Machine learning-based prediction of luminal breast cancer subtypes using polarised light microscopy.

Abstract

Background: Routine histopathology cannot distinguish between clinically diverse luminal A and B breast cancer subtypes (LBCS), often requiring ancillary testing. Mueller matrix polarimetry (MMP) offers a promising approach by analysing polarised light interactions with complex breast tissues. This study explores the efficacy of using MMP for luminal subtype differentiation.

Methods: We analysed 30 polarimetric and 7 clinical parameters from 116 unstained breast core biopsies, LBCS classified using the BluePrint® molecular assay. These features were used to train various machine learning models: logistic regression, linear discriminant analysis, support vector machine, random forest, and XGBoost to distinguish luminal subtypes. Receiver operating characteristic curve (ROC) analysis was used to each to assess diagnostic performance using area under the curve, accuracy, sensitivity, and specificity.

Results: Using the top six most prognostic polarimetric (three) and clinical (three) biomarkers ranked by feature importance, the best-performing random forest model achieved an accuracy of 81% (area under ROC = 86%), with both sensitivity and specificity at 75% on an unseen test set, indicating moderately promising, clinically informative performance.

Conclusions: MMP, particularly its selected Mueller matrix elements, combined with clinical biomarkers show promise in distinguishing LBCS as validated against BluePrint®. By detecting subtle differences in tissue morphology, this approach may enhance breast cancer prognosis and help guide treatment decisions.