Answers are grounded in this paper and cite their sources.
Background: The routine clinical implementation of patient-specific dosimetry in 177 Lu-PSMA radiopharmaceutical therapy can be challenging due to the need for multiple-time-point posttherapy SPECT/CT imaging to estimate time-integrated activity (TIA). Single-time-point (STP) approaches offer a more practical alternative, although their accuracy has yet to be fully established. This study aimed to evaluate existing analytical STP methods and to assess the feasibility and performance of a machine learning (ML) framework. Patients and Methods: Thirty-four patients with metastatic castration-resistant prostate cancer undergoing 177 Lu-PSMA-617 therapy were retrospectively analyzed. Reference TIAs were derived from multiple-time-point 177 Lu SPECT/CT imaging at 3 or 4 time points (TPs) postinjection (TP1: 0.7-2.1 h, TP2: day 1–2, TP3: day 3–4, TP4: day 5–7) in cycle 1 for 147 tumors, 64 kidneys, and 65 parotid glands using monoexponential or biexponential fitting. Analytical STP methods, proposed by Hänscheid and Madsen, were evaluated across different TPs. For each organ, a ML model was developed using a residual learning approach, in which the log-transformed ratio between reference TIAs and analytical STP TIA estimates was modeled. Input features included combinations of STP SPECT-derived parameters, baseline 68 Ga-PSMA-11 PET-derived biomarkers (for individual structures and whole-body tumor volume), and clinical variables. Model performance was evaluated by using multiple quantitative metrics within a 10-times repeated 5-fold nested cross-validation framework and compared with purely analytic STP approaches. Model interpretability was assessed using SHapley Additive exPlanations. Results: Analytical methods showed strong performance at optimal imaging TPs (TP2 for kidney/parotid, and TP3 for tumor), with the largest errors observed at early imaging (TP1). Overall, ML models provided modest improvement over analytical methods across structures at most TPs. The best performance was achieved using a ML model combining PET/SPECT imaging features with analytical TIA estimates. The mean percentage errors of the models, across TPs, ranged from −1.7 ± 17.9% (TP2) to 7.4 ± 65.3% (TP1) for kidneys, −0.4 ± 12.1% (TP2) to 4.5 ± 26.7% (TP4) for parotid glands, and 1.4 ± 14.8% (TP3) to −4.1 ± 100.5% (TP1) for tumors. Within the present cohort and modeling framework, the inclusion of clinical variables did not significantly improve model performance. SHapley Additive exPlanations analysis showed that the ML models applied a positive correction to the analytical TIA estimates at early imaging TPs and a negative correction at late TPs. Conclusions: Analytical STP methods yield accurate TIA estimates when applied at appropriate imaging TPs. ML models that integrate analytical estimates with imaging features provide modest improvements in accuracy with less sensitivity to TP selection. However, optimal timing remained structure-dependent, requiring a compromise when a single acquisition is used. These findings support the potential of ML-driven STP dosimetry in 177 Lu-PSMA-617 radiopharmaceutical therapy to enable more flexible and clinically feasible workflows, pending further validation with larger and more diverse data sets.
No takes yet. Share an insight, caveat, or question.
Mansouri et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: