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

Area of Science:

• Biomedical Optics
• Computational Pathology
• Cancer Diagnostics

Background:

• Distinguishing luminal A and B breast cancer subtypes (LBCS) is crucial for treatment but challenging with routine histopathology.
• Ancillary testing is often required, highlighting the need for improved diagnostic methods.
• Mueller matrix polarimetry (MMP) analyzes polarized light interactions with tissue, offering a novel approach for breast cancer analysis.

Purpose of the Study:

• To evaluate the efficacy of Mueller matrix polarimetry (MMP) in differentiating luminal A and B breast cancer subtypes.
• To explore the potential of MMP combined with clinical parameters for improved breast cancer subtyping.
• To assess the diagnostic performance of machine learning models trained on polarimetric and clinical data.

Main Methods:

• Analysis of 30 polarimetric and 7 clinical parameters from 116 breast core biopsies classified by BluePrint® assay.
• Training of machine learning models (logistic regression, LDA, SVM, random forest, XGBoost) to distinguish LBCS.
• Receiver operating characteristic (ROC) curve analysis to evaluate model performance (AUC, accuracy, sensitivity, specificity).

Main Results:

• The top six biomarkers (3 polarimetric, 3 clinical) were identified using feature importance.
• A random forest model achieved 81% accuracy and 86% area under the ROC curve.
• The best model demonstrated 75% sensitivity and 75% specificity on an independent test set.

Conclusions:

• MMP, particularly selected Mueller matrix elements, shows potential for distinguishing LBCS when validated against BluePrint®.
• This optical approach may improve breast cancer prognosis by detecting subtle morphological differences.
• The findings suggest MMP can aid in guiding breast cancer treatment decisions.