Prior-Guided Feature Refinement Network with Zero Initialization for Robust Breast Tumor Segmentation in DCE-MRI

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is a critical modality for the clinical diagnosis of breast cancer. However, existing 3D convolutional neural networks (CNNs) are often constrained by restricted local receptive fields and remain highly susceptible to false-positive enhancements from background structures, such as blood vessels and benign glandular tissue. To address these limitations, this paper proposes a novel multi-stage segmentation framework. Our approach leverages a lesion spatial prior probability map, generated via a You Only Look Once (Yolo) detection model, to compensate for the global localization deficiencies inherent in traditional segmentation networks. Furthermore, built upon a multi-phase feature fusion backbone, we introduce a Priorguided 3D Spatial Attention (PSA3D) module. By employing a shallow tandem injection strategy coupled with a zeroinitialization mechanism, the proposed segmentation network (PFR-Net) maintains high training stability and robust fault tolerance, even when conditioned on inaccurate prior detections. Comprehensive evaluation on the large-scale, multi-center MAMAMIA dataset-comprising the DUKE, ISPY1, ISPY2, and NACT cohorts-demonstrates the robust efficacy of our framework. Quantitative results yield a Dice Similarity Coefficient(DSC) of 78.61%, a Jaccard index of 68.05%, an Average Symmetric Surface Distance (ASSD) of 2.71 mm, and a 95% Hausdorff Distance (HD95) of$\mathbf{1 5. 0 5 ~ m m}$. Ultimately, this work provides an efficient and robust computational paradigm for the computeraided diagnosis of breast cancer.