A Fully Automatic Patch-Based Deep Learning Pipeline for Mass and Microcalcification Detection in Mammography: A Preliminary Study

This study proposes a fully automatic, patch-based pipeline for the classification and localization of masses and microcalcifications in mammographic images. Leveraging the EfficientNetB6 convolutional neural network, the method classifies image patches into one of three categories, healthy tissue, mass, or calcification, based on their lesion content. The CBIS-DDSM dataset was used, containing over 3500 annotated images. Preprocessing included removing irrelevant content, enhancing image quality through normalization, filtering, and contrast enhancement. Patches (528 × 528) were extracted with 50% overlap and labeled if they contained at least 20% of a lesion. The network was trained using transfer learning followed by fine-tuning, with data augmentation applied to improve generalizability. During inference, patches were extracted with 85% overlap, classified, and reassembled into full-size probability maps, which were smoothed to aid visualization. The model achieved robust performance, with an average accuracy of 81%, F1-score of 81%, and area under the ROC curve (AUC) of 94% across all classes. These results highlight the effectiveness of EfficientNetB6 in extracting meaningful features from mammograms, even with limited lesion visibility. The generated probability maps serve as visual aids to support clinical decision-making and enable integration with techniques like Grad-CAM for further interpretability. The approach enables efficient processing (~3 min per image), making it viable for real-time applications. This framework represents a step towards an automated, interpretable, and efficient mammographic reporting system.