Hierarchical Semantic Concept Modeling for Generalizable Myocardial Pathology Segmentation on Multisequence CMR...

Myocardial pathology segmentation (MyoPS) aims to accurately quantify myocardial scar and edema, which is critical for precise assessment of myocardial infarction (MI) severity. Recent deep learning (DL) methods have shown promising performance. However, due to cross-domain distribution shifts, many methods exhibit limited generalizability because it is challenging to learn domain-invariant pathology information, resulting in degraded performance on unseen data. This work proposes a hierarchical semantic concept segmentation framework, referred to as HSCM-Net. This framework uses concept modeling of multisequence cardiac magnetic resonance (CMR) images and labels to integrate complementary information and enhance the generalizability of MyoPS. Specifically, the CMR images are first decomposed into three independent concept variables, and hierarchical priors are assigned to explicitly force them to model the sequence-irrelevant concept of shape, the semantic concept of pathology, and the domain-specific concept of appearance. Then, the anatomy segmentation and pathology segmentation are modeled as global and local concepts only related to shape and pathology, respectively. Finally, a variational inference (VI) framework is developed to approximate the posteriors of these conceptual variables, and the inference process is implemented with a deep neural network. Results on three-domain multisequence CMR datasets proved the advantageous generalizability of our HSCM-Net, which achieves a MyoPS Dice score of 0.577 on unseen target domains. Our code is released at https://github.com/Dongjinweicn/HSCMNet.