Fully Automated Segmentation of [ 18 F]FDG and [ 68 Ga]/[ 18 F]PSMA PET/CT Images via Data-Centric Deep Learning

The purpose of this study was to develop and validate lesion identification in oncologic nuclear imaging (LION), an open-source PET-only tumor segmentation pipeline for [18F]FDG and prostate-specific membrane antigen (PSMA)–targeted PET/CT, and to investigate how training data characteristics influence segmentation performance. Methods: In this retrospective multicenter study, 5209 [18F]FDG PET/CT scans spanning 19 disease types and 2046 PSMA-targeted PET/CT scans were used to train PET-only segmentation models. Tumor segmentation incorporated organs with physiologic uptake as auxiliary classes to enable PET-only inference. Tumor occurrence maps (TOMs) quantified tumor spatial diversity across the training data. For [18F]FDG, disease-specific and mixed-disease models trained on progressively larger subsets were compared to test whether increasing spatial diversity improves generalization. Scanner-related domain shift was analyzed using DINOv2 embeddings. Models were evaluated on multicenter holdout cohorts (616 [18F]FDG scans across 4 diseases; 443 PSMA-targeted prostate cancer scans) and compared with 3 open-source tools. Results: Organ context improved median Dice from 0.62 to 0.71 for [18F]FDG and from 0.75 to 0.83 for PSMA, on the complete holdout cohorts. Spatial diversity measured by TOMs was strongly associated with Dice (Spearman ρ = 0.80, P = 0.003). A mixed-disease model trained on 500 patients matched the performance of a lymphoma specialist model trained on 3031 cases. DINOv2 embeddings revealed scanner-induced domain shift between same-disease cohorts. LION achieved median Dice scores of 0.71 for [18F]FDG and 0.85 for PSMA and outperformed other open-source approaches on the model-comparison test set, which excluded the AutoPET test cases. Conclusion: LION enables PET-only automated segmentation for [18F]FDG and PSMA-targeted PET. Training data composition, particularly spatial diversity quantified by TOMs, was strongly associated with segmentation performance.