Swin-DAFU: denoising-augmented feature fusion unit for breast lesion segmentation and classification in 2D mammograms

Abstract

Breast lesion segmentation and benign/malignant classification from 2D mammographic images remain challenging because lesion boundaries are often subtle, irregular, and affected by variations in breast density, image contrast, and acquisition conditions. This study presents Swin-DAFU, a Swin Transformer-based encoder-decoder framework with a Denoising-Augmented Feature Fusion Unit for mammography-based breast lesion analysis. The proposed model uses a Swin Transformer encoder to extract hierarchical multi-scale image features, followed by a lightweight perturbation-denoising feature refinement unit that improves the fused latent representation before decoder-based segmentation. The proposed DAFU module is not a conventional generative diffusion model and does not perform multi-step image-level or mask-level diffusion sampling. Instead, it applies a single-step Gaussian perturbation to the fused latent feature representation, followed by learned residual denoising using a lightweight convolutional block. This design is used as a noise-regularized feature refinement mechanism to reduce unstable activations and improve lesion-boundary representation with limited computational overhead. The framework was evaluated on CBIS-DDSM and RTM mammography datasets for lesion segmentation and benign/malignant classification. The model achieved a Dice coefficient of 0.948 on CBIS-DDSM and 0.941 on RTM with classification accuracy values of 0.959 and 0.962, respectively. The corresponding AUC-ROC values were 0.995 and 0.992 under the evaluated data split and preprocessing protocol. Although these results are promising, they should be interpreted within the present experimental setting. Independent external validation was not performed, and the reported AUC values may be dataset- and split-dependent. Grad-CAM visualizations are provided as qualitative model-behavior examples, while supplementary pointing-game accuracy, Brier score, and expected calibration error are reported as quantitative model-behavior and calibration analyses under the evaluated test setting. Further validation on independent mammography cohorts is required before broader clinical generalization can be claimed. Implementation and reproducibility resources for this study can be sourced using the following link: ( https://github.com/narayanam-prasanthi/Swin-DAFU ).