Deep Learning-Based Breast Cancer Classification in Mammography: A Lesion-Specific Approach With Hybrid Ensemble

Most mammographic computer-aided diagnosis systems are trained on all lesion types together, without accounting for the fundamental morphological differences between masses and calcifications. This study proposes and evaluates three ensemble strategies for malignancy classification on the Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM) under strict patient-level splitting across 1,428 patients and 2,704 images: unified models trained on all lesion types, lesion-specific models trained separately for masses and calcifications, and a hybrid ensemble that adaptively combines both via a validation-optimized fusion weight ( $\alpha ^{*} = 0.4$ ). Four convolutional neural network (CNN) architectures (ResNet50, DenseNet121, EfficientNet-B0, and ConvNeXt-Tiny) were evaluated under an identical preprocessing pipeline and patient-level benchmark, ensuring leakage-free comparison across all strategies. The results revealed a lesion-type-dependent relationship between training strategy and performance: mass-specific models outperformed unified models despite 46 % less training data $(AUC: 0.754 ~\text {vs.} \ 0.748, \Delta AUC{\,}={\,}0.006)$ , whereas calcification-specific models underperformed (AUC: 0.711). At the ensemble level, the hybrid model achieved statistically equivalent overall discrimination to the unified strategy (AUC: 0.782 vs. 0.783, $p$ = not significant), while reducing false positives by 11.9 % at 90 % sensitivity in this dataset (104 vs. 118), and significantly outperformed the lesion-specific ensemble ( $p \lt 0.001$ ). These findings suggest that training strategy selection may benefit from being lesion-type-specific rather than uniform and provide a reproducible patient-level benchmark for the CBIS-DDSM for future comparisons.