Abstract PS3-06-27: Deep Learning to Predict Pathological Complete Response in Patients Receiving NAC using Pre-treatment Clinical and Imaging Features

Abstract Background: Pathological complete response (pCR) after Neoadjuvant Chemotherapy (NAC) is a strong predictor of survival, making its early prediction clinically valuable. Current research is developing machine learning (ML) models that use pre- and post-treatment MRIs to predict pCR non-invasively. However, there is a critical need for models that rely solely on pre-treatment data to help assess if NAC is appropriate. While some ML models have been developed for this purpose, they have substantially lower accuracy compared to models that add post-treatment data. We address this gap by presenting a novel ML architecture that uses only pre-treatment MRIs while nearing the accuracy of methods that use post-treatment MRIs. Our method uses the pre-treatment MRI to predict radiomic features of the post-treatment MRI, which is then substituted for actual post-treatment imaging. We use the I-SPY2 dataset to train and validate our model. Methods: Our method is 3 steps. Step 1. A Convolutional Neural Network (CNN) is trained using both pre- and post-treatment MRIs. This CNN learns the important radiomic features in the MRI to predict pCR.Step 2. A second CNN is trained using the pre-treatment MRI to predict the radiomic features of the post-treatment MRI, essentially predicting the patient’s post-treatment MRI. The CNN predicts the changes in the radiomic features from the pre- to post-treatment MRI, as learned by the first CNN. Step 3. The predicted post-treatment radiomic features from Step 2 are substituted into Step 1, resulting in a model that does not need any post-treatment data. Results: Our method increases accuracy over a CNN that was trained using only pre-treatment MRIs while maintaining 87% of the accuracy obtained by the CNN that uses both treatment time-points. On the ISPY2 dataset, the Area Under the Curve (AUC) for the pre-treatment only CNN is 0.53, the pre- and post-treatment CNN is 0.71, and our proposed method is 0.62. If clinical data (molecular subtype and age) are added as inputs to the model, the AUC for the pre-treatment only CNN is 0.70, the pre- and post-treatment CNN is 0.78, and our proposed method is 0.73, a 3% increase over the model using only pre-treatment MRI. Conclusion: Our study presents a novel deep learning architecture that predicts pCR using only pre-treatment data with an accuracy of 3-9% higher than other pre-treatment methods. This method raises the AUC from 0.53 to 0.62, and up to 0.73 with clinical data, approaching the accuracy of models that use post-treatment data. By eliminating the need for post-treatment scans, our approach could help identify patients who might benefit from alternative or non-operative strategies early in their treatment. Leveraging the robust I-SPY2 dataset, our findings support the potential of radiomics-driven ML models to guide more personalized, less invasive care in breast cancer. Citation Format: R. Gifford, J. Hawley, C. Taylor, B. Griffith, A. Cubbison, S. Beyer, T. Andraos, R. Young, J. Eckstein, S. Jhawar, S. Krening. Deep Learning to Predict Pathological Complete Response in Patients Receiving NAC using Pre-treatment Clinical and Imaging Features abstract. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS3-06-27.