AI in personalized internal dosimetry for radiopharmaceutical therapy

Source article: Artificial intelligence in personalized internal dosimetry for radiopharmaceutical therapy: a narrative mini-review

Personalized internal radiation dosimetry supports individualized radiopharmaceutical therapy but remains difficult to implement because workflows require repeated quantitative imaging, segmentation, registration, kinetic modeling, and absorbed-dose calculation. This mini-review summarizes current evidence on artificial intelligence across the personalized dosimetry workflow and considers clinical readiness. A narrative PubMed search was completed in early 2026 and covered peer-reviewed studies published from 20…

By the Truvace deskOctober 20, 20262 min read1 sourcePeer-reviewedTRV-2026-2002· cite & verify

Trace impact readingContested

P 71

Both sides are scored from claims and sources, not community votes.

G 69

In brief

A narrative mini-review surveyed peer-reviewed studies from 2015 to early 2026 on machine learning and deep learning applied to personalized internal dosimetry. It covered image reconstruction, segmentation, registration, time-activity curve estimation, absorbed-dose calculation and decision support. Reported results included large drops in reconstruction error and faster registration.

Dosimetry-guided radiopharmaceutical therapy depends on a multi-step workflow that is hard to run in routine practice, so tools that shorten or simplify steps could widen access to individualized dosing. The review finds readiness uneven: reconstruction and segmentation look closer to clinical use, while decision-support tools lack prospective, multicenter evidence. Because this is a narrative review rather than a systematic one, the evidence base may be incompletely captured.

Main points

Deep learning lowered Y-90 SPECT reconstruction error from 149.5% ± 31.2% to 9.2% ± 1.7% in the cited work.
Lesion dose-rate error was 8.6% ± 5.7% in virtual-patient testing.
Learning-based registration cut computation time from about 50 s to under 6 s.
Automated segmentation shows improving reproducibility, and machine learning has been explored for predicting organ effective half-life.
Reconstruction and segmentation appear nearer clinical readiness than decision-support applications.

The gain

Deep learning cut Y-90 SPECT reconstruction error and sped up image registration, easing workflow burdens that make individualized radiopharmaceutical dosimetry hard to implement.

The problem

AI-based clinical decision support for dosimetry rests mainly on retrospective, single-center evidence and still needs prospective, multicenter validation before routine clinical use.

The rundown

The review identifies the barriers to personalized dosimetry as repeated quantitative imaging, segmentation, registration, kinetic modeling and absorbed-dose calculation. It examines AI at each of these steps, prioritizing studies that reported quantitative performance or validation results.

The headline accuracy figures come from specific tasks. The Y-90 SPECT reconstruction improvement is a reported reduction in error, and the 8.6% lesion dose-rate error was measured in virtual-patient testing rather than in a clinical cohort.

Single-time-point approaches may simplify imaging requirements, and recent machine learning work has explored predicting organ effective half-life. Automated segmentation reproducibility is described as improving.

On clinical readiness, the authors conclude that reconstruction and segmentation applications are closer to use, whereas decision-support tools remain supported mostly by retrospective, single-center studies.

What this doesn’t fix

The evidence for AI-based clinical decision support is mainly retrospective and single-center, and decision-support tools still need prospective, multicenter validation.