186 - AUTOMATIC EVALUATION OF CORONARY ARTERY DISTENSIBILITY IN KAWASAKI DISEASE FROM OCT IMAGES

Background: Kawasaki disease (KD) is a complication which results in an inflammation of coronary arteries, which can lead to aneurysm formation, stenosis, and even myocardial infarction. In the context of KD, affected coronary arteries often lose their distensibility: it results in a thickening of the intima and a reduction in the size of the media. Optical Coherence Tomography (OCT) imaging has been proposed to visualize luminal changes, and to evaluate the coronary arteries's distensibility, however, visual identification and tracking on the OCT imaging can be challenging.

Methods: A new segmentation approach based on deep learning U-Net architecture has been trained to evaluate the distensibility in coronary arteries (CA). A database of 10 stationary sequences of intracoronary cross-sectional images obtained from different pediatric patients with KD were used. Each stationary sequence contained 121 images. To start, 96 images were annotated to characterize the lumen. Then, based on these annotated images, the segmentation model was trained to recognize the lumen from intracoronary OCT images. After having characterized the lumen on images, two different approaches were used to evaluate the distensibility of the lumen over time. The first one was to determine the area of the segmented lumen and to follow its changes across the time. The second one was to monitor the evolution of the minimal diameter of the lumen. The average distensibility percentage of the lumen for each patient was computed, which assess the ability of an artery to change diameter in a cardiac cycle.

Results: The U-Net model successfully segmented the lumen with a Dice coefficient equals to 0.83 and Intersection Over Union (IOU) coefficient equals to 0.91. The distensibility of each sequence showed a percentage of distensibility ranging from 6% to 43%, mean = 21.5%, median = 20%, standard deviation = 13.6%.


Conclusion: This study paves the road for automatic assessment of CA distensibility in KD from OCT imaging. For future work, a larger database of KD, along with proper manual annotation might lead to better understand the physical properties of CA.