266. extension of fdg-pet-based radiomic models on identification of patients with residual esophageal cancer after neoadjuvant chemoradiotherapy

Abstract For esophageal cancer patients, detection of residual tumor after neoadjuvant chemoradiotherapy (nCRT) is important for guiding treatment decisions regarding standard esophagectomy or active surveillance. Prediction models incorporating clinical variables and FDG-PET quantitative radiomic features have been previously developed in a single-center setting (‘development cohort’, Beukinga et al. 2018) to identify patients with residual tumor after nCRT. The aim was to improve generalizability by redeveloping the models (i.e. ‘model extension’) using a multicenter prospectively acquired cohort. The development cohort was combined with a patient cohort collected from a prospective multicenter study in four Dutch institutes. Patients underwent nCRT according to the CROSS regimen followed by esophagectomy 6–14 weeks after nCRT between 2013–2019. Outcome was tumor regression grade (TRG) 1–2 (0–10% tumor) vs TRG 3–4 (>10% tumor). FDG-PET/CT was performed 6–12 weeks after nCRT. Post-treatment FDG-PET-radiomic features were calculated from gross tumor volumes. Features were standardized per scanner. A bootstrapped least absolute shrinkage and selection (LASSO) model using clinical variables and radiomic features was developed and a model incorporating clinical variables was developed for reference. Some 262 patients were included. Median age was 65 years (IQR 60–70), 158 of 262 were male (82%) and outcomes were 129/262 TRG 1–2 (49%) and 133/262 TRG 3–4 (51%). The extended LASSO model included seven radiomic features and the clinical variables cT stage and histology (internally validated AUC, 0.59). At a probability threshold to obtain 90% sensitivity for detection of TRG 3–4 (benchmark with bite-on-bite biopsies + fine-needle aspiration of suspected lymph nodes), specificity was 36% and accuracy 63%. A LASSO clinical reference model (including cT, sex, histology) achieved an internally validated AUC of 0.58 (Fig. 1). The extended internally validated LASSO model incorporating post-treatment radiomic features and clinical variables yielded moderate discriminative ability. The improvement in predictive performance with this model compared with a model using only clinical variables was marginal. The application of radiomics to post-treatment FDG-PET/CT scans is of no help in decision-making in individual patients regarding the choice for active surveillance after nCRT.