Lightweight Dual-Task Deep Learning Framework for Malignancy Classification and Fuzzy Severity Estimation in Mammography

Abstract

Early detection of breast cancer is crucial for improving survival, with mammography serving as the primary screening tool despite challenges such as tissue overlap, subtle lesions, and breast density variation. Although deep learning has improved automated analysis, most existing methods focus on binary classification and single-view imaging, with limited use of clinical metadata or joint malignancy classification and fuzzy severity estimation frameworks. This study proposes a lightweight dual-task framework for simultaneous malignancy classification and fuzzy severity estimation using paired Cranio-Caudal (CC) and Medio-Lateral Oblique (MLO) views along with patient metadata. A structured preprocessing pipeline enhances data reliability through CC–MLO pairing, normalization, resizing, and memory-efficient loading. The model is based on an EfficientNet-B0 backbone with dual-view fusion and metadata integration, trained using a two-phase fine-tuning strategy and fuzzy severity encoding. Experiments on 7,808 images from the Mini-DDSM dataset achieved 97.83% accuracy and 0.98 ROC–AUC, with severity-level ROC–AUC values ranging from 0.90 to 0.96. Comparative results against ResNet50, DenseNet121, and Vision Transformer (ViT) showed improved performance with lower computational cost. The proposed framework integrates multi-view imaging, metadata, and fuzzy severity estimation within a unified lightweight architecture for efficient and reliable computer-aided breast cancer screening.