MP-SAM: A Novel Multi-Box Prompt-Guided and CNN-Augmented Method for Subtype Segmentation of Intracerebral Hemorrhage

Intracerebral hemorrhage (ICH) is a life-threatening condition, and precise lesion segmentation is crucial for clinical intervention. Existing methods still face challenges in terms of accuracy and robustness, especially when dealing with small targets and blurred boundary regions in ICH subtypes. To address these issues, this paper proposes MP-SAM, a multi-class automatic segmentation framework for ICH based on the vision foundation model, Segment Anything Model (SAM). The framework introduces a multi-box prompt mechanism to alleviate the class imbalance issue and enhance the detection capability of small and heterogeneous hematomas. It also incorporates a direction-aware PConv module with parallel CNN branches to strengthen local feature modeling, compensating for the limitations of the ViT encoder in capturing local features. Furthermore, low-rank adaptation and efficient parameter tuning strategies are employed to reduce model complexity. Experimental results on the BHSD dataset demonstrate that MP-SAM outperforms existing methods in metrics such as Dice coefficient and Hausdorff distance. Ablation studies further validate the effectiveness of the proposed mechanisms in improving segmentation accuracy and boundary recognition.