3373 - Using High-Repeatable Radiomic Features Improves the Cross-Institutional Generalization of Prognostic Model in Esophageal Cancer Receiving Definitive Chemoradiotherapy

J. Gong¹, Q. Wang², J. Cai³, and L. Zhao⁴; ¹Department of Radiation Oncology, Xijing Hospital, Air Force Medical University, Xian, China, ²Department of Radiation Oncology, Sichuan Cancer Hospital and Institution, Sichuan Cancer Center, School of Medicine, University of Electronic Science and Technology of China, Radiation Oncology Key Laboratory of Sichuan Province, Chengdu, China, ³Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong, China, ⁴Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, Xian, China

Purpose/Objective(s): To investigate the repeatability of radiomic feature (RF) via image perturbation and its impact on the cross-institutional generalizability of the prognostic model for predicting local recurrence-free survival (LRFS) and overall survival (OS) in esophageal cancer (EC) receiving definitive chemoradiotherapy (dCRT). Materials/

Methods: 792 patients from hospital 1 and 120 patients from hospital 2 were included in this study as the training and external validation set respectively. Image perturbations including spatial rotation and contour randomization were applied to contrast-enhanced computed tomography (CECT) to generate perturbed images. In total, 6,510 RFs were extracted from the original and the perturbed images separately and the repeatability of each RF was evaluated by the intraclass correlation coefficient (ICC). The volume-independent RFs were equally grouped into high-repeatable and low- repeatable RF groups by the median ICC, which were further analyzed separately by multi-step feature selection and multivariate Cox proportional hazards regression model construction for predicting LRFS or OS. The Concordance Index (C-index) and Kaplan-Meier analysis were used to evaluate the prognostic performance of the models.
Results: First-order statistical features were more repeatable than texture features (median ICC: 0.70 vs. 0.42-0.62). RFs from LoG filters had better repeatability than that of wavelet filters (median ICC: 0.70-0.84 vs. 0.14-0.64). Features with smaller bin width had higher repeatability (median ICC of 8-128: 0.65-0.47). 5,970 RFs with the volume correlation coefficient of less than 0.6 were equally grouped into high-repeatable and low- repeatable RF groups. For both LRFS and OS, the performance of the models based on high-repeatable RFs and low-repeatable RFs remained stable in the training set with similar C-index (LRFS: 0.65 vs. 0.67, P = 0.958; OS: 0.64 vs. 0.65, P = 0. 651), while the performance of the model based on low-repeatable group was significantly lower than that based on the high-repeatable group in the external validation set (C-index for LRFS: 0.61 vs. 0.67, P = 0.013; C-index for OS: 0. 56 vs. 0.63, P = 0. 013).
Conclusion: Applying high-repeatable RFs in modeling could safeguard the cross-institutional generalizability of prognostic model for helping clinical application in EC.