Automated Tissue Classification of Intracardiac Optical Coherence Tomography Images (conference Presentation)

Yu Gan,David Tsay,Syed A. Bin Amir,Charles C. Marboe,Christine P. Hendon

DOI: https://doi.org/10.1117/12.2212796

2016-01-01

Abstract:Remodeling of the myocardium is associated with increased risk of arrhythmia and heart failure. Our objective is to automatically identify regions of fibrotic myocardium, dense collagen, and adipose tissue, which can serve as a way to guide radiofrequency ablation therapy or endomyocardial biopsies. Using computer vision and machine learning, we present an automated algorithm to classify tissue compositions from cardiac optical coherence tomography (OCT) images. Three dimensional OCT volumes were obtained from 15 human hearts ex vivo within 48 hours of donor death (source, NDRI). We first segmented B-scans using a graph searching method. We estimated the boundary of each region by minimizing a cost function, which consisted of intensity, gradient, and contour smoothness. Then, features, including texture analysis, optical properties, and statistics of high moments, were extracted. We used a statistical model, relevance vector machine, and trained this model with abovementioned features to classify tissue compositions. To validate our method, we applied our algorithm to 77 volumes. The datasets for validation were manually segmented and classified by two investigators who were blind to our algorithm results and identified the tissues based on trichrome histology and pathology. The difference between automated and manual segmentation was 51.78 ± 50.96 μm. Experiments showed that the attenuation coefficients of dense collagen were significantly different from other tissue types (P < 0.05, ANOVA). Importantly, myocardial fibrosis tissues were different from normal myocardium in entropy and kurtosis. The tissue types were classified with an accuracy of 84%. The results show good agreements with histology.