VOLUMETRIC ASSESSMENT OF INTRACORONARY PATHOLOGICAL FORMATIONS USING OCT IMAGING IN KAWASAKI DISEASE

Abstract

Kawasaki disease (KD) can lead to coronary artery aneurysms with subsequent intimal hyperplasia, neovascularization, fibrosis, calcification, and macrophage accumulation. Intracoronary Optical Coherence Tomography (OCT) was found promising for characterization of coronary artery tissue layers in KD patients. Cross-sectional OCT images of coronary arteries play a crucial role to estimate the thickness variations of arterial wall layers, but the visualization is limited on a 2D plane. Automatic volumetric assessment of pathological formations of each tissue would be of high interest to quantify pathological formations. Hence, the goal of this study is to perform automatic registration of 2D OCT cross-sections for volumetric analysis of pathological formations in KD patients. In this study, 26 OCT cross-sectional pullbacks were obtained from different pediatric patients with KD using ILUMIEN ® OCT system. Each pullback sequence consists of ∼100 images per patient. We designed a model to compound 2D images of coronary artery tissues using a Convolutional Neural Network (CNN). CNNs were found to be very effective techniques to extract automatically specific and discriminative attributes of various tissues from OCT images. A fully-automated tissue classification model is used to characterize intima, media, neovascularization, fibrosis, and calcification using CCN and deep learning algorithms. Similarity between deep features extracted from different coronary artery tissues is used for registration. Paired t-test is applied between every two consequent frames of each pullback to validate tissue alignment before and after applying CNN-based registration. Alpha is set to 0.05 and significant difference was observed if the p-value was less than 0.05. The 3D models (figure) show a good alignment of intracoronary tissue layers and pathological formations with very low p-values (-log10 (p-value) = 34.35 ± 26.41 and significant difference of 0.03 ± 0.01 for all patient’s pullback). We applied our technique to create 3D reconstruction and analyze volume variations of coronary artery layers and pathological formations caused by KD on coronary artery tissues. By doing so, we can create an automated diagnostic framework providing clinicians with an operator-independent diagnosis of histological coronary lesions. From a visual 3D rendering of qualitative findings, we can generate a quantitative diagnostic output. This study contributes to estimate volumetric variations of various pathological formations caused by KD on coronary artery tissue layers. A visual 3D rendering of qualitative findings, together with a quantitative diagnostic report can greatly enhance patient care and management.