Optimized ResNet-50 framework for mammogram-based breast cancer classification: a comparative evaluation with EfficientNet-B0

Abstract

Breast cancer remains one of the most prevalent malignancies worldwide, underscoring the need for accurate and reliable mammographic interpretation. Computer-aided diagnosis (CAD) based on deep learning has emerged as a promising approach to improve both screening performance and diagnostic consistency, yet fairness-driven comparisons between popular convolutional backbones on public mammogram benchmarks remain limited. This study provides a statistically validated, fairness-driven comparison of two widely used convolutional neural network architectures, ResNet-50 and EfficientNet-B0, for mammogram-based breast cancer classification under a rigorously controlled, clinically motivated protocol. The proposed “optimized ResNet-50” framework is defined by patient-level stratified undersampling, paired 5-fold cross-validation with identical partitions, harmonized augmentation and training configurations, and dual statistical testing (paired t-tests and Wilcoxon signed-rank tests), emphasizing methodological rigor rather than architectural novelty. Across MIAS and CBIS-DDSM benchmarks, the models demonstrated complementary strengths, with EfficientNet-B0 excelling in screening-oriented tasks (normal vs. abnormal) and ResNet-50 offering more robust performance for diagnostic-oriented tasks (benign vs. malignant). These findings highlight the value of fairness-driven evaluation protocols in CAD research and support the feasibility of integrating lightweight convolutional neural networks (CNNs) into tiered clinical workflows, where different backbones are strategically deployed for initial screening and confirmatory assessment.