ContraFormer: a boundary-contrastive CNN-transformer framework with uncertainty-guided refinement for breast mass segmentation

Breast mass segmentation of mammograms is an essential and critical task for computer-aided diagnosis (CAD) as it can deliver reliable information for the early detection and assessment of breast cancer. Existing deep learning methods still struggle with variations in mass appearance, indistinct lesion boundaries, and low contrast between the lesions and the surrounding tissues. To address these limitations, this paper proposes a novel network architecture, ContraFormer, which is based on a CNNTransformer model and integrates with boundary-aware refinement modules to achieve robust breast mass segmentation by exploiting local structural information and global contextual information. The proposed framework utilizes a CNN-Transformer hybrid encoder to capture hierarchical representations by combining local structural information with global contextual information. A Boundary-Contrastive Attention Module (BCAM) is proposed to further distinguish between lesion boundaries and surrounding breast tissue by learning contrastive features to improve boundary localization. Furthermore, an Uncertainty-Guided Boundary Refinement (UGBR) module is designed to identify boundaries of uncertain predictions and gradually improve the refinement of boundaries, thereby reducing the segmentation error around ambiguous lesion contours. The proposed method is evaluated in a cross-dataset scenario using the CBIS-DDSM dataset for training and the INbreast dataset for testing. Experimental results indicate that ContraFormer achieves competitive performance compared to recent state-of-the-art segmentation methods with a Dice coefficient of 93.32% and an HD95 of 0.98 pixels. The results show that the suggested framework provides accurate and reliable breast mass segmentation and has good generalization performance on mammographic datasets.