JS01.3.A FULLY AUTOMATED EVALUATION OF FET PET IN BRAIN TUMOR PATIENTS - AN INTERNATIONAL, MULTICENTER STUDY

Abstract BACKGROUND Amino acid PET using the tracer O-(2-[18F]fluoroethyl)-L-tyrosine (FET) has become a valuable modality for brain tumor diagnostics, complementing MRI. Despite existing guidelines, variations persist in the assessment of quantitative metrics such as the maximum and mean tumor-to-brain ratios (TBRmax; TBRmean) and the metabolic tumor volume (MTV). To address this, we developed a fully automated workflow for the evaluation of FET PET and compared the performance to human experts in a multicenter setting. MATERIAL AND METHODS A total of 817 FET PET scans from 672 patients were evaluated. These comprised 740 scans from two institutions in Germany (3 scanners; 65% glioma; 54% male; median injected activity 222 MBq), and 77 scans from one institution in the USA (1 scanner, 74% glioblastoma; 66% male; median injected activity 703 MBq). Our approach builds on the previously developed Juelich Segmentation Tool for Brain Tumor PET (JuST_BrainPET, PMID: 37562802) to derive the initial tumor mask, followed by post-processing of the brain-extracted scan to obtain a reference region. With these, SUV-based thresholding is applied, and diagnostic metrics are computed on the summed images from 20 to 40 minutes post injection. The mean background SUV (BG-SUVmean), TBRmean, TBRmax, and MTV were calculated and compared with the available expert-rated parameters, using Pearson correlation coefficients and Bland-Altman plots for assessing bias. Detection accuracy was measured by taking the physician findings as reference standard. RESULTS Automated assessment on the German dataset obtained a sensitivity of 92% and an F1-score of 93%. In this process, 106 lesions with an MTV below 0.5 mL were classified as non-measurable disease according to PET RANO 1.0 and were therefore not included in further evaluation. The correctly identified lesions showed a high correlation between automated and expert assessment for BG-SUVmean, TBRmean and TBRmax (Pearson correlation, 0.98; 0.86; 0.96, respectively), with a slight bias towards overestimation (mean bias, 0.01; 0.10; 0.17, respectively). For the USA dataset, the workflow obtained a sensitivity of 92% and an F1-score of 95%, with a high correlation between the automated and manual assessment of the MTV as well as TBRmax (Pearson correlation, 0.95; 0.83, respectively). A slight bias towards underestimation was noted for both MTV and TBRmax (mean bias, 4.30 mL; 0.38, respectively). CONCLUSION The presented fully automated workflow demonstrated its value for robust and standardized evaluation of FET PET scans in brain tumor patients on different scanners from multiple institutions. FUNDING This work was supported by project 428090865 (SPP2177) of the German Research Foundation (DFG).