Comparative evaluation of feature-selection strategies for machine learning-based histological grading of breast cancer using tumor and fibroglandular-tissue DCE-MRI features.
Authors
Affiliations (5)
Affiliations (5)
- Professor of Radiology, President of the Iranian Society of Radiology, Tehran, Iran.
- Department of Medical Physics, School of Medicine Isfahan University of Medical Sciences, Isfahan, Iran. Electronic address: [email protected].
- Department of Radiology, School of Medicine, Shahid Beheshti University of Medical Sciences, Tehran, Iran. Electronic address: [email protected].
- Resident Of Radiology, School of Medicine, Department of Radiology, Taleghani Hospital, Shahid beheshti University of Medical Sciences, Tehran, Iran.
- Department of Radiology, School of Medicine, Shahid Beheshti University of Medical Sciences, Tehran, Iran.
Abstract
Nottingham histological grading is central to breast cancer prognosis and treatment planning, but conventional pathological assessment is labor-intensive and subject to inter-observer variability. Radiomics and machine learning may support noninvasive preoperative grade prediction. To compare mutual-information, chi-squared, and LASSO feature-selection strategies for binary Nottingham grade classification using tumor- and fibroglandular-tissue (FGT)-derived DCE-MRI features. This retrospective secondary analysis included 576 patients from the Duke-Breast-Cancer-MRI collection in The Cancer Imaging Archive, comprising 288 non-high-grade and 288 high-grade tumors. A total of 529 publicly released precomputed imaging features across 10 categories were analyzed, including volume, morphology, enhancement kinetics, texture, spatial heterogeneity, and temporal variation. No additional segmentation, resampling, normalization, or feature extraction was performed. Feature selection was followed by classifier development and Bayesian hyperparameter optimization. Models were evaluated using stratified 5-fold cross-validation, and SHAP analysis was used for feature interpretation. Because feature selection and model configuration were not embedded within a fully nested cross-validation framework, performance estimates were considered exploratory. Chi-squared selection with a neural network achieved the highest mean accuracy (74.78%; AUC 0.77). LASSO provided the strongest discrimination, with a pooled out-of-fold AUC of 0.82 and relatively balanced sensitivity and specificity. Tumor volumetric, morphological, and enhancement-kinetic features were the most influential predictors, while FGT-derived features contributed complementary but more limited information. Tumor- and FGT-derived DCE-MRI features showed preliminary potential for binary Nottingham grade classification. Fully nested model development and independent multi-institutional validation are required before clinical application.