Fully Automated Segmentation of [<sup>18</sup>F]FDG and [<sup>68</sup>Ga]/[<sup>18</sup>F]PSMA PET/CT Images via Data-Centric Deep Learning.
Authors
Affiliations (24)
Affiliations (24)
- QIMP Team, Medical University of Vienna, Vienna, Austria; [email protected].
- QIMP Team, Medical University of Vienna, Vienna, Austria.
- Santa Croce e Carle Hospital, Medical Physics, Cuneo, Italy.
- Department of Biomedical Imaging and Image-Guided Therapy, Division of Nuclear Medicine, Medical University of Vienna, Vienna, Austria.
- Division of Nuclear Medicine, Azienda Ospedaliero Universitaria Careggi, Florence, Italy.
- Medical University of Vienna, Vienna, Austria.
- Department of Nuclear Medicine, West German Cancer Center, University Hospital Essen, University of Duisburg-Essen, Essen, Germany.
- German Cancer Consortium, Partner Site, University Hospital Essen, Essen, Germany.
- Institute for AI in Medicine, University Hospital Essen, Essen, Germany.
- Department of Radiology, LMU University Hospital, LMU Munich, Munich, Germany.
- Munich Center for Machine Learning, Munich, Germany.
- relAI-Konrad Zuse School of Excellence in Reliable AI, Munich, Germany.
- Melbourne Theranostic Innovation Centre, North Melbourne, Victoria, Australia.
- St Vincent's Hospital, Department of Medicine, The University of Melbourne, Fitzroy, Victoria, Australia.
- Department of Nuclear Medicine and Clinical Molecular Imaging, University Hospital Tuebingen, Tuebingen, Germany.
- Werner Siemens Imaging Center, Department of Preclinical Imaging and Radiopharmacy, Eberhard Karls University Tuebingen, Tuebingen, Germany.
- IRIS, Institut Curie, Inserm, CNRS, Université Versailles Saint-Quentin, Universite PSL, Orsay, France.
- Department of Nuclear Medicine, LMU University Hospital, LMU Munich, Munich, Germany.
- Department of Nuclear Medicine, Medical University of Innsbruck, Innsbruck, Austria.
- Department of Radiology, University of California, Davis, Davis, California.
- Nuclear Medicine Unit, South Egypt Cancer Institute, Assiut University, Assiut, Egypt.
- Department of Nuclear Medicine, Veterans Health Service Medical Center, Seoul, Korea.
- Turku PET Centre, University of Turku and Turku University Hospital, Turku, Finland; and.
- Department of Clinical Physiology and Nuclear Medicine, Turku University Hospital, Turku, Finland.
Abstract
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 [<sup>18</sup>F]FDG and prostate-specific membrane antigen (PSMA)-targeted PET/CT, and to investigate how training data characteristics influence segmentation performance. <b>Methods:</b> In this retrospective multicenter study, 5209 [<sup>18</sup>F]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 [<sup>18</sup>F]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 [<sup>18</sup>F]FDG scans across 4 diseases; 443 PSMA-targeted prostate cancer scans) and compared with 3 open-source tools. <b>Results:</b> Organ context improved median Dice from 0.62 to 0.71 for [<sup>18</sup>F]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, <i>P</i> = 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 [<sup>18</sup>F]FDG and 0.85 for PSMA and outperformed other open-source approaches on the model-comparison test set, which excluded the AutoPET test cases. <b>Conclusion:</b> LION enables PET-only automated segmentation for [<sup>18</sup>F]FDG and PSMA-targeted PET. Training data composition, particularly spatial diversity quantified by TOMs, was strongly associated with segmentation performance.