AI-driven radiopharmaceutical dosimetry optimization for theranostics: a systematic review and meta-analysis.
Authors
Affiliations (2)
Affiliations (2)
- Departamento de Investigación, Ministerio de Salud, Abha, Saudi Arabia. Electronic address: [email protected].
- Departamento de Investigación, Ministerio de Salud, Abha, Saudi Arabia.
Abstract
Personalised radioligand therapy (RLT) depends on accurate absorbed-dose estimation. Standard dosimetry requires multiple post-therapy single-photon emission computed tomography/computed tomography (SPECT/CT) acquisitions, imposing logistical burdens. Machine learning (ML) and deep learning (DL) have been proposed to streamline organ segmentation, generate synthetic imaging timepoints, and predict patient-specific doses. We evaluated the performance, methodological quality, and translational readiness of artificial intelligence (AI)-driven dosimetry in the clinical theranostics literature. In this systematic review and meta-analysis, we searched PubMed, Web of Science, Scopus, and regional databases (Jan 1, 2014, to March 1, 2025). Studies applying AI to dosimetry in human RLT recipients were eligible. Two reviewers screened and extracted data. Risk of bias was assessed with the Prediction model Risk Of Bias ASsessment Tool (PROBAST) and the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2). Meta-analyses of organ segmentation (Dice similarity coefficient [DSC]) and dose prediction (concordance correlation coefficient [CCC]) used restricted maximum-likelihood (REML) random-effects models restricted to studies with independent validation sets. CRD420261349805 (https://www.crd.york.ac.uk/prospero/display_record.php?RecordID=1349805). 58 studies met inclusion criteria covering 1847 patients (1523 [82%] male, 324 [18%] female). 23 studies with independent validation were eligible for meta-analysis. Pooled DSC was 0·87 (95% CI 0·83-0·90; 16 studies). Pooled voxel dosimetry relative absolute error (RAE) was 6·1% (4·6-8·1%; 12 studies). Pretherapy kidney dose prediction yielded pooled CCC 0·79 (0·72-0·85; 8 studies). PROBAST identified high risk of bias in 71% of prediction studies. Reducing timepoints from ≥3 to 1-2 scans increased RAE by 8·2 percentage points (5·1-11·4). AI reduced median processing time from 180 to 25 minutes. On independent validation data, AI-driven dosimetry shows promising accuracy for organ-level dose estimation and substantially reduces processing time. However, the evidence base remains immature: 71% of prediction studies carry high risk of bias, only four studies report external validation, and most pooled outcomes were graded as low or very low certainty. External multicentre validation and prospective randomised trials comparing AI-guided against fixed-activity RLT are needed before clinical adoption. Current evidence primarily demonstrates technical feasibility rather than proven clinical utility.