MRI-Based Digital Models Forecast Patient-Specific Treatment Responses to Neoadjuvant Chemotherapy in Triple-Negative Breast Cancer

Abstract

Triple-negative breast cancer (TNBC) is persistently refractory to therapy, and methods to improve targeting and evaluation of responses to therapy in this disease are needed. Here, we integrate quantitative MRI data with biologically based mathematical modeling to accurately predict the response of TNBC to neoadjuvant systemic therapy (NAST) on an individual basis. Specifically, 56 patients with TNBC enrolled in the ARTEMIS trial (NCT02276443) underwent standard-of-care doxorubicin/cyclophosphamide (A/C) and then paclitaxel for NAST, where dynamic contrast-enhanced MRI and diffusion-weighted MRI were acquired before treatment and after two and four cycles of A/C. A biologically based model was established to characterize tumor cell movement, proliferation, and treatment-induced cell death. Two evaluation frameworks were investigated using: (i) images acquired before and after two cycles of A/C for calibration and predicting tumor status after A/C, and (ii) images acquired before, after two cycles, and after four cycles of A/C for calibration and predicting response following NAST. For Framework 1, the concordance correlation coefficients between the predicted and measured patient-specific, post-A/C changes in tumor cellularity and volume were 0.95 and 0.94, respectively. For Framework 2, the biologically based model achieved an area under the receiver operator characteristic curve of 0.89 (sensitivity/specificity = 0.72/0.95) for differentiating pathological complete response (pCR) from non-pCR, which is statistically superior (P < 0.05) to the value of 0.78 (sensitivity/specificity = 0.72/0.79) achieved by tumor volume measured after four cycles of A/C. Overall, this model successfully captured patient-specific, spatiotemporal dynamics of TNBC response to NAST, providing highly accurate predictions of NAST response.

Significance: Integrating MRI data with biologically based mathematical modeling successfully predicts breast cancer response to chemotherapy, suggesting digital twins could facilitate a paradigm shift from simply assessing response to predicting and optimizing therapeutic efficacy.

Figure 1.. The two Frameworks to predict patient-specific response to NAST.

Panel A shows the timeline of treatment administration and data acquisition for each patient. Panel B illustrates the processing-modeling pipeline to generate patient-specific digital twins. Two frameworks are established to evaluate the predictive ability of the digital twins. Framework 1 (Panel C) employs digital twins to predict the outcome of the doxorubicin and cyclophosphamide (A/C) regimen. Patient-specific images from visits 1 (V1) and 2 (V2) along with the schedule of A/C provide the input to which the digital twin is calibrated. Once calibrated, the digital twin outputs a prediction of the spatiotemporal development of the tumor in response to A/C. The prediction is then directly compared to the V3 images. Framework 2 (Panel D) employs digital twins to predict the outcome of the entire NAST. Images from V1, V2, and V3, along with the schedule of both A/C and paclitaxel, are given as input, and the digital twin outputs a prediction of whether the patient will achieve a pCR. The prediction is then directly compared to the post-surgical pathological response.

Figure 2.. Flow charts of MRI data processing.

Panel A shows an example set of DW-MRI and DCE-MRI data acquired at one visit of a patient. Panels B-D illustrates the three steps of the processing pipeline, respectively. In panel B, multiparametric images are trimmed to the same field-of-view (FOV) and registered. In panel C, images from V1 and V3 are registered to those from V2. In panel D, tissue segmentation and calculation of tumor cellularity (from the DW-MRI data) are performed. These steps prepare the data for calibration with the biologically-based mathematical model and establishing each patient’s digital twin.

Figure 3.. Temporal accuracy of patient-specific predictions of the response of TNBC to A/C.

Panels A and B show the time courses of calibrated therapeutic efficacies over the A/C regimen in two representative patients, respectively. Panels C and D present the temporal dynamics predicted by the digital twins of the same two patients, in which subpanels (i) and (ii) represent the change of tumor cellularity (TTC) and tumor volume (TTV), respectively. In each panel, red circles present the measured TTC or TTV at certain time points, while blue curves and shadows present the predicted median and range of dynamics, respectively. Very small differences are observed between the measured and predicted changes of TTC and TTV over time in the example patients. Panels E and F plot the correlation between the measured and predicted changes of TTC and TTV (CCC = 0.95 and 0.94), respectively, in the cohort. These results indicate high precision and accuracy for predicting patient-specific temporal dynamics of TNBC in response to A/C.

Figure 4.. Spatial accuracy of patient-specific predictions of the response of TNBC to A/C.

Panels A and B show the measured and predicted tumor cell distributions on the central tumor slice for two patients. Panels C and D show the 3D renderings of the measured and predicted change of those two tumor shapes. Very small differences are observed between the measured and predicted tumor cell distributions or tumor shapes in the patients. Panel E presents the difference between the measured and predicted change of tumor cell distributions in the cohort. The median (red circle) and interquartile range (blue bar) of difference within each patient’s tumor region are presented. The difference across all patients has a mean (95% CI) of 0.20% (−20.35% – 20.75%). These results indicate a high accuracy of the digital twins for predicting patient-specific spatial dynamics of TNBC in response to A/C.

Figure 5.. Accuracy of patient-specific prediction of final pathological response.

Panels A and B show the time courses of calibrated therapeutic efficacies during NAST in two example patients, respectively. Panels C and D present the temporal dynamics predicted by the digital twins for the same two example patients, in which subpanels (i) and (ii) represent the change of tumor cellularity (TTC) and tumor volume (TTV), respectively. Panel E presents the ROC analysis of differentiating pCR from non-pCR based on predicted (blue) and measured (red) TTC; Similarly, panel F presents the ROC analysis based on predicted and measured TTV. The larger AUCs of blue curves compared to red curves in both panels E and F indicates superior accuracy for predicting final pathological response.