Myocardial Pathology Segmentation of Multi-modal Cardiac MR Images with a Simple but Efficient Siamese U-shaped Network

Segmentation of multi-modal myocardial pathology images is a challenging task, due to factors such as the heterogeneity caused by large inter-modality and intra-modality intensity variations in multi-modal images, and the diversity of location, shape, and scale of lesion regions. Existing methods based on multi-modal segmentation cannot effectively integrate and utilize complementary information between multiple modalities, leading to the difficulty in segmenting edema and discontinuous scars. In this paper, we propose a simple but efficient U-shaped network, named Siamese U-Net, to solve these problems. There are two aspects to our method. First, we adopt a multi-modal complementary information exploration network (MCIE-Net) to explore the correlations across multi-modal images and simultaneously segment cardiac structures and myocardial pathology. This method is able to fully leverage complementary information between different modalities. Second, to obtain accurate and continuous segmentation of edema and scars, we use a lesion refinement network (LR-Net) with the same architecture as the MCIE-Net, which extracts lesion features to enhance the fusion of lesion information. We conducted extensive experiments on the MyoPS 2020 and MS-CMRSeg 2019 datasets to demonstrate the effectiveness of our proposed approach. We obtained an average Dice score of 0.734 ± 0.088 for the myocardial edema + scars on the MyoPS 2020 test set, a result which outperformed the state-of-the-art method. These results are a 0.9% improvement over the segmentation results of our previous work, and exceed the results of the winner of the MyoPS 2020 challenge by 0.3%.