Related Experiment Video
Updated: Sep 19, 2025
15:48
Tracking the Mammary Architectural Features and Detecting Breast Cancer with Magnetic Resonance Diffusion Tensor Imaging
Published on: December 15, 2014
22.6K
Robust evaluation of tissue-specific radiomic features for classifying breast tissue density grades.
Vincent Dong1, Walter Mankowski2, Telmo M Silva Filho3
1University of Pennsylvania, Department of Bioengineering, Philadelphia, Pennsylvania, United States.
Journal of Medical Imaging (Bellingham, Wash.)
|June 2, 2025
Summary
This study introduces a robust RFE-SHAP method for classifying breast density using radiomic features from digital breast tomosynthesis. The approach accurately captures density progression and generalizes well to validation datasets.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence in Healthcare
- Biomedical Data Science
Background:
- Accurate breast density assessment is crucial for breast cancer risk evaluation, as dense tissue can mask lesions.
- Radiologist assessment of breast density shows significant variability despite standardized guidelines.
- Current deep learning tools for automated breast density assessment face challenges in robustness and interpretability.
Purpose of the Study:
- To assess the robustness of a Recursive Feature Elimination with SHapley Additive exPlanations (RFE-SHAP) methodology for classifying breast density grades.
- To identify highly predictive and influential tissue-specific radiomic features using RFE-SHAP.
- To evaluate the classification performance and intrinsic separability of breast density grades using logistic regression and unsupervised clustering.
Main Methods:
- Radiomic features were extracted from raw central projections of digital breast tomosynthesis (DBT) screenings.
- A Recursive Feature Elimination with SHapley Additive exPlanations (RFE-SHAP) method was employed for feature selection.
- Logistic regression (LR) classifiers were used for performance assessment, with cross-validation and external validation conducted.
- Unsupervised clustering was utilized to investigate the separability of density grades based on selected features.
Main Results:
- Cross-validated AUCs for density grades A, B, C, and D were 0.909, 0.858, 0.927, and 0.890, respectively.
- An overall AUC of 0.936 was achieved for classifying nondense versus dense breasts.
- External validation yielded AUCs of 0.880 (A), 0.779 (B), 0.878 (C), 0.673 (D), and 0.823 for nondense/dense classification.
- Unsupervised clustering confirmed that the selected radiomic features effectively characterize different breast density grades.
Conclusions:
- The RFE-SHAP methodology demonstrates robustness in classifying breast tissue density using radiomic features from DBT.
- The selected features accurately capture the progression of breast density grades and generalize well to validation datasets.
- These findings support further research correlating selected radiomic features with clinical descriptors of breast tissue density.
