A Spatiotemporal Explainable Model for Predicting Pathological Complete Response to Neoadjuvant Chemotherapy Using Breast DCE-MRI

Motivation: Current MRI-based deep learning approaches for predicting pathological complete response (pCR) often fail to fully leverage 3D spatial information and temporal relationships, limiting their predictive accuracy and interpretability. Goal(s): We propose ATVE-3DCNN, an Adaptive Thresholding and Visually Explainable 3D CNN，to improve pCR prediction and interpretability. Approach: We design multiscale modules to capture local and global feature in 3D space. Pre- and post-treatment features are combined to model temporal relationships. Finally, we utilize 3D Grad-CAM for interpretability. Results: Validated on a public dataset, our method achieve median AUC of 75.61%, accuracy of 75.41%, sensitivity of 70%, and specificity of 80.49%. Impact: This proposed method provides a more comprehensive understanding of the dynamic changes within the tumor, thereby improving the effectiveness of response assessment and offering valuable technical support for predicting neoadjuvant chemotherapy efficacy in breast cancer.