Explainable Precision Medicine in Breast MRI: A Combined Radiomics and Deep Learning Approach for the Classification...

Area of Science:

• Radiology
• Artificial Intelligence
• Biomarkers

Background:

• Background parenchymal enhancement (BPE) in DCE-MRI is a key biomarker for breast cancer risk and treatment.
• Current visual classification of BPE (BI-RADS) suffers from inter-reader variability.
• A standardized, objective classification method for BPE is needed.

Purpose of the Study:

• To develop and evaluate a deep neural network (DNN) for automated, standardized classification of BPE into BI-RADS categories.
• To assess the accuracy and explainability of the DNN model using radiomic features and Shapley values.

Main Methods:

• Retrospective analysis of DCE-MRI scans from 27 healthy female subjects.
• Extraction of radiomic features from segmented BPE regions.
• Training a DNN on latent representations of radiomic features for BPE classification.
• Utilizing Shapley values for model interpretability at the feature level.

Main Results:

• The DNN achieved a BPE classification accuracy of 84 ± 2% (p < 0.00001).
• Misclassifications primarily occurred between adjacent BI-RADS classes.
• Shapley values revealed distinct radiomic features driving the classification for each BPE category.

Conclusions:

• An accurate and explainable AI pipeline for BPE classification was developed.
• The method reduces reliance on subjective visual assessment and user-dependent feature selection.
• This automated approach offers a standardized tool for evaluating a crucial breast cancer biomarker.