Automated lesion tracking and treatment response classification on [⁶⁸Ga]Ga-PSMA-11 PET/CT across PPP, RECIP 1.0 and TLW frameworks.

Abstract

This study aimed to validate an automated lesion-tracking algorithm for [68Ga]Ga-PSMA-11 PET/CT against expert reference matches and to evaluate its integration into three PSMA-specific treatment response frameworks-the Response Evaluation Criteria in PSMA PET/CT (RECIP 1.0), PSMA PET Progression (PPP) criteria, and the Traffic Light Workflow (TLW)-in men with biochemically recurrent (BCR) prostate cancer (PCa). Paired baseline and follow-up [68 Ga]Ga-PSMA-11 PET/CT scans from 132 BCR PCa patients were analysed using a registration-based pipeline that automatically assigned one-to-one lesion correspondences. Lesions were labelled as corresponding (CL), disappearing (DL) or new (NL). Algorithm-derived RECIP 1.0, PPP, and TLW categories were compared with expert reference labels and tested for overall survival (OS) association using univariable Cox regression. A total of 1,066 lesions were identified (559 baseline, 507 follow-up), including 371 CLs, 188 DLs, and 136 NLs, yielding 695 matching decisions. The lesion matching algorithm achieved 95.1% matching accuracy. Lesion mismatches were associated with poorer registration quality. Algorithm-derived patient-level response classifications showed high concordance with ground-truth frameworks, with classification accuracy of 95.5% for RECIP 1.0, 96.2% for PPP, and 96.2% for TLW. All frameworks retained prognostic significance for OS, with progressive disease conferring hazard ratios (HR) of-RECIP 1.0 HR = 2.57 (95% CI 1.33-4.98), PPP HR = 2.02 (95% CI 1.04-3.90), and TLW HR = 2.29 (95% CI 1.22-4.28). The automated lesion-tracking algorithm reproduced expert matches with high accuracy and preserved the prognostic utility of RECIP 1.0, PPP, and TLW in a BCR cohort, supporting its use as a reproducible lesion-matching method for quantitative response assessment on PSMA PET/CT. ACTRN ACTRN12615000608561. Registered 11 June 2015. Retrospectively registered.