Objective  To evaluate a mammographic image-based deep learning model for the prediction of the benign and malignant nature of breast imaging reporting and data system (BI-RADS) category four microcalcifications.

Methods We retrospectively analyzed 543 mammographic images from 321 patients with pathologically confirmed BI-RADS 4 breast lesions at Tianjin Medical University Cancer Institute & Hospital between January 2018 and January 2023. We extracted radiomics features (including microcalcification shapes, gray-scale values, and point densities) and integrated them into a ResNet34 deep learning model incorporating a spatial attention mechanism (SpatialResNet) to comprehensively predict benign or malignant microcalcifications. We randomly split the dataset into training and test sets at a 7:3 ratio and evaluated model performance using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1-score. We performed subgroup analyses for BI-RADS 4A, 4B, and 4C lesions, and applied gradient-class activation mapping for model interpretability.

Results The SpatialResNet model achieved an AUC, accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1-score of 0.884, 0.783, 0.773, 0.816, 0.500, 0.939, and 0.870 on the test set, respectively, outperforming the ResNet34 model without the attention mechanism across all metrics. Further subgroup analysis demonstrated that the model maintained robust diagnostic performance across BI-RADS 4 subcategories with varying risks of malignancy.

Conclusions The fusion model of combined radiomics features and deep learning spatial attention mechanism we established in this study allows for better distinction of the benign and malignant nature of breast BI-RADS type 4 microcalcifications.