Weakly Supervised Training with Explainable Artificial Intelligence to Predict Breast-Cancer Response to Neoadjuvant Chemotherapy

Abstract

This study investigated the feasibility of weakly-supervised deep regression for predicting patient responses to neoadjuvant chemotherapy (NAC) using Maximum Intensity Projection (MIP) images. We used radiological tumor volume ratio (RTVR) and Residual Cancer Burden (RCB) to represent radiological and pathological responses to NAC, respectively. We conducted three experiments, two with single-task regression of RTVR and RCB and one with multi-task regression. Each experiment involved training a model based on a resnet14t architecture to minimize Batch Monte-Carlo (BMC) loss designed for imbalanced regression. We evaluated the performance of each model using Spearman’s correlation and Bland–Altman analysis. Spearman’s correlation coefficients were calculated for the hold-out test set and were ρ = 0.47 for the RTVR single-task model, ρ = 0.23 for the RCB single-task model, and ρ = 0.61 and ρ = 0.34 for RTVR and RCB respectively, in the multi-task model. Despite the multi-task model showing a slightly better correlation, we observed a statistically significant difference neither for predicting RTVR values (P = 0.49) nor for RCB scores (P = 0.55). Deep SHapley Additive exPlanations (SHAP) provided insight into the models’ decision-making processes. The results indicated that the current method could not provide clinically meaningful outputs. We discussed potential reasons for this poor performance and possible future research directions.