Deep Learning for Multi-Level Detection and Localization of Myocardial Scars Based on Regional Strain Validated on Virtual Patients

How well the heart is functioning can be quantified through measurements of myocardial deformation via echocardiography. Clinical assessment cardiac function generally focused on global indices relative shortening; however, segmental strain have been shown to abnormal in regions disease such as scarring. In this work, we propose a single framework predict scars at global, territorial, and levels using regional traces input convolutional neural network (CNN). An anatomically meaningful representation data from clinically standard bullseye multi-channel 2D image proposed, thus enabling use state-of-the-art configurations. A Fully Convolutional Network (FCN) trained detect localize scar left ventricular (LV) traces. Simulated controlled dataset virtual patients with varying degrees locations used for training validation. The proposed method successfully detects localizes 98% 5490 ventricle segments 305 test set only. Due sparse existence dataset, only 10% LV are scarred. Taking imbalance into account, class balanced accuracy calculated 95%. proves successful cohort offers potential solve detection problem real patient cohorts.