68/100

Moderate evidence

Cite
More
EmailLinkedInX

Adding PET/CT reports to AI cut false-positive prostate cancer lesions by about 4 per patient

Study title: Prostate-Specific Membrane Antigen PET/CT Segmentation Guided by Vision Language Models Embedding Clinical Report Semantics

Evidence type
Controlled or validated comparison
Sample
n = 284
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: Prostate-Specific Membrane Antigen PET/CT Segmentation Guided by Vision Language Models Embedding Clinical Report Semantics

Researchers built a 3-dimensional segmentation model that combines the text of nuclear medicine reports with 68Ga-PSMA-11 PET/CT images, so that it can better tell cancer from normal physiologic uptake. They tested four ways of encoding report text against an image-only baseline.

Read the study abstract

Whole-body prostate-specific membrane antigen (PSMA) PET/CT segmentation remains challenging: manual contouring is unsustainable, and image-only artificial intelligence models often misclassify physiologic uptake as malignant, causing high false-positive rates. These limitations underscore the need for segmentation strategies that incorporate the nuclear medicine expertise encoded in PET/CT reports but not used by current models. We investigated whether routine PET/CT reports improve automated lesion segmentatio…

Prostate-Specific Membrane Antigen PET/CT Segmentation Guided by Vision Language Models Embedding Clinical Report Semantics
"Interventional radiology A" by Navy Medicine, public domain.

The gain

In a study of 284 prostate cancer patients, an AI model that read PET/CT report text alongside the scans cut false-positive lesions by nearly 4 per patient with BiomedVLP versus an image-only model.

In brief

With BiomedVLP, false-positive lesions fell by nearly 4 per patient, and by about 3 per patient with RadGraph and GPT-raw. Average symmetric surface distance dropped from 105.4 mm for the image-only model to 23.4 mm with BiomedVLP.

When the authors checked only correctly detected lesions, no language-guided model beat the image-only model on any surface metric. They conclude that the boundary gains came from removing false positives, not from tracing tumours more accurately.

Main points

  1. Of 320 patients scanned with 68Ga-PSMA-11 PET/CT, 284 were analysed: 243 for development and a fixed held-out test set of 41.
  2. Four report-encoding strategies were compared with an image-only baseline: BiomedVLP, RadGraph, GPT-raw and GPT-engineered.
  3. The 95th percentile Hausdorff distance improved from 455.0 mm to 123.8 mm with BiomedVLP.
  4. Precision rose by 0.099 with BiomedVLP and 0.060 with RadGraph, while voxel-level recall dipped by 0.051 with BiomedVLP.
  5. Dice improved by 0.032 but was only borderline significant; lesion-level sensitivity changed by no more than 0.04.
  6. Total lesion PSMA stayed highly correlated with manual measurements (r of 0.85 or higher), but tumour volume correlations varied (r = 0.37-0.70).

The full story

The model is a language-guided SegResNet that fuses report semantics with image features. The text was encoded four ways: BiomedVLP, RadGraph, GPT-raw and GPT-engineered. Of 320 patients, 13 studies were excluded for missing reports or incomplete image-mask pairs, leaving 284 split at the patient level.

Performance was judged on Dice, intersection over union, precision, recall, surface distances and lesion-level false positives under three matching criteria. Significance used paired one-sided Wilcoxon tests with Holm adjustment.

The gains in whole-mask surface metrics vanished once only matched true-positive lesions were analysed (all Holm-adjusted P = 1.0). The authors say this shows false-positive suppression, not better contouring. The trade-off was a modest drop in recall.

The authors say the approach may reduce manual cleanup and support more reliable quantitative workflows, and call for external validation and PSMA-specific language pretraining.

Sources

1 on file
  1. Journal of Nuclear Medicinedoi:10.2967/jnumed.126.272702
    Peer-reviewed2026-11-12

Talk it over

Three questions to start on, two more if you have time
The model cut false-positive lesions by nearly 4 per patient with BiomedVLP, while voxel-level recall dipped by 0.051. How would you weigh fewer false alarms against a small loss in recall?

Start with a question

Adding PET/CT reports to AI cut false-positive prostate cancer lesions by about 4 per patienthttps://truvace.com/item/prostate-specific-membrane-antigen-pet-ct-segmentation-guided-by-vision-language

  1. The model cut false-positive lesions by nearly 4 per patient with BiomedVLP, while voxel-level recall dipped by 0.051. How would you weigh fewer false alarms against a small loss in recall?

    Answer in the thread
  2. The test set held only 41 cases from one dataset. What could results from that set tell us about how the model would do on scans from other hospitals?

    Answer in the thread
  3. When only correctly detected lesions were checked, no language-guided model beat the image-only model on any surface metric. What finding would convince you the model traces tumours better, not just removes false positives?

    Answer in the thread
If you have more time
  • If you were a clinician reading these scans, what would you want to check before trusting a model that reads the report text alongside the images?

    Answer in the thread
  • Look up what external validation would involve for a model like this. Which kinds of patients and scanners would you want in it?

    Answer in the thread

What readers say