Automatic Extraction of PET RANO Criteria with an Externally Validated Deep Learning Model: Application to [18F]FDOPA PET Imaging

BACKGROUND: Automatic segmentation of gliomas on amino acid PET is essential for quantitative tumor assessment, a pillar in monitoring gliomas under treatment. This study aimed to develop a deep learning model for the automated extraction of PET RANO criteria from [18F]FDOPA PET, with external validation. METHODS: A total of 635 static [18F]FDOPA PET scans from three European centers were retrospectively included for glioma diagnosis, recurrence assessment, or treatment monitoring. The training cohort comprised 530 scans from Nancy Hospital, with external validation and test sets from Pitié-Salpêtrière Hospital (n = 66) and Turin Hospital (n = 39). Ground truth segmentations followed international guidelines. A 3D U-Net was trained to segment tumor and healthy brain volumes. Performance was evaluated using the Dice coefficient using the whole tumor volume. Quantitative agreement for PET RANO criteria 1.0 parameters, tumor-to-background ratios (TBRmean, TBRmax) and metabolic tumor volume (MTV), was assessed at the lesion-level. RESULTS: Tumor segmentation achieved Dice of 0.925 [0.841; 0.970] in training, 0.885 [0.829; 0.925] in validation, and 0.851 [0.733; 0.911] in the test set. At lesion level, agreement with expert quantification was high, with low bias and strong reliability for MTV (2.293 [-4.734; 9.321] mL), TBRmax (0.056 [-0.189; 0.301]), and TBRmean (-0.139 [-0.424; 0.146]) and intra-class correlation coefficients superior to 0.93. Measurable lesions were correctly identified in more than 97% of cases. CONCLUSION: Our [18F]FDOPA PET deep learning model (available at https://github.com/IADI-Nancy/FDOPA-PET-GliomaSeg) demonstrates robust multicenter performance and enables fully automated, reproducible quantification, supporting broader clinical adoption of amino-acid PET in neuro-oncology.