MRI-based digital twins to improve treatment response of breast cancer by optimizing neoadjuvant chemotherapy regimens

Abstract

We developed a practical framework to construct digital twins for predicting and optimizing triple-negative breast cancer (TNBC) response to neoadjuvant chemotherapy (NAC). This study employed 105 TNBC patients from the ARTEMIS trial (NCT02276443, registered on 10/21/2014) who received Adriamycin/Cytoxan (A/C)-Taxol (T). Digital twins were established by calibrating a biology-based mathematical model to patient-specific MRI data, which accurately predicted pathological complete response (pCR) with an AUC of 0.82. We then used each patient's twin to theoretically optimize outcome by identifying their optimal A/C-T schedule from 128 options. The patient-specifically optimized treatment yielded a significant improvement in pCR rate of 20.95-24.76%. Retrospective validation was conducted by virtually treating the twins with AC-T schedules from historical trials and obtaining identical observations on outcomes: bi-weekly A/C-T outperforms tri-weekly A/C-T, and weekly/bi-weekly T outperforms tri-weekly T. This proof-of-principle study demonstrates that our digital twin framework provides a practical methodology to identify patient-specific TNBC treatment schedules.

Fig. 1. Overview of digital twin-based optimization of TNBC response to NAC.

For individual TNBC patients (a), longitudinal MRIs (b) are collected pre- and early in the course of NAC, and then used to calibrate a biology-based mathematical model of tumor growth and treatment response, thereby establishing digital twins (c). The digital twins capture patient-specific tumor characteristics as well as response behaviors, so they can be used to systematically optimize NAC schedules (d). This digital twin-based optimization schema has the potential to improve pCR rates of TNBC (e, top), and to identify opportunities for therapy de-escalation (e, bottom) on a patient-specific basis.

Fig. 2. Using the patient-specific digital twins to predict response to various therapeutic schedules.

The framework (a) consists of data preparation (i.e., image collection and processing pipeline), model personalization (i.e., calibration of the biology-based model using longitudinal MRI data), and response prediction. The prediction accuracy was evaluated by comparing the predicted final pathological status to the actual pathological status via ROC analysis (b). Applying both the actual schedule (navy curves in (c)) and the alternative schedule (red curves in (c)) of NAC to the digital twin allows for predicting the dynamics of tumor response (measured by the change of tumor volume over time; d) to each intervention. In this illustrative case, the actual therapeutic schedule leads to a predicted tumor volume after NAC (TV_T) larger than the pCR/non-pCR differentiating threshold (TV_T,J, determined from the optimal cutoff of ROC; see “Methods” section “Establishment of patient-specific digital twin to predict TNBC response to NAC” for details); thus, we predicted this patient as a non-pCR. In contrast, the alterative schedule leads to a predicted TV_T less than the TV_T,J; thus, we predicted the alternative schedule would lead to a pCR for this patient.

Fig. 3. Illustration of the three optimization strategies.

In the multi-step optimization strategy (a), patient-specific response is predicted for all candidate A/C schedules (light blue area represents the range of tumor volume, TV, predicted from various A/C schedules), with the schedule yielding the minimized TV_A/C identified as the optimal A/C schedule. Based on the minimized TV_A/C, patient-specific response is then predicted for all candidate T schedules (light green area), and the one yielding the minimal TV_T is identified as the optimal T schedule. Together, the optimal A/C and T schedules form the optimal NAC schedule (red area). In the simultaneous optimization strategy (b), patient-specific response is predicted for all candidate A/C and T schedules, with the one yielding the minimal TV_T identified as the optimal NAC schedule (red area). In the midway optimization strategy (c), the A/C schedule is fixed to the actual schedule (black curve). Based on the TV achieved from the actual A/C schedule, patient-specific response is predicted to all candidate T schedules, with the one yielding the minimal TV_T identified as the optimal T schedule.

Fig. 4. Outcome from three optimization strategies in three example patients.

a–c and d–f show the results of first (i.e., A/C optimization) and second (i.e., T optimization) steps, respectively, of the multi-step optimization for each patient. In (a–c), the blue curves represent the outcomes from all candidate A/C schedules, with the black and red dots indicating the actual and the optimized outcomes, respectively. In (d–f), the light green curves/surfaces represent the outcomes of the candidate T schedules with 12 cycles, the navy curves/surfaces represent the outcomes of the candidate T schedules with 4 cycles, and the gray flat plane depicts the threshold of residual tumor volume (TV_T,J) for differentiating pCR from non-pCR. g–i show the results of the simultaneous optimization in the same three patients, in which the light green surfaces represent the outcomes from all A/C and T candidates with 12 cycles, and the navy surfaces represent the outcomes from all A/C and T candidates with 4 cycles. j–l show the results of the midway optimization, in which the light green curves represent the outcomes from the candidate T schedules with 12 cycles, and the navy curves represent the outcomes from the candidate T schedules with 4 cycles. Patient 1 had a non-pCR with their actual treatment, and optimization of therapeutic schedule suggests this could be improved to pCR. Patient 2 had a pCR with their actual treatment, and optimization of schedule suggests an opportunity for de-escalation. Patient 3 had a non-pCR either with the actual or optimized therapeutic schedules.

Fig. 5. Treatment escalation and de-escalation opportunities through patient-specific, multi-step optimization.

a compares the duration of A/C that the individual patient actually received (horizontal axis) versus the duration optimized for the patient (vertical axis). Similarly, (b) compares the duration of Taxol that the individual patient actually received, and the duration optimized for the patient. Each gray dot represents one patient. Note that when multiple schedules lead to the optimal outcome for the patient (like Patient 2 in Fig. 4b), the “optimized duration” is shown as the longest one among the optimal schedules. The dots in the light blue areas indicate patients who had the opportunity for treatment de-escalation, whereas the dots in the light red areas indicate patients had the opportunity for treatment escalation. Dots falling on the diagonal indicate no change in the treatment schedules was suggested for the corresponding patients.

Fig. 6. Treatment outcomes with and without optimization.

a–d shows the improvement in outcomes for the whole cohort (n = 105) with the first step of multi-step optimization, second step of multi-step optimization, simultaneous optimization, and midway optimization, respectively. In each panel, the light blue and light red violin plots depict the distribution of predicted outcomes with the actual schedule and the optimized schedule, respectively. The boxplots show the median (red line) and IQR (blue box) of the corresponding outcomes. All three optimization strategies lead to a significant reduction of TV_T when compared to the treatment schedule the patient actually received.