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

Abstract

Simple Summary Breast lesions can be difficult to distinguish on mammograms because they vary considerably in size, shape, density, contrast, and visibility. This study developed MSF-Swin-ICFNet, an artificial-intelligence model that examines mammograms at multiple image scales and combines information from these scales to classify breast lesions as benign or malignant while also indicating their approximate location. The model was evaluated using two mammography datasets, CBIS-DDSM and RTM, with patients separated between training and testing to reduce information leakage. It achieved classification accuracies of 94.9% on CBIS-DDSM and 96.8% on RTM and produced lesion-localization masks. These findings suggest that combining image information from multiple scales can support breast-lesion analysis using mammograms alone. However, the model has been evaluated retrospectively, and independent multicentre validation and radiologist reader studies are required before considering its use in clinical practice.