Integrated CT and 18F-FDG PET Based Radiomic Prediction Models for Pathological Response and Outcome in Esophageal Cancer Patients Treated with Neoadjuvant Chemoradiotherapy and Esophagectomy

Abstract

Pre-treatment computed tomography (CT) and 18F-fluorodeoxyglucose Positron Emission Tomography (PET) based radiomics exploratory analysis to investigate potential imaging biomarkers predictive of pathological response and clinical outcome in esophageal cancer patients treated with neoadjuvant chemo-radiotherapy (NACRT). Pre-treatment CT and PET images of 68 operable esophageal cancer patients treated with NACRT followed by esophagectomy were analyzed. A total of 126 features were extracted from the segmented primary disease individually from PET and CT images using an in-house algorithm. Radiomic features were compared with pathological response using tumor regression grade (TRG) and clinical parameters. Primary model were built using logistic regression, and similar models were also built and validated using random forest and libSVM methods. A composite feature-based signature formula, α PETf+ (1-α) CTf, where constant α (0<α<1) was optimized for each individual predictive model. Leave-one-out validation (LOOV) was performed. Locoregional control (LRC), recurrence free survival (RFS), metastasis free survival (MFS), and overall survival (OS) were estimated by Kaplan-Meier analysis and compared using log-rank test. Pathological complete response (pCR) was observed in 34 (50%) patients. Estimated actuarial 5-year RFS, LRC, MFS, and OS were 56.6% (mean: 67.8, 95% CI 53.7-81.8 months), 83.1% (mean: 92.8, 95% CI 81.4-104.2 months), 64.9% (mean: 76.4, 95% CI 62.2-90.7 months), 39.5% (median: 40, 95% CI 24.59-55.40), respectively. Near complete response (TRG 0-1) was associated with improved LRC [p = 0.04; Hazard Ratio (HR) 4.24 (1.01-19.01)]. Three separate predictive models were built for CT, PET, and combined CT+PET and analyzed separately for clinical endpoints. Integrated CT+PET features resulted in best predictive power. For pCR prediction, the signature used composite features with optimized α=0.8 resulted in AUC=0.87 and 76.5% accuracy in LOOV. For loco-regional recurrence, α=0.7, AUC=0.90 and 89.7% accuracy and for distant recurrence, α=0.4, AUC=0.83, and 83.8% accuracy in LOOV. For overall recurrence, α=0.5, AUC=0.80 and 73.5% accuracy in LOOV. For survival, α=0.7, AUC=0.85, accuracy 75% in LOOV. Stratifying patients by median radiomic signature predicted RFS: 80.1 vs. 40.1 months; 2-year RFS 77.4% vs. 52.6% (HR: 2.24 , 95% CI 1.01-5.58), respectively. The developed pre-treatment composite radiomic signatures were highly predictive of pCR, loco-regional recurrence, metastasis, and overall survival following NACRT plus esophagectomy. This is provocative data and further prospective validation study is in progress.