Deep Transfer Learning for Mammographic Masking Level Classification with Explainability Analysis

Abstract

Mammographic masking levels were classified using deep transfer learning architectures according to the Masking Potential index. Because mammographic masking can obscure breast lesions and affect mammogram interpretation, accurate, reliable, and interpretable classification is important. Three transfer learning models (VGG16, ResNet50, and InceptionV3) were trained using the Synthetic CSAW 100k Mammograms dataset, which was generated from the CSAW-M dataset using a latent diffusion model. Five-fold cross-validation, Receiver Operating Characteristic (ROC) analysis, and the area under the ROC curve (Area Under the Curve (AUC)) were used to evaluate the models’ performance, while Gradient-weighted Class Activation Mapping (Grad-CAM) was used to enhance model interpretability by highlighting regions contributing to the classification decisions. InceptionV3 outperformed the other evaluated models, achieving a mean Accuracy of 98.82% and a mean AUC of 99.94%. These findings suggest that deep transfer learning models may provide a suitable approach for the automated classification of mammographic masking levels and may also support the development of computer-aided tools for mammographic image analysis.