Dynamic Convolution Based Multi-Source Domain Generalizable Myocardial Pathology Segmentation

Myocardial pathology segmentation is essential for the clinical diagnosis and prognosis of myocardial infarction, as it can provide valuable pathological information from multi-sequence cardiac magnetic resonance (CMR) images. In practical applications, domain generalized medical image segmentation requires model to learn from multiple source domains and achieve robust performance on unseen target domains. However, in reality, medical images are typically collected from various hospitals and annotated by different experts with varying levels of expertise, often resulting in significant distribution gaps between the source and target domains. Addressing such domain shifts poses considerable technical challenges. In this paper, we introduce dynamic domain convolution and dynamic content convolution based domain generalization approach into our improved nnUNet pipeline. The experimental results indicate that this method shows great potential in myocardial pathology segmentation. In the MyoPS++ validation set, our method achieved mean Dice scores of 0.6394 and 0.684 in LV myocardial scar and the union of scar and edema, respectively.