Dual-Path Feature Aggregation Network Combined Multi-layer Fusion for Myocardial Pathology Segmentation with Multi-sequence Cardiac MR

Abstract

Deep convolutional neural networks have shown great potential in medical image segmentation. However, automatic cardiac segmentation is still challenging due to the heterogeneous intensity distributions and indistinct boundaries in cardiac magnetic resonance (CMR) images, especially for myocardial pathology segmentation. In this paper, we present a dual-path feature aggregation network combined multi-layer fusion (MF&DFA-Net) to overcome these misclassification and shape discontinuity problems in myocardial pathology segmentation. The proposed network is aimed to maintain a realistic shape of the segmentation results and predict the position of myocardial pathology, which network is divided into two parts: the first part is a non-downsampling multiscale nested network (MN-Net) which restrains the cardiac shape and maintains the global information, and the second part is multiscale symmetric encoding and decoding network (MSED-Net) that can retain details. Three sequences of CMR images were adopted for multi-layer fusion training which included three inputs and one output. We can segment left ventricular (LV) blood pool, right ventricular (RV) blood pool, LV normal myocardium, LV myocardial edema, LV myocardial scars simultaneously with MF&DFA-Net. We randomly took the 90% data for training and 10% data for verification which data provided by the organizer of the 2020 Medical Image Computing and Computer Assisted Interventions (MICCAI) myocardial pathology segmentation challenge (MyoPS 2020). Compared with inter-observer, we increased the Dice value of myocardial scar segmentation by 8.08%.