EfficientSeg: A Simple But Efficient Solution to Myocardial Pathology Segmentation Challenge

摘要

Myocardial pathology segmentation is an essential but challenging task in the computer-aided diagnosis of myocardial infraction. Although deep convolutional neural networks (DCNNs) have achieved remarkable success in medical image segmentation, accurate segmentation of myocardial pathology remains challenging, due to the low soft-tissue contrast, irregularity of pathological targets, and limited training data. In this paper, we propose a simple but efficient DCNN model called EfficientSeg to segment the regions of edema and scar in multi-sequence cardiac magnetic resonance (CMR) data. In this model, the encoder uses EfficientNet as its backbone for feature extraction, and the decoder employs a weighted bi-directional feature pyramid network (BiFPN) to predict the segmentation mask. The former has a much improved image representation ability but with less computation cost than traditional convolutional networks, while the latter allows easy and fast multi-scale feature fusion. The loss function of EfficientSeg is defined as the combination of Dice loss, cross entropy loss, and boundary loss. We evaluated EfficientSeg on the Myocardial Pathology Segmentation (MyoPS 2020) Challenge dataset and achieved a Dice score of 64.71% for scar segmentation and a Dice score of 70.87% for joint edema and scar segmentation. Our results indicate the effectiveness of the proposed EfficientSeg model for myocardial pathology segmentation.