefore, it is imperative that techniques be developed to help the interventionist properly characterize and diagnose this form of CVD.;In IVUS images, the lumen is typically adjacent to the imaging catheter, and the coronary artery vessel wall mainly appears as three layers: intima, media, and adventitia. In clinical research, the two inner layers are of principal concern. The identification of the border between the lumen and intima, as well as, the identification of the border separating the media and adventitia are of vital interest to clinicians for the proper diagnosis of artery stenosis pre-operatively and for proper follow-up post-operatively. Several segmentation techniques have been developed for these purposes. Generally, these methods are not fully automatic and they segment IVUS images in the spatial domain, rather than analyzing complete sequences of images in the spatio-temporal domain.;The objective of this work is to propose an automatic algorithm that targets the successful extraction of the lumen border with respect to the coronary wall using all IVUS images in a given dataset. We suppose that the image dataset was acquired by a transducer having a range above 30 MHz. This minimum frequency reveals the apparent border separations, as well as the texture of the inner walls of the artery. The proposed algorithm performs texture analysis followed by a classification.;First, a pre-processing of the IVUS images is performed in order to transform the images into polar coordinates. During this step, the calibration markers visible on the images are removed. Also, images are selected by ECG-gating, that is, images falling on the same cardiac phase are extracted for analysis. Finally, the catheter visible in the polar images is removed as well.;One of the most commonly used medical devices for the identification of coronary artery stenosis is Intravascular Ultrasound (IVUS). A small transducer on the tip of a coronary catheter is moved, with the help of a guidewire, inside the coronary arteries (using high frequency sound waves) in order to visualize the interior walls of the artery. The sound waves that are emitted from the catheter tip are usually in the 20-40 MHz range. The catheter also receives and conducts the return echo information which constructs and displays a real time 2D ultrasound image of a thin section of the blood vessel currently surrounding the catheter tip. The guidewire is kept stationary and the ultrasound catheter tip is slid backwards, usually under motorized control at a pullback speed of 0.5 mm/s. This is useful as we can visualize the artery from the inside out, thus making it possible to quantify the severity of stenosis present. It is a means of showing the physician where the normal artery wall ends and the plaque begins.;Secondly, texture analysis is performed in order to create a characteristic space for each pixel of the image. Three dimensional co-occurrence matrices are used to perform the analysis. The matrices are calculated using cubes. These cubes are created using the current image coupled with a few images preceding and following it. Then, the characteristic space is created using a convolution on each pixel. The convolution consists of calculating the co-occurrence matrix for the cube, along several angular directions. Eight statistical characteristics are evaluated for each matrix. If many angular directions are considered for the analysis, the complexity of a given characteristic space increases and therefore a principle component analysis is performed on the data to reduce the space.;Third, a hybrid k-means classifier is used to separate the characteristic space into two distinct classes: (i) the lumen and (ii) the intima, media and adventitia region. The advantage of this classifier is that it is an unsupervised learning algorithm compared to the others such as the support vector machine algorithm for example.;Lastly, a post treatment is applied on the final results from the classification. A region growing algorithm is used to identify the border between the lumen and the other tissue regions. The image representing the classification is decomposed into binary values, where a value of one is assigned to pixels making up the lumen. Several morphological filters are applied to smooth the final image. The resulting lumen boundary is superimposed on the original IVUS images by using active contours in order to minimize interpolation errors when reconverting the image back to cartesian coordinates from the polar coordinates.;Validation is performed on two clinical IVUS datasets. In total, 270 images were analyzed and the area correlation between manual and automatic segmentation is 0.90. When removing images containing severe artifacts (i.e. catheter echo is present in the image or the lumen is not visible) the area correlations increase to 0.97, whereas the mean Euclidean distance between manual and automatic identified borders was 0.08 mm with a standard deviation 0.095 mm. The final results are comparable to published data. The proposed method automatically segments the lumen border by analyzing the different textures present in the image and by using a hybrid k-means classifier. Future work will rely on the classification of the morphology of the plaque present inside the coronary artery.