Development of a radiomic signature for prediction of pathological complete response following neoadjuvant chemoradiotherapy in esophageal cancer.

Abstract

106 Background: Pivotal to personalized therapy is accurate pre-treatment patient selection based on predictive biomarkers. We developed a radiomic signature model to predict pathological complete response (pCR) following neoadjuvant chemo-radiotherapy (NACRT) in esophageal cancers. Methods: We retrospectively identified 68 patients with esophageal cancer, treated with NACRT followed by esophagectomy, with baseline 18F-fluorodeoxyglucose Positron Emission Tomography (PET) and computed tomography (CT) scans performed at our institution. In-house data-characterization algorithm was used to extract 3D radiomics features from the segmented primary disease. Prediction models were constructed and internally validated (2/3rd training, 1/3rd validation) using machine learning software. Cox regression hazard used to assess association with relapse free survival (RFS). Results: pCR was observed in 34 (50%) patients. A total of 126 features were extracted individually from PET and CT scans. On univariate analysis, 16 PET and 14 CT features were found to independently predict pCR. Three separate predictive models were built for CT, PET and combined CT+PET. Individually, both CT and PET had a good predictive power with AUC 0.74 and 0.730, respectively. Combined CT+PET radiomic features resulted in a significant improvement of predicting power, with AUC 0.856. Good discrimination (82.6%) was observed in validation dataset. Final model was: 13.439-34.632(CT-eccentricity)-2.469(CT-LRLGE)+10.626(CT-Gnorm_HIE)-2.094(CT-SD_Lac2)-19.337(CT-FD_SD)-0.344(PET-Contrast), where eccentricity is shape feature, LRLGE a gray level run length feature, Gnorm_HIE gray-level size zone feature, SD_Lac2 a fractal dimension feature. Stratifying patients with low vs. high radiomic signature score predicted RFS: median RFS 80.1 vs. 40.1 months; and 2-year RFS 77.4% vs. 52.6% [Hazard Ratio: 2.24, 95% CI 1.01-5.58], respectively. Conclusions: The developed radiomics signature models were highly predictive of pCR following NACRT. Further validation in larger data sets is needed to determine whether the model can predict clinical outcomes.