MSF-Swin-ICFNet: An Image-Derived Multi-Scale Swin Transformer Framework with Cross-Scale Feature Fusion for Breast Lesion Classification and Segmentation in Mammograms.

Background

Accurate breast lesion classification and localization from mammograms remain challenging because lesions vary in size, morphology, density, contrast, and boundary clarity. This study proposes MSF-Swin-ICFNet, an image-derived Multi-Scale Swin Transformer framework with Image-Derived Cross-Scale Fusion (ICF) for joint benign-malignant classification and lightweight lesion localization.

Methods

The proposed framework extracts hierarchical representations from four Swin Transformer stages and aligns, adaptively recalibrates, and fuses these multi-scale features into a shared representation for classification and lesion-mask prediction. The model was evaluated on CBIS-DDSM, comprising 891 patients and 1592 mammograms, and RTM, comprising 573 patients and 10,070 mammograms. Patient-level stratified partitioning was used to ensure that mammograms and associated annotations from the same patient remained within a single data partition. Classification was evaluated using accuracy, precision, recall, F1-score, and AUC-ROC, while lesion localization was assessed using the Dice coefficient.

Results

MSF-Swin-ICFNet achieved strong mammogram-level classification performance on the patient-disjoint test cohorts. On CBIS-DDSM, the model achieved 0.949 accuracy, 0.931 F1-score, and 0.985 AUC-ROC. On RTM, it achieved 0.968 accuracy, 0.956 F1-score, and 0.992 AUC-ROC. The corresponding Dice coefficients for lesion localization were 0.786 on CBIS-DDSM and 0.878 on RTM. Ablation experiments further supported the contribution of cross-scale fusion, scale recalibration, feature refinement, and the boundary-aware localization head.

Conclusions

MSF-Swin-ICFNet demonstrates promising retrospective performance for image-based breast lesion classification with complementary lesion localization. The results support the technical potential of adaptive hierarchical feature fusion for mammographic analysis. However, the model is not intended as a standalone diagnostic system, and independent multicentre validation and radiologist reader studies are required before clinical deployment.