Comparation of Lightweight CNN and Transformer-Based Models for Mammogram Image Classification on CBIS-DDSM

Abstract

Detecting malignancy in mammograms requires models capable of recognizing local lesion features while preserving broader breast-tissue context. This study compares three lightweight architectures, DenseNet121, EfficientNetV2-S, and MobileViT-S, for binary classification of benign and malignant lesions in the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM). Region-of-interest images were converted to single-channel grayscale inputs of 224×224 pixels, enhanced using contrast-limited adaptive histogram equalization, and augmented through geometric and intensity transformations. Patient-level stratified group five-fold cross-validation was used to prevent data leakage. Probabilities from the five fold-specific models were averaged through cross-validation bagging and evaluated on an independent test set of 298 images form 145 patients. EfficientNetV2-S achieved the highest accuracy, macro-F1 score, and specificity at 0.6913, 0.6837, and 0.7412, respectively. DenseNet121 achieved the highest sensitivity, ROC-AUC, and PR-AUC at 0.6953, 0.7446, and 0.7127, respectively, indicating fewer false-negative predictions. MobileViT-S achieved 0.6510 accuracy and 0.7368 ROC-AUC with 4.94 million parameters, 2.8250 GFLOPs per model, and an end-to-end bagging latency of 11.8658 ms per image, compared with 15.0013 ms for EfficientNetV2-S and 16.3840 ms for DenseNet121. Overall, EfficientNetV2-S provided the best classification balance, DenseNet121 was preferable when malignancy sensitivity was prioritized, and MobileViT-S offered the greatest computational efficiency for resource-constrained devices. The standardized evaluation framework supports reproducible comparisons across architectures while accounting for patient-level variation. These findings show that model selection should consider predictive performance, computational cost, and the clinical consequences of classification errors.