ESTRO 2020 Abstract book

S901 ESTRO 2020

PO-1571 Radiomics for prediction of metastatic melanoma patient survival after immunotherapy H. Gabrys 1 , L. Basler 1 , S. Hogan 2 , M. Pavic 1 , M. Bogowicz 1 , D. Vuong 1 , S. Tanadini-Lang 1 , R. Förster 1 , M. Huellner 3 , R. Dummer 2 , M. Guckenberger 1 , M.P. Levesque 2 1 University Hospital Zurich, Department of Radiation Oncology, Zurich, Switzerland ; 2 University Hospital Zurich, Department of Dermatology, Zurich, Switzerland ; 3 University Hospital Zurich, Department of Nuclear Medicine, Zurich, Switzerland Purpose or Objective Immune checkpoint inhibition has achieved substantial improvements in survival of metastatic melanoma patients; however, reliable biomarkers for patient selection are lacking. This study aims to assess the predictive potential of PET/CT-based radiomics for modeling patient overall survival (OS) in metastatic melanoma patients receiving checkpoint inhibitors. Material and Methods A retrospective single-institution cohort of 112 patients with metastatic melanoma was the basis of this study. Patients were treated with either single-checkpoint- inhibition using anti-PD-1-antibodies or dual-checkpoint- inhibition in combination with anti-CTLA-4-antibodies. PET/CT imaging was done before the treatment (TP0) and at 3 months after the treatment initiation (TP1). All individual metastases (n=716 over all patients) were segmented on all images for calculation of radiomic features (n=172) describing shape, intensity, and texture of the metastases. To perform the analysis on a patient level, radiomic features extracted from multiple lesions were aggregated via weighted arithmetic mean, where the weight corresponded to a lesion volume. Logistic regression models were built using the data available at TP0 and delta-radiomic features calculated between TP0 and TP1. The performance metric was the area under the receiver operating characteristic curve (AUC). The generalization performance of the models was estimated with a nested cross-validation (inner loop: 10 times repeated 10-fold cross-validation, outer loop: 10 times repeated 5-fold cross-validation). Model calibration has been evaluated with calibration curves and the Brier score. The radiomic models were benchmarked against models based only on metastases volume. Results The OS at 12 and 36 months was 84% and 55%, respectively (Fig. 1). Prognostic models based on radiomic features achieved AUCs of 0.87 +/- 0.08 and 0.74 +/- 0.13 at 12 and 36 months, respectively. Volume-based models performed slightly worse with AUCs of 0.83 +/- 0.15 and 0.71 +/- 0.13, at 12 and 36 months, respectively. Nevertheless, the radiomics models were more stable and better calibrated than the volume-based models (Fig. 2). Inclusion of PET- based radiomic features did not improve performance of radiomic models in either of the investigated scenarios. Considering only the largest lesion instead of the weighted mean of all lesions resulted in AUC scores smaller by 0.16 on average.

Results Analysis process leadings to results, is depicted in Fig.2. Robustness threshold of the 1972 analyzed dosomics features has been set to 0.5, leading to the following results. Regarding the statistical features, a R fe > than 0.5 was detected for: kurtosis in 10 patients (5 PTVs and 7 OARs – 2 bowel, 2 bladder, 2 optic chiasm, 1 rectum ); skewness in 11 patients (7 PTVs, 2 rings and 1 OAR – optic chiasm); intensity uniformity in 3 patients (4 OARs – 3 lungs and esophagus); minimum intensity in 2 patients (OARs – 2 lungs);10th intensity percentile in 1 patient (1 ring); intensity-based energy in 1 patient (2 OARs – Bladder Rectum). For morphologic features, a R fe > than 0.5 was detected for: compactness 1 in 4 patients (4 PTVs). In relation to textural features, a R fe > than 0.5 was detected for: cluster shade in 6 patients (5 PTVs and 1 OARs – bowel); cluster prominence in 1 patient (1 PTV); cluster tendency in 1 patient (1 PTV).

Conclusion These preliminary results lead to define a group of robust dosomic features potentially useful for improve model prediction. Robustness evaluation of the second level features will be performed, as well as the dependency on electron density map slice thickness and other clinical/planning parameters.