Deep learning-based classification of benign and malignant breast microcalcifications in mammography

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Machine Learning for Diagnostics

Background:

• Accurate classification of mammographic microcalcifications is crucial for effective breast cancer screening.
• Deep learning models, particularly convolutional neural networks (CNNs), show promise but require comparative analysis of different architectures.
• Existing research lacks systematic comparisons of advanced CNN architectures for microcalcification classification.

Purpose of the Study:

• To systematically compare the performance of ResNet and EfficientNet architectures for classifying malignant versus benign microcalcifications in mammograms.
• To evaluate the trade-offs between classification accuracy, AUC, F1-score, and inference time for different deep learning models.
• To determine the optimal deep learning model for integration into breast cancer screening workflows.

Main Methods:

• A five-fold cross-validation framework was applied to 3,674 mammographic slices.
• Two ResNet variants (ResNet-50, ResNet-101) and five EfficientNet models (B0-B4) were evaluated.
• Performance was assessed using accuracy, AUC, and weighted F1-score, with statistical significance determined by pairwise Wilcoxon signed-rank tests.

Main Results:

• All EfficientNet models significantly outperformed ResNet variants in F1-score (p < 0.05).
• EfficientNet-B3 achieved the highest metrics (accuracy 86.9%, AUC 0.998, F1 0.869), though differences within EfficientNets were not statistically significant.
• EfficientNet-B0 demonstrated comparable performance with significantly faster inference times compared to other models.

Conclusions:

• EfficientNet models offer superior performance for mammographic microcalcification classification compared to ResNet architectures.
• EfficientNet-B0 presents a favorable balance of high accuracy and computational efficiency, making it suitable for real-world diagnostic applications.
• Lightweight EfficientNet variants are recommended for integration into clinical breast cancer screening workflows.