A Non-contrast CT-based Multi-Task Learning Framework for Co-segmenting Hematoma and Edema and Predicting Hematoma Expansion in Spontaneous Intracerebral Hemorrhage.

OBJECTIVES: To develop a multi-task learning (MTL) framework for the simultaneous segmentation of spontaneous intracerebral hemorrhage (sICH) and perihematomal edema (PHE), along with the prediction of hematoma expansion (HE). METHODS: This multicenter retrospective study collected baseline non-contrast CT (NCCT) images of 981 patients with sICH from four centers, with 429 cases for training, 183 cases for internal validation, and 369 cases for external testing. A multi-task sICH model (MTsICHM) was designed to simultaneously segment hematoma and PHE and predict HE, which employs a hard parameter-sharing architecture based on a 3D U-Net backbone. The model processes dual-window NCCT inputs and generates three outputs: a binary hematoma mask, a binary PHE mask, and a binary HE label. For comparative evaluation and ablation studies, four control models were constructed: a multi-task segmentation comparative model, a multi-task classification comparative model, a single-task segmentation comparative model, and a single-task classification comparative model. RESULTS: For hematoma segmentation, MTsICHM achieved mean Dice of 0.909 and Intersection over Union of 0.872; for PHE segmentation, it attained 0.686 and 0.644, respectively, consistently outperforming the comparative models. For HE prediction, it yielded an AUC of 0.888, with corresponding accuracy, sensitivity, and specificity of 0.765, 0.949, and 0.710. CONCLUSIONS: The MTsICHM accurately segments hematoma and PHE while effectively predicting HE, demonstrating substantial potential for clinical application. ADVANCES IN KNOWLEDGE: MTsICHM simultaneously quantifies hematoma and PHE and predicts HE from a single NCCT, with high consistency to manual segmentation and clinical utility for risk identification.