Explainable Deep Learning for Breast Lesion Classification in Digital and Contrast-Enhanced Mammography

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Artificial intelligence (AI) shows promise for breast cancer screening.
• Integration of AI across mammographic modalities like Digital Mammography (DM) and Contrast-Enhanced Spectral Mammography (CESM) needs further exploration.
• Understanding AI model behavior across DM and CESM is crucial for equitable clinical adoption.

Purpose of the Study:

• To evaluate and compare the performance of three deep learning convolutional neural network (CNN) architectures (ResNet-18, DenseNet-121, EfficientNet-B0) on DM and CESM breast lesion classification.
• To analyze the decision-making patterns of these AI models across both mammographic modalities using SHapley Additive exPlanations (SHAP).

Main Methods:

• Utilized the public CDD-CESM dataset comprising 2006 images for binary classification tasks (Normal vs. Benign, Benign vs. Malignant, Normal vs. Malignant).
• Trained CNN models separately on DM and CESM images using transfer learning, weighted binary cross-entropy loss, and a 3-fold cross-validation scheme.
• Employed SHAP analysis to visualize and interpret model decisions at a pixel level.

Main Results:

• CESM demonstrated superior performance in Normal vs. Benign and Benign vs. Malignant classifications.
• Digital Mammography (DM) achieved the highest discriminative ability in the Normal vs. Malignant comparison (EfficientNet-B0: AUC = 97%, Accuracy = 93.15%).
• SHAP analysis revealed consistent, anatomically relevant decision patterns for both modalities, with CESM providing sharper focus and DM showing broader attention.

Conclusions:

• CESM enhances lesion discrimination via functional contrast, while DM offers highly accurate and explainable results with CNNs, despite simpler acquisition.
• Consistent SHAP-based relevance across modalities confirms that both DM and CESM preserve clinically meaningful information for AI analysis.
• This study is the first to directly compare DM and CESM using explainable deep learning models under identical conditions.