Breast Cancer Detection from Mammograms Using ResNet-50 Transfer Learning and Physics-Informed Neural Networks

Early and reliable breast cancer detection remains a major challenge in medical imaging. We present a novel hybrid framework that combines a ResNet-50 backbone pretrained on ImageNet with a Physics-Informed Neural Network (PINN) branch enforcing the X-ray attenuation law. Trained end-to-end on the CBIS-DDSM mammography dataset, the model uses a composite objective: binary cross-entropy for lesion classification and a physics loss derived from the Beer–Lambert law applied to predicted attenuation coefficients. Data preprocessing includes orientation standardization, ROI cropping, augmentation, and transfer learning to mitigate limited-data issues. On held-out test data, our method achieves 95.8% accuracy, 95.4% precision, 96.2% recall, and ROC AUC of 0.97, outperforming a baseline ResNet-50 by 1.7%. Embedding domain knowledge into deep networks enhances generalization and robustness, suggesting a viable path for next-generation CAD systems.