JC-MTL: Multi-task Learning with Joint Consistency Loss for Improved Subtype Segmentation of Intracerebral Hemorrhage

Deep learning-assisted diagnosis of intracerebral hemorrhagic (ICH) stroke has attracted increasing attention in recent years, owing to its capability to rapidly identify and quantify hemorrhages. However, accurate segmentation of hemorrhage subtypes remains a challenging task due to the scarcity of annotated subtype data, high intra-class variability, and ambiguous hematoma boundaries. To address these challenges, we propose a novel multi-Task learning framework with joint consistency loss (JC-MTL) for hemorrhage subtype segmentation. The framework integrates the tasks, including classification (CLS), boundary detection (BDE), and image reconstruction (REC), into the segmentation (SEG) pipeline to exploit shared and complementary representations via MTL, thereby improving data efficiency and model generalization. A joint consistency loss is introduced to enforce low-dimensional consistency across tasks, encouraging subtype-aware feature learning. Furthermore, a boundary enhancement module is incorporated into the BDE task to refine boundary representations and guide the SEG task, mitigating the impact of blurry hematoma edges. Experiments conducted on the BHSD dataset demonstrate the effectiveness of our JC-MTL framework, achieving superior performance compared to previous medical segmentation models. The results indicate promising potential for future clinical applications in ICH diagnosis.