Deep learning-based PSMA PET segmentation repeatability: A post-hoc analysis of a single-center, prospective,
Jake Kendrick1,2,3, Roslyn J Francis4,5,6,7, Ghulam Mubashar Hassan8,4
1School of Physics, Mathematics and Computing, The University of Western Australia, 35 Stirling Highway, Mailbag M013, Crawley, Perth, WA, 6009, Australia. jake.kendrick@uwa.edu.au.
La Radiologia Medica
|October 30, 2025
Summary
Artificial intelligence (AI) models show reliable test-retest repeatability for PSMA-positive tumor volume in prostate cancer patients using the same PET tracer. Consistent results support using AI for response assessment, but different tracers yield wider variability.
Area of Science:
- Nuclear Medicine
- Oncology
- Artificial Intelligence in Medical Imaging
Background:
- Prostate-specific membrane antigen (PSMA) positron emission tomography (PET) is crucial for staging and monitoring prostate cancer.
- Artificial intelligence (AI) offers automated analysis of PSMA PET scans, but its test-retest repeatability needs validation.
- Assessing the performance of AI models across different PSMA radiotracers, such as [68Ga]Ga-PSMA-11 and [18F]F-PSMA-1007, is essential for clinical application.
Purpose of the Study:
- To quantify the test-retest repeatability of AI-derived PSMA PET imaging biomarkers.
- To evaluate the performance of an AI segmentation model trained on [68Ga]Ga-PSMA-11 PET scans when applied to [18F]F-PSMA-1007 PET scans.
- To compare the repeatability of AI biomarkers between intra-tracer and inter-tracer scenarios in metastatic prostate cancer patients.
Main Methods:
- A post-hoc analysis of a prospective, single-center, test-retest trial involving 17 patients with metastatic prostate cancer.
- Patients underwent two PSMA PET scans, either with the same tracer (intra-tracer group, n=9) or different tracers (inter-tracer group, n=8).
- AI-driven and semi-automated delineation were used to quantify subject-level repeatability of four imaging biomarkers, including PSMA-positive tumor volume.
Main Results:
- Poorer repeatability was observed for all biomarkers in the inter-tracer group compared to the intra-tracer group.
- In the intra-tracer group, AI-derived PSMA-positive tumor volume showed a repeatability coefficient of 13.8% for patients with higher disease volume.
- The AI model demonstrated comparable lesion-level positive predictive value between [68Ga]Ga-PSMA-11 and [18F]F-PSMA-1007 PET scans.
Conclusions:
- AI-based PSMA-positive tumor volume calculations exhibit repeatability limits suitable for response assessment (RECIP 1.0 criteria) in higher volume disease when using the same PET tracer.
- Significant variability in repeatability between different tracers underscores the importance of using a consistent radiotracer for response assessment in metastatic prostate cancer.
- These findings support the use of AI for quantitative PSMA PET analysis, provided consistent imaging protocols are maintained.
