A novel deep learning architecture named ContraFormer provides accurate and reliable breast mass segmentation on mammographic datasets, achieving competitive performance compared to recent state-of-the-art segmentation methods. In cross-dataset experimental evaluations, the framework demonstrated robust generalization performance, recording a Dice coefficient of 93.32% and a 95% Hausdorff distance of 0.98 pixels.
Breast mass segmentation of mammograms is an essential and critical task for computer-aided diagnosis systems because it delivers reliable information for the early detection and assessment of breast cancer. However, existing deep learning methods continue to struggle with variations in mass appearance, indistinct lesion boundaries, and low contrast between lesions and surrounding breast tissues.
To address these challenges, ContraFormer integrates boundary-aware refinement modules within a CNN-Transformer architecture to achieve robust breast mass segmentation. The network uses a CNN-Transformer hybrid encoder to capture hierarchical representations by directly combining local structural information with global contextual information.
To improve boundary localization, the framework introduces a Boundary-Contrastive Attention Module. This module distinguishes between lesion boundaries and surrounding breast tissue by learning contrastive features, thereby sharpening the delineation of mass margins.
The architecture also incorporates an Uncertainty-Guided Boundary Refinement module designed to identify boundaries with uncertain predictions and gradually improve boundary refinement, reducing segmentation error around ambiguous lesion contours. The method demonstrated its generalization capability in a cross-dataset evaluation using the CBIS-DDSM dataset for training and the INbreast dataset for testing.
Subject of Research: Deep learning framework for breast mass segmentation in mammograms
Article Title: ContraFormer: a boundary-contrastive CNN-transformer framework with uncertainty-guided refinement for breast mass segmentation
Article References: Dai, B., Ji, X.-Y., Ijaz, A., Sarfraz, T., Ahmed, N., Saeed, M., Saddozai, U. A. K., & Wang, M.-Y. (2026). ContraFormer: a boundary-contrastive CNN-transformer framework with uncertainty-guided refinement for breast mass segmentation. BMC Cancer. https://doi.org/10.1186/s12885-026-17231-4
Image Credits: AI Generated
DOI: 10.1186/s12885-026-17231-4
Keywords: breast cancer, mammography, image segmentation, deep learning, CNN-Transformer, ContraFormer, computer-aided diagnosis, medical imaging, boundary-contrastive, framework, uncertainty-guided, refinement
News Source: Ophelia Keating. (November 10, 2026). Hybrid CNN-Transformer Framework Improves Breast Mass Boundary Segmentation in Mammograms. Scienmag.



