DCC-Net: depth-wise contextual attention with 3d connectivity constraints for intracerebral hemorrhage segmentation

Abstract
Intracerebral hemorrhage (ICH) often leads to high disability and mortality rates, making accurate segmentation of hematoma regions critical for clinical assessment and treatment planning. Although deep learning has advanced ICH segmentation techniques, it still faces challenges due to the complexity of hematoma morphology and location. ICH-related a priori knowledge provides additional information about hematoma morphology, location, and evolution, which helps the model more accurately identify and segment hematoma regions. In this study, we propose the Depth-wise Contextual Attention with 3D Connectivity Constraints Network (DCC-Net), an architecture that enhances intracerebral hemorrhage segmentation accuracy through a synergistic dual-component design. The Depth-wise Contextual Attention (DCA) module leverages hematoma anatomical spatial continuity by employing a depth-wise local window attention mechanism to adaptively aggregate adjacent slice information, thereby strengthening intracranial hematoma feature representation. Additionally, the 3D Connectivity Constraint (3D-CC) loss ensures three-dimensional structural integrity through differentiable topological computations. This synergistic design collectively enhances model robustness for complex hematoma morphologies. We evaluated the proposed DCC-Net and compared it with current state-of-the-art methods. Tested using rigorous intracerebral hemorrhage evaluation metrics, the experimental results show that our method improves over the underlying model on most evaluation metrics and outperforms other existing techniques in key performance indicators.