Decoding breast cancer Histopathology: Machine Learning-Enhanced Advanced Mathematical Models in Multi-B-Value MR Diffusion Imaging

Abstract

Abstract This study aims to advance breast cancer (BC) subtype classification by employing machine learning algorithms to identify key diffusion parameters from apparent diffusion coefficient (ADC0-800) histogram, intravoxel incoherent motion (IVIM), and diffusion kurtosis imaging (DKI). A total of 148 newly diagnosed BC patients were enrolled, confirmed through core needle biopsy, and subjected to histopathological analyses to assess key molecular markers such as estrogen and progesterone receptors (collectively termed hormone receptors), HER2, and ki67. These markers were then used to classify BC subtypes. Utilizing advanced post-processing techniques on multi-b-value MR Images, the study employed a diverse set of machine learning (ML) algorithms (supervised, unsupervised, and deep learning techniques) to quantitatively assess their diagnostic utility and subsequently identify algorithmically refined diffusion signatures. Machine learning algorithms demonstrated varying efficacies in the classification of BC subtypes. Key diffusion parameters were prioritized based on feature importance values from the ML models with the highest mean AUC and were further validated using group comparison tests and univariate logistic regression. In conclusion, our findings underscore the importance of tailored ML algorithms in classifying BC subtypes and advocate for a synergistic approach in personalized oncology and precision medicine.