Abstract

More than 50% of triple negative breast cancer (TNBC) patients do not respond well to the standard-of-care neoadjuvant therapy (NAT). Therefore, methods capable of predicting treatment response will be highly useful to optimize intervention and outcomes for TNBC patients. To address this problem, we aim to integrate quantitative magnetic resonance imaging (MRI) with biology-based mathematical modeling and deep learning to make patient-specific predictions of TNBC response to NAT using only pretreatment data. TNBC patients (n = 150) enrolled in the ARTEMIS trial (NCT02276443) received doxorubicin/cyclophosphamide (A/C) followed by paclitaxel. MRI exams were acquired for each patient at the following timepoints: (1) before initiation of NAT, (2) after two A/C cycles, (3) after four A/C cycles, and (4) at the conclusion of NAT. Using patient-specific MRI data from the first two exams, we calibrated a biology-based mathematical model to characterize migration, proliferation, and treatment-induced death of tumor cells. We then used this model as a digital twin to predict spatiotemporal tumor response. While effective, this approach requires the patient to have completed at least part of their NAT regime before we are able to predict therapeutic response. To relax this requirement, we have developed an approach that combines deep learning and biology-based mathematical modeling to predict the response of TNBC to NAT before treatment initiation. Specifically, we integrated a U-Net-based convolutional neural network with our mathematical model to regress between pre-treatment data and the model parameters obtained from a training set. Using parameters from learning a network with a subset of 68 patients, our mathematical model yielded concordance correlation coefficients between the predicted and measured patient-specific changes in tumor cellularity and volume at the third imaging point of 0.95 and 0.92, respectively. Spatially, we obtain the median difference between predicted and measured percent change in cellularity from visit one to visit three for each patient, giving a mean (95% confidence interval) of -6.51% (-7.13%, -5.90%) across all patients. These encouraging results may be further improved using methods such as expanding to a spatially-resolved proliferation rate, including genetic and/or histological data, and extending the deep learning framework to the end of the treatment course to predict pathological response. This approach allows us to obtain patient-specific predictions of response before NAT commences, thereby providing the opportunity to optimize interventions and patient outcomes. Citation Format: Casey E. Stowers, Chengyue Wu, Sidharth Kumar, Ernesto A.B.F. Lima, Xhan Zu, Clinton Yam, Jong Bum Son, Jingfei Ma, Jonathan I. Tamir, Gaiane M. Rauch, Thomas E. Yankeelov. Developing MRI-based digital-twins via mathematical modeling and deep learning to predict the response of triple-negative breast cancer to neoadjuvant therapy [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 845.