Predicting Pathologic Response to Neoadjuvant Chemotherapy in Breast Cancer Using Machine Learning of MRI Radiomics.

OBJECTIVE

To develop and validate a machine learning-based MRI radiomics model for predicting the risk of non-pathological complete response (non-pCR) following neoadjuvant chemotherapy in breast cancer, to identify independent influencing factors.

METHODS

A total of 400 consecutive breast cancer patients were randomly divided into training and validation sets at a 7:3 ratio. Baseline clinicopathological variables, functional magnetic resonance imaging (MRI) parameters, tumor-infiltrating lymphocytes (TILs), and longitudinal MRI radiomics scores were collected. Variables were selected using univariate analysis and least absolute shrinkage and selection operator (LASSO) regression. Random forest, support vector machine, and logistic regression models were constructed and evaluated using area under the curve (AUC), calibration, decision curve analysis, and SHapley Additive exPlanations (SHAP) interpretation.

RESULTS

Multivariate logistic regression confirmed all six factors (Ki-67 proliferation index, maximum tumor diameter, apparent diffusion coefficient (ADC) value, initial enhancement rate, tumor-infiltrating lymphocytes (TILs), and radiomics score) as independent predictors of non-pCR (P < 0.05). The random forest model showed the numerically highest AUC, with an AUC of 0.835 [95% confidence interval(CI): 0.767-0.904] in the training set and 0.745 (95% CI: 0.597-0.893) in the validation set, outperforming logistic regression (training: 0.805; validation: 0.717) and support vector machine (training: 0.792; validation: 0.702). SHAP analysis revealed that feature contributions, in descending order, were as follows: radiomics score, ADC value, initial enhancement rate, TILs, Ki-67 proliferation index, and maximum tumor diameter. SHAP analysis identified radiomics score, ADC value, and initial enhancement rate as the leading contributors.

CONCLUSIONS

The random forest model constructed in this study, integrating MRI radiomics and functional imaging parameters, exhibits acceptable discrimination and calibration for predicting pathological response to neoadjuvant chemotherapy in breast cancer.