Abstract 2804: Impact of deformable registration methods for prediction of treatment response to neo-adjuvant chemotherapy in breast cancer

Abstract

Abstract Purpose: This study aimed to evaluate the effect of the registration accuracy of several deformable registration methods on the predictive value of radiomic features to model recurrence-free survival (RFS) in patients undergoing neoadjuvant chemotherapy (NAC) for breast cancer. Methods: From the I-SPY 1 cohort, 130 patients had clinical data and imaging data from the first two visits available, including 38 events (death or recurrence). We computed voxel-wise kinetic maps (peak enhancement, wash-in slope, wash-out slope, and signal enhancement ratio) from both pre-treatment and early-treatment MR images. For each of six different deformable registration methods (ANTs, DRAMMS, ART, NiftyReg, NMI-FFD, and SSD-FFD), we calculated the transformation field and used it to warp the kinetic maps obtained from early treatment MR images. Using these, for each kinetic feature, for each registration method, the parametric response map (PRM) at each voxel computed the difference between the warped kinetic feature and the kinetic feature from the pre-treatment image. We extracted 104 radiomic texture features from each PRM kinetic map, using the CAPTK toolkit, applied principal component (PC) analysis to the 104-dimensional feature vector, and retained the first four PCs for modeling (one covariate for every 10 events). We modeled RFS via Cox proportional hazards, comparing eight models: 1) baseline covariates of age, race, and hormone receptor status (model F1); 2) the covariates in model F1 plus functional tumor volume at the early-treatment visit (FTV2) (model F2); 3-8) the covariates in F2 with the addition of the radiomic feature PCs derived from each registration method. We evaluated model predictive performance using the C-statistic, and model fit via Kaplan-Meier plots and the log-rank test. Results: The baseline model (model F1) provided a C-statistic of 0.54, and model F2 gave 0.66. Among the automated registration methods, the F2+ANTs model had the highest performance with a C-statistic of 0.72. F2+DRAMMS gave 0.70, followed by F2+NiftyReg (0.68), F2+NMI-FFD (0.67), and F2+SSD-FFD (0.67). F2+ART had the lowest performance with a C-statistic of 0.66. The Kaplan-Meier curve for model F2 (baseline + FTV2) gave p = 0.0013 for separation between patients above and below median hazard as compared to the model F1 (p = 0.31). Including FTV2 and radiomic features, all models yielded p < 0.001 except the F2+ART model. Conclusion: The radiomic features of PRM maps derived from warping the DCE-MRI kinetic maps using the ANTs registration method significantly improved early prediction of survival during NAC as compared to other registration methods. Citation Format: Snekha Thakran, Eric Cohen, Nariman Jahani, Susan P. Weinstein, Lauren Pantalone, Nola Hylton, David Newitt, Christos Davatzikos, Despina Kontos. Impact of deformable registration methods for prediction of treatment response to neo-adjuvant chemotherapy in breast cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 2804.