The classification of malignant versus benign microcalcifications in mammograms remains a critical yet challenging task in breast cancer screening. Deep learning models, particularly convolutional neural networks, have demonstrated promising results; however, few studies have systematically compared different architectures within this domain. We evaluated the classification performance of two ResNet variants (ResNet-50 and ResNet-101) and five EfficientNet models (B0 to B4) using a five-fold cross-validation framework on 3,674 mammographic slices labelled with BI-RADS 1-2 or 5-6. Performance metrics included accuracy, area under the curve (AUC), and weighted F1-score. We further applied pairwise Wilcoxon signed-rank tests to assess the statistical significance of differences between the models. All EfficientNet models significantly outperformed the ResNet variants in terms of the F1 score (p < 0.05). Among the EfficientNet models, although B3 achieved the highest overall metrics, (accuracy = 86.9%, AUC = 0.998, weighted F1 = 0.869), the performance differences within the EfficientNet group were not statistically significant. EfficientNet-B0 provided comparable performance with much faster inference time. EfficientNet models exhibit superior performance compared to traditional ResNet architectures in the classification of mammographic calcifications. While B3 demonstrated slightly superior performance, B0 may provide a more favourable trade-off between accuracy and inference efficiency. These findings support the integration of lightweight EfficientNet variants into real-world diagnostic workflows.
Wei-Chung Shia (Fri,) studied this question.