LiON Liver CT Study Tests AI as an Additional Reader
Research brief · Peer-reviewed paper published 19 August 2026
LiON, the Liver DiagnOsis Network, was evaluated as an additional AI reader for contrast-enhanced CT. The Nature Medicine study combines retrospective validation with a prospective single-arm deployment, providing evidence about performance inside an imaging workflow.
The evidence comes in separate stages
LiON study stages and their sizes
| Stage | Patients | Purpose |
|---|---|---|
| Training | 6,443 | Develop the model |
| Retrospective validation | 22,251 | Evaluate recorded imaging cohorts |
| Single-arm clinical trial | 10,333 | Evaluate an additional AI reader in routine workflow |
In the trial, malignancy-diagnosis AUC was 0.952 (95% confidence interval 0.942–0.961), meeting its predefined primary endpoint. AI–human review identified 51 previously overlooked lesions, including 15 malignancies, and prompted 37 amended reports. These are the study’s observed workflow findings. Source: paper abstract.
The authors call for prospective comparative studies across different health systems to assess clinical outcomes. They also state that the complete LiON source cannot be released because of proprietary infrastructure and pending patent protection. The linked repository contains key components, not the entire deployed system. Source: conclusions and code availability.
What the public code tells researchers
The repository identified by the paper describes the Pixel-Lesion-pAtient Network, or PLAN, and links to a 2023 MICCAI study. Its approach combines pixel-level segmentation, lesion-level instance segmentation and patient-level classification in an end-to-end network.
The README describes outputs and evaluation at each of those levels, alongside configuration and data-preparation instructions. That makes the repository useful for understanding algorithmic components. Its own description should not be stretched into a claim that downloading it reproduces the entire 2026 clinical deployment.
What the trial cannot settle alone
Editorial interpretation: a single-arm study can reveal what happened after a tool entered a workflow, but it does not provide the same concurrent comparison as a randomized trial. A finding that triggered review is also different from an established improvement in survival, quality of life or net clinical benefit.
The AUC should remain labeled as AUC. Rewriting 0.952 as “95.2% diagnostic accuracy” would give readers a different metric without evidence. Likewise, the numbers of lesions, patients and amended reports should not be combined as if they counted the same event.
For readers following medical imaging AI, LiON is worth examining as a deployment study with explicit workflow outcomes and reproducibility limits. The next useful question is how this approach compares with usual practice under a defined evaluation, including any additional investigations it prompts. This research brief is not diagnostic advice.