Feature-selection methods for MRI-based breast cancer grade prediction compared

Dynamic contrast-enhanced MRI radiomics showed preliminary potential for predicting Nottingham grade in 576 breast cancers; LASSO selection gave pooled out-of-fold AUC 0.82, while chi-squared+neural net achieved 74.78% accuracy (AUC 0.77); tumor features mattered most.

• Retrospective secondary analysis of 576 patients (288 high-grade, 288 non-high-grade) from the Duke-Breast-Cancer-MRI collection in The Cancer Imaging Archive.
• Chi-squared selection with a neural network achieved highest mean accuracy 74.78% (AUC 0.77); LASSO gave strongest discrimination with pooled out-of-fold AUC 0.82 and relatively balanced sensitivity/specificity.
• Tumor volumetric, morphological, and enhancement-kinetic features were most influential; FGT-derived features added limited complementary information. Performance is exploratory because feature selection and model configuration were not nested in cross-validation.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.