ESTRO 2021 Abstract Book

S719

ESTRO 2021

Conclusion This study evaluates our DL dose prediction model in a broader patient referral context and demonstrates its ability to be used for clinical decisions. Future research should focus on the detection of outliers to reduce the impact of dose prediction error on the clinical decision for those patients. PD-0882 PET radiomics for predicting PFS in patients with esophageal cancer who are treated with CRT N. Takahashi 1 , R. Umezawa 1 , T. Yamamoto 1 , K. Takeda 1 , Y. Ishikawa 1 , Y. Suzuki 1 , K. Kishida 1 , S. Teramura 1 , K. Jingu 1 1 Tohoku University Graduate School of Medicine, Radiation Oncology, Sendai, Japan Purpose or Objective The aim of this study was to determine whether radiomic parameters are predictors of progression-free survival (PFS) in patients with esophageal cancer who are treated with definitive chemoradiotherapy (dCRT). Materials and Methods Patients with stage II – III esophageal cancer who underwent FDG-PET/CT within 45 days before dCRT between 2005 and 2014 were included in this study. There were 107 patients and they were randomly assigned to a training set (71 patients) and a validation set (36 patients). Pretreatment FDG-PET/CT was used and radiomic parameters inside the area of FDG standard uptake value ≥ 3 were calculated. The open-source software 3D slicer v.4.10.2 was used for segmentation and calculating radiomic parameters. In the training set, radiomic parameters were divided by their median values. Predictors of PFS including radiomic parameters and general information (gender, age, T stage, N stage and total radiation dose) were investigated by Cox hazards model. Correlations among significant parameters in univariate analysis were checked and parameters without strong correlations were included in multivariate analysis. In the validation set, significant parameters in the raining set were applied to Kaplan-Meier curves. The median value in training set was used as a cut-off value in the validation set. JMP pro v.15.2.0 (SAS Institute) was used for statistical analysis. p < 0.05 was defined as significant. Results The median follow-up periods in the training set were 28.1 months for all patients and 54.7 months for survivors. In multivariate analysis in the training set, gender (male vs female, hazard ratio [HR]: 0.34, 95% confidence interval [CI]: 0.10 – 0.84, p = 0.017), first order feature variance (HR: 1.99, 95% CI: 1.02 – 4.17, p = 0.042) and gray level size zone matrix (GLSZM) large area low gray level emphasis (HR: 0.49, 95% CI: 0.23 – 0.96, p = 0.036) were significant predictors of PFS (Table 1). In the validation set, the log-rank test showed that the low variance group had significantly better PFS than did the high variance group (p = 0.048). There was no significant difference between the high and low GLSZM large area low gray level emphasis groups (Figure 1).