Abstract 463: Convocals: A Convolutional Neural Network To Predict Symptoms And Major Secondary Cardiovascular Events Based On High-resolution Scanned Histological Slides

Abstract

Background: Atherosclerosis is a primary risk factor underlying myocardial infarction and stroke, and there remains an unmet need to identify atherosclerotic plaque subtypes to improve clinical outcomes. Histological examination is used to manually characterize the extent and type of atherosclerosis, and histological traits are associated with major adverse cardiovascular events (MACE). Recent applications of automated machine learning approaches may improve scalability and precision in predicting symptoms and cardiovascular events, however this has yet to be fully evaluated in the setting of atherosclerosis. Methods: CONVOCALS uses a convolutional neural network (CNN), feature embedding and attention-based pooling model to develop an internal representation of whole slide images (WSI) and learn position and scale invariant structures in the data. We evaluated this model on human arterial H&E histology WSI from the Genotype Tissue Expression Project (GTEx), consisting of 688 WSI of normal and atherosclerotic coronary, tibial and aortic artery tissues. Results and Conclusions: We evaluated the performance of a base model of CONVOCALS on the GTEx artery H&E WSI dataset, which achieved a mean AUC of 0.86 and accuracy of 0.72 on a test dataset of 68 WSI images. Principal component analysis demonstrated clear separation of the 4 distinct atherosclerosis stages. We also visualized different pathological hallmarks (e.g., calcification) using attention scores overlaying the WSI. We further evaluate CONVOCALS performance on more granular feature extraction and demonstrate improved WSI segmentation. Next we plan to evaluate the performance on a large-scale AtheroExpress dataset to predict symptoms and MACE.