Moderate evidence
More
A 72-patient study found AI outlines overlapped experts' by 89.2%; it marked 711 lesions vs 848
Study title: Evaluation of the detection accuracy of a fully automated AI-based lesion segmentation tool in whole-body FDG PET/CT for lymphoma
- Evidence type
- Observational study
- Sample
- n = 72
- Sources
- 1 sourcePeer-reviewed
- Published
- 1 min read
- Truvace record
- TRV-2026-1222cite or check it
Study title: Evaluation of the detection accuracy of a fully automated AI-based lesion segmentation tool in whole-body FDG PET/CT for lymphoma
By September 2026, researchers retrospectively evaluated a fully automated AI-based lesion segmentation tool using a hybrid 2D-3D nnUNet with ResNet18 backbone on whole-body FDG PET/CT scans from 72 lymphoma patients. Compared with manual delineations by experienced physicians, the model segmented 711 lesions versus 848 manual lesions and achieved DSC 89.2%, sensitivity 82.9, PPV 96.5, with strong correlations for PET metrics.
Read the study abstract
Background This study evaluated an AI-based lesion segmentation tool using a hybrid 2D-3D deep learning model based on the nnUNet architecture with a ResNet18 backbone, incorporating post-processing to reduce false positives and segmentation-related quantitative errors. Methods Whole-body FDG PET/CT scans from 72 lymphoma patients (mean age 39.2 ± 18.6 years; 33 females, 39 males; 43.1% Hodgkin lymphoma, 56.9% non-Hodgkin lymphoma) acquired on a uMI550 digital PET/CT system were retrospectively analysed. Automat…

56Gain side
How the two sides scoreClose call
53Problem side
The gain
In 72 lymphoma patients' scans, the AI tool's lesion outlines overlapped expert manual outlines by 89.2% (DSC) with strong agreement on PET measurements (r > 0.95); workload savings are only suggested.
The problem
The tool segmented 711 lesions against 848 outlined by experts, with 82.9% sensitivity, so it missed lesions; it was tested on a single scanner and its clinical impact is unknown.
A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.
What the scores mean
G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.
P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.
In brief
High concordance suggests potential to reduce manual segmentation burden and standardize quantitative PET metrics in lymphoma care, but the lower lesion count and 82.9% sensitivity indicate missed lesions. As a single-scanner retrospective study, generalizability, region-specific performance, and impact on clinical decision-making remain uncertain and require further refinement.
Main points
- Retrospective analysis of 72 lymphoma patients (mean age 39.2 ± 18.6 years; 33 females, 39 males; 43.1% Hodgkin lymphoma, 56.9% non-Hodgkin lymphoma) scanned on uMI550 digital PET/CT.
- AI used hybrid 2D-3D deep learning model based on nnUNet architecture with ResNet18 backbone and post-processing to reduce false positives.
- 711 AI-segmented lesions versus 848 manually segmented lesions compared using F1-score, sensitivity, PPV, Pearson correlation, concordance correlation, ICC, and Bland-Altman analysis.
- DSC showed moderate negative correlation with liver SNR (r = -0.40)
The full story
The study analyzed whole-body FDG PET/CT scans from 72 lymphoma patients acquired on a uMI550 digital PET/CT system, with 43.1% Hodgkin and 56.9% non-Hodgkin lymphoma. Manual delineations by experienced physicians were the reference, covering 848 lesions versus 711 AI lesions.
Performance metrics included DSC 89.2%, sensitivity 82.9 and PPV 96.5, with strong quantitative agreement (r > 0.95, ICC > 0.94, 95% CI: 0.93-0.99) and minimal bias on Bland-Altman. Authors noted moderate negative correlation between DSC and liver SNR and called for region-specific training.
Sources
1 on file- Zeitschrift für Medizinische Physikdoi:10.1016/j.zemedi.2026.09.002
Talk it over
The AI tool's outlines overlapped experts' by 89.2%, yet it marked 711 lesions against 848 and had 82.9% sensitivity. How do you weigh the agreement against the missed lesions?
A 72-patient study found AI outlines overlapped experts' by 89.2%; it marked 711 lesions vs 848https://truvace.com/item/evaluation-of-the-detection-accuracy-of-a-fully-automated-ai-based-lesion-segmen
The AI tool's outlines overlapped experts' by 89.2%, yet it marked 711 lesions against 848 and had 82.9% sensitivity. How do you weigh the agreement against the missed lesions?
Answer in the threadIf you were a physician reading a scan from the 72 lymphoma patients, which would you want first: the AI outlines as a draft you correct, or your own outlines checked by the AI? Why?
Answer in the threadThe study used one scanner, the uMI550 digital PET/CT, and looked back at past scans. What would a study need to show to tell you whether the tool changes treatment decisions for lymphoma patients?
Answer in the thread
If you have more time
DSC showed a moderate negative correlation with liver SNR (r = -0.40). Which kinds of scans or patients would you want tested next to see where the tool struggles?
Answer in the threadThe researchers called for region-specific training. How would a hospital check which lesions the tool missed among the 848 before relying on it, and who should do that check?
Answer in the thread
ace
What readers say