Mamba‑MFNet: A hierarchical supervised network based on the fusion of axial and cross-modal attention for breast DCE‑MRI tumor segmentation.

Early detection and accurate segmentation of breast cancer are critical for improving cure rates and optimizing treatment strategies. However, existing 3D segmentation methods face significant limitations in global context modeling and efficient multimodal feature fusion. To address these challenges, this study proposes Mamba‑MFNet, a 3D breast tumor segmentation framework based on a state-space model, which introduces three key innovations. First, a cross-modal attention mechanism dynamically fuses six DCE-MRI temporal modalities with one breast anatomical mask and adaptively adjusts the contribution of each channel using a learnable similarity matrix and weight parameters. Second, a three-dimensional axial attention module computes multi-head attention along the coronal, sagittal, and transverse planes, and reorganizing features through convolutions to model global dependencies with linear complexity, effectively filling in a blind spot of the local convolution receptive field in capturing long-range contex. Finally, Hierarchical loss strategy is adopted in the high-resolution layer, combining multi-scale deep supervision and boundary-aware loss, introducing focal loss in the middle layer, and focusing on global semantics in the low-resolution layer, which significantly improves the detection accuracy of tiny lesions and blurred boundaries. Experimental results on the breast DCE-MRI dataset validate the effectiveness of the proposed method: compared with baseline UMamba, Mamba-MFNet improved the Dice coefficient, Precision, Recall, and IoU by 5.07%, 2.84%, 4.75%, and 3.39%, respectively. Additionally, nine clinical experts commended highly the model's accuracy, boundary clarity, and clinical decision support capability in a subjective evaluation, further demonstrating the reliability and usefulness of Mamba-MFNet in real-world diagnostic and treatment scenarios.