MRI-BASED RADIOMICS FOR PREDICTING NEOADJUVANT CHEMOTHERAPY RESPONSE IN BREAST CANCER USING MACHINE LEARNING AND AI

Abstract

Current prediction methods for pathological complete response (pCR) after neoadjuvant chemotherapy (NACT) for breast cancer are limited to post-surgical histopathological assessment of the tumor and doesn’t give actionable advice during the treatment, remaining a clinically unsolved challenge. Current MRI radiomic models almost exclusively use the features from a single pre-treatment MRI, ignoring the dynamic changes in morphology and textural features within the tumor microenvironment that occur over multiple cycles of chemotherapy. This paper proposes a novel framework named longitudinal Delta-Radiomic Hybrid Intelligence (DRHI) to represent a change-based radiomic signature across three serial MRI timepoints baseline (T0), post-cycle 1 (T1), and post-cycle 2 (T2) and combines it with deep convolutional features learned from a pre-trained ResNet-50 backbone. The prediction models are developed independently for three molecular subtypes: HER2-enriched, Triple-Negative (TNBC) and Hormone Receptor-positive/HER2-negative (HR+/HER2−) instead of using a single generalized model for all kinds of patients and their biological heterogeneity. To bridge this interpretability, disconnect that hinders the adoption of AI-powered oncology tools, a SHAP based explainability module is integrated into each of the subtype pipelines to explain the influential features per prediction. Experiments are conducted on the publicly available I-SPY 2 TRIAL dataset. To our knowledge, there is no previous study that involves using a unified prediction framework to combine all these four aspects, namely, longitudinal delta radiomics, hybrid handcrafted-deep feature fusion, molecular subtype stratification, and SHAP explainability. The proposed approach shows clinically relevant gain in AUC for all three subtypes on top of the existing state-of-the-art baseline methods.