Cascaded Framework with Complementary CMR Information for Myocardial Pathology Segmentation

Abstract

Myocardial pathology segmentation in cardiac magnetic resonance (CMR) is an important step for patients suffering from myocardial infarction. In this paper, we present a cascaded framework with complementary information for infarcted and edema regions segmentation in CMR sequences. Specifically, instead of using all the three CMR sequences as joint inputs, we first use a 2D U-Net with balanced-Steady State Free Precession (bSSFP) cine sequence to segment the whole heart (left ventricle and myocardium) because bSSFP can capture cardiac motions and present clear boundaries. Then, we crop the whole heart as a region of interest (ROI). Finally, we segment the scar and edema regions in the late gadolinium enhancement (LGE) and T2 CMR sequence ROI. We evaluate the proposed method on MICCAI 2020 MyoPS testing set and achieve Dice scores 0.6283 ± 0.2772 for scar and 0.5419 ± 0.2406 for the combination of edema and scar, which is better than the inter-observer variation of manual scar segmentation (0.5243 ± 0.1578).