70/100

Moderate evidence

Cite
More
EmailLinkedInX

Review of 58 studies shows AI cut radioligand therapy dose processing from 180 to 25 minutes

Study title: AI-driven radiopharmaceutical dosimetry optimization for theranostics: a systematic review and meta-analysis

Evidence type
Systematic review or meta-analysis
Sample
n = 1,847
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: AI-driven radiopharmaceutical dosimetry optimization for theranostics: a systematic review and meta-analysis

A systematic review and meta-analysis of 58 studies covering 1,847 patients examined AI used to estimate absorbed radiation dose in radioligand therapy. Across the studies, AI reduced median processing time from 180 to 25 minutes.

Read the study 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 artific…

AI-driven radiopharmaceutical dosimetry optimization for theranostics: a systematic review and meta-analysis
"CA Dean Hospital CT Scanner, X-Ray, and Snowplow with USDA Help" by USDAgov, public domain.

The gain

A systematic review of 58 studies covering 1,847 patients found AI cut median dosimetry processing time for radioligand therapy from 180 to 25 minutes, with pooled organ segmentation accuracy (Dice) of 0.87 on independent validation data.

In brief

Standard dosimetry needs several post-therapy SPECT/CT scans, which burdens patients and hospitals. AI methods aim to segment organs, generate synthetic imaging timepoints and predict patient-specific doses, which could ease that workload.

The evidence is still early. Most prediction studies carry a high risk of bias and few have external validation, so the findings show technical feasibility rather than proven benefit for patients.

Main points

  1. The review searched PubMed, Web of Science, Scopus and regional databases from January 2014 to March 2025 and included 58 studies with 1,847 patients (82% male).
  2. Of those, 23 studies with independent validation sets entered the meta-analysis.
  3. Pooled organ segmentation accuracy was a Dice coefficient of 0.87 (95% CI 0.83-0.90) across 16 studies.
  4. Pooled voxel dosimetry relative absolute error was 6.1% (4.6-8.1%) across 12 studies.
  5. Pretherapy kidney dose prediction reached a pooled concordance correlation coefficient of 0.79 (0.72-0.85) across 8 studies.
  6. Cutting imaging from at least 3 scans to 1-2 scans raised relative absolute error by 8.2 percentage points (5.1-11.4).

What the review found when it combined the studies

These are the review’s own pooled results, copied from its abstract. Truvace did not recalculate them.

  • Organ segmentation overlap (Dice similarity coefficient) of AI on independent validation sets in radioligand therapy

    Dice similarity coefficient 0·87 (95% CI 0·83 to 0·90)

    16 studies · search to March 1, 2025

    The review’s own words
    Pooled DSC was 0·87 (95% CI 0·83-0·90; 16 studies).

Code checks every number above against the review’s text before it is shown. The short labels are an AI model’s reading of that text, so the review’s own sentence sits under each one. A pooled figure is only as strong as the studies behind it. How Truvace uses pooled results

The full story

Two reviewers screened studies and extracted data. Bias risk was assessed with PROBAST and QUADAS-2, and random-effects models limited to studies with independent validation pooled segmentation and dose prediction results.

On independent validation data, AI dosimetry showed promising accuracy for organ-level dose estimates and much shorter processing. Reducing the number of scans to one or two came at a cost of higher error, an increase of 8.2 percentage points.

The authors say PROBAST found high risk of bias in 71% of prediction studies, only four studies reported external validation, and most pooled outcomes were graded low or very low certainty. They call for multicentre external validation and prospective randomised trials comparing AI-guided with fixed-activity therapy before clinical adoption.

Sources

1 on file
  1. Peer-reviewed2026-11-13

Talk it over

Three questions to start on, two more if you have time
AI cut median dosimetry processing time from 180 to 25 minutes, yet 71% of prediction studies had high risk of bias. How do you weigh the speed against that weakness in the evidence?

Start with a question

Review of 58 studies shows AI cut radioligand therapy dose processing from 180 to 25 minuteshttps://truvace.com/item/ai-driven-radiopharmaceutical-dosimetry-optimization-for-theranostics-a-systemat

  1. AI cut median dosimetry processing time from 180 to 25 minutes, yet 71% of prediction studies had high risk of bias. How do you weigh the speed against that weakness in the evidence?

    Answer in the thread
  2. Only four of the 58 studies reported external validation, and most pooled outcomes were rated low or very low certainty. What does that tell us about how far the 0.87 Dice figure can be trusted in other hospitals?

    Answer in the thread
  3. Cutting imaging from at least 3 scans to 1-2 scans raised error by 8.2 percentage points. If you were a patient, how much extra error would you accept to have fewer scans?

    Answer in the thread
If you have more time
  • The authors call for prospective randomised trials comparing AI-guided with fixed-activity therapy. What outcomes would you want such a trial to measure?

    Answer in the thread
  • The 1,847 patients were 82% male. Who might be missing, and how could that affect dose predictions?

    Answer in the thread

What readers say