Accurate Myocardial Pathology Segmentation with Residual U-Net

Abstract

Accurate assessment of myocardial viability in multi-sequence cardiac magnetic resonance (CMR) images is desired to automate disease diagnosis. To classify myocardial pathology automatic segmentation methods are necessary. In this paper we propose to use an automatic segmentation for each slice in the short-axis view with convolutional neural network architecture based on U-Net. We compare performances of two different networks to segment myocardial pathologies. The best performance is obtained by using the U-net convolutional neural network architecture built from residual units trained by augmentation operations, showing that it is a practical approach for segmentation. The network performances are assessed on MyoPS 2020 challenge dataset consists of three-sequence CMR images from 45 patients. A five-fold cross-validation strategy is utilized to assess performance of the proposed method.