Abstract No. 
2020 San Antonio Breast Cancer Symposium
Dec 8-11

Serial MRI and pathology combined to select candidates for therapy de-escalation in the I-SPY 2 TRIAL

Venters SJ, Li W, Wolf DM, Carter JM, Klein ME, Singh K, Rabe K, I Ocal T, Newitt D, Yau C, Onishi N, Gibbs J, Sahoo S, Harada S, Khazai L, Harigopal M, Borowsky AD, Krings G, Balassanian R, Chen Y-Y, Cole K, Shad S, LeStage B, Delson A, Finestone S, Brown-Swigart L, I-SPY 2 Imaging Working Group, I-SPY 2 TRIAL Consortium, Esserman L, van ‘t Veer L, Symmans WF, Hylton NM

Background: The I-SPY 2 TRIAL, open to patients with locally advanced, molecular high-risk breast cancer, aims to bring each patient to pathologic complete response (pCR) with a minimum of toxicity.  Here we test the hypothesis that imaging (MR volume predictors) combined with core biopsy may be used to accurately select candidates who show early response and provide an option of treatment de-escalation at mid-therapy (12 weeks).

Methods: 87 I-SPY 2 patients with core biopsies at the inter-regimen time point (~12 weeks through treatment), pCR data, and serial MR images were considered in this study. Eleven I-SPY 2 TRIAL pathologists independently provided a digital assessment of the presence or absence of residual invasive cancer from H&E stained, and any requested ancillary IHC, images from imaging-guided core biopsies. Pathology predicts pCR if there is a consensus of no invasive residual disease. We generated predictions for all (55) unique pairs over the 11 pathologists, where pCR is predicted if both pathologists find no invasive cells. MRI pCR prediction models were previously developed on an independent dataset of ~990 I-SPY 2 patients, and applied to this cohort. Volume-based prediction models were previously optimized within each subtype and predicted probability thresholds were selected over a range of positive predictive value (PPV). In this study, MR predicts pCR (positive test) if the predicted probability is above a threshold that yields a given PPV value.  For each pathologist pair, we combined pathology-based and MR-based predictors into a predictive-RCB (pre-RCB); and pre-RCB predicts a patient as pCR (RCB0) if both MR and pathology predicts pCR.  Predictive performance is assessed by calculating the mean and range of PPV and sensitivity.

Results: 39% (34/87) of the patients in this study achieved pCR.  Over all pairs of pathologists, on average 80% of pathology-only predicted pCRs were true pCRs (mean PPV = 80% [range: 69-92%]), and 74% of patients who achieved pCR were predicted pCR by pathology alone (mean sensitivity = 74% [65-82%]).  We assessed combinations with MR probability thresholds at PPV levels 50%-70%; and observed the best balance of PPV and sensitivity for the pre-RCB when MR thresholds were set at 50% PPV level. At this threshold setting, the pre-RCB achieved a PPV = 92% [83-100%], meaning on average 92% of predicted pCRs were true pCRs, and this improvement in positive predictive performance over pathology alone is achieved with a lower but still-reasonable 53% sensitivity [33-62%].  

Conclusion: pre-RCB, which predicts a patient as pCR if both MR and inter-regimen pathology predicts pCR, provides clinically actionable accuracy for treatment de-escalation for early responders (PPV>90%).   Adding a final MR review at the time of early surgery may further improve performance.

View original