Quick takeaways
- LiON analyzes contrast-enhanced CT scans and clinical information for possible liver malignancy.
- In routine practice, AI-human review identified 51 overlooked lesions, including 15 malignancies.
- Some findings led to amended reports, specialist-team review and changes in patient management.
- The clinical study lacked a randomized control group, so the outcome benefit remains uncertain.
A radiologist can read thousands of images in a normal workweek. Subtle liver lesions sometimes disappear into that volume, especially on complicated scans or examinations ordered for another reason.
A 2026 study tested the Liver DiagnOsis Network, or LiON, as an additional reader. LiON combined multiphase CT imaging with clinical information to identify possible liver malignancies. The researchers also studied what happened when it disagreed with reports in routine care.
From validation to the clinical workflow
LiON was trained using 6,443 patients and retrospectively evaluated across more than 22,000 patients from multiple cohorts. It maintained strong diagnostic performance in patients with fatty liver disease and cirrhosis, two settings that can make liver imaging harder to interpret (LiON investigators).
The researchers then placed the system into routine practice as an additional reader for 10,333 patients. The radiologist still issued the report, and LiON could flag a lesion for another look.
What the system caught
AI-human collaboration identified 51 previously overlooked lesions. Fifteen were malignant. The review led to 37 amended radiology reports and 22 cases being escalated to multidisciplinary teams. Some patients had their clinical management changed.
The 15 malignancies represent a small share of 10,333 patients. For those patients, though, a second check may have changed when treatment started and which options were available.
I can see a practical role here. LiON reviewed every eligible scan without requiring a second radiologist to reread the entire list.
What this study cannot tell us
The clinical portion was a single-arm trial. Everyone received the AI-supported workflow, so there was no randomized group experiencing normal care at the same time. We do not know how many reports might have been corrected later through ordinary follow-up, or whether earlier detection improved survival.
A safety-net system can also create false alarms. Every flag takes attention, and follow-up testing can expose patients to cost, radiation and worry. A useful second reader has to catch enough meaningful misses without overwhelming clinicians with low-value disagreements.
The study was conducted within particular hospitals and patient populations. Independent validation elsewhere will be needed before assuming the same performance across scanners, reporting styles and rates of liver disease.
My Thoughts
I trust this use of AI more than systems that try to make the whole decision. The radiologist makes the main interpretation, LiON checks every case, and a clinician remains responsible for the report.
I want the next trial to compare this workflow with ordinary care and report how often alerts change treatment, how many extra tests they create, and whether cancers are treated earlier. The 15 malignancies found here justify running that trial.