A new AI liver cancer system helped identify 51 liver lesions that had previously been overlooked on CT scans, including 15 that turned out to be malignant, according to a large clinical study.
Called the Liver Diagnosis Network, or LiON, the system analyses contrast-enhanced CT scans and incorporates clinical information to assess whether a liver lesion may be malignant. Researchers tested it across multiple centers before putting it into routine clinical practice in a single-arm trial involving more than 10,000 patients.
The results suggest that AI could serve as an additional safety net when doctors review medical images. But the study does not show that the technology improves survival or that it can replace radiologists. Its most interesting contribution may be simpler: giving doctors another chance to notice something that was initially missed.
Why Liver Lesions Can Be Difficult to Spot
Liver malignancies are often evaluated with contrast-enhanced CT, where the contrast agent helps differentiate tissues and makes abnormal areas easier to identify.
The challenge is that radiologists may have to review large volumes of complex images. Some abnormalities are obvious, while others are small or difficult to distinguish from surrounding liver tissue.
LiON was designed with this problem in mind.
Instead of relying on a single CT image, the system can process multiple phases of contrast-enhanced scans and integrate imaging information with clinical data. The researchers wanted to see whether that approach could work not just on carefully selected images but also across different hospitals and, eventually, within routine clinical care.
LiON Performed Strongly Across Different Groups
The researchers first trained LiON using data from 6,443 patients.
They then tested it retrospectively on 22,251 patients from multicentre and real-world cohorts. The system achieved an area under the receiver operating characteristic curve, or AUC, of 0.975 for detecting liver malignancy.
An AUC is a way of measuring how well a diagnostic system distinguishes between people with and without a condition. A score of 1 represents perfect discrimination, while 0.5 would be no better than chance.
LiON also maintained strong performance in patients with conditions that can complicate liver imaging. Its AUC was 0.971 in patients with hepatic steatosis (fatty liver) and 0.924 in those with cirrhosis.
Those results provided the basis for the researchers' next step: seeing what happened when LiON was actually incorporated into clinical practice.
The AI Became an Extra Reader
The researchers conducted a single-arm clinical trial involving 10,333 patients undergoing routine care.
In this setting, LiON did not replace the radiologist. Instead, it functioned as an additional AI reader within the existing workflow. The system reviewed the scans and provided information that clinicians could consider alongside their own assessment.
The system achieved an AUC of 0.952 in the clinical trial, meeting the study's predefined primary endpoint.
But the most tangible results came from cases in which AI and human readers worked together.
AI Helped Find 51 Previously Overlooked Lesions
During the trial, AI-human collaboration identified 51 lesions that had previously been overlooked.
Fifteen of those lesions were malignant.
The findings led to 37 amended radiology reports, and 22 cases were escalated to multidisciplinary teams. Clinical management was changed for some patients as a result.
That is where the study becomes particularly interesting.
An impressive accuracy score can show that an AI system performs well under testing conditions. Finding an abnormality that had already slipped through a real-world clinical workflow tells a different story.
It suggests the technology could potentially act as a second set of eyes — not making the final decision, but drawing attention to findings that deserve another look.
The Results Do Not Mean AI Can Replace Doctors
Despite the promising findings, LiON was not tested as an autonomous replacement for radiologists.
The study was single-arm, meaning there was no separate control group receiving standard care without the AI system. Because of that design, the researchers cannot determine from this trial alone how much LiON improved diagnostic performance compared with ordinary clinical practice.
The study also did not establish that using LiON leads to better survival, fewer complications or longer lives.
Those are important questions because identifying a cancer earlier only matters clinically if that earlier discovery ultimately changes what happens to the patient.
The researchers therefore say further prospective comparative studies across different healthcare systems are needed to determine whether the technology improves clinical outcomes.
What Makes This AI System Different?
AI has already been investigated for analyzing medical images, including liver lesions. What stands out about LiON is the scale of the evaluation and the attempt to integrate the system directly into an existing clinical workflow.
The researchers did not simply ask whether a computer could identify cancer in a collection of scans.
They asked whether it could work alongside clinicians in routine practice — and whether that collaboration could uncover findings that might otherwise remain unnoticed.
In this study, the answer was promising enough to justify further investigation.
The 51 overlooked lesions provide a particularly concrete example of what that collaboration could look like. The AI did not diagnose every patient on its own. Instead, it helped draw attention to abnormalities, prompting radiologists and clinical teams to take another look.
What Happens Next?
The biggest unanswered question is whether these additional detections will ultimately translate into better outcomes for patients.
Researchers will need to compare AI-assisted care directly with standard care and follow patients over time. They will also need to see whether LiON performs consistently in hospitals with different equipment, patient populations, and clinical workflows.
That evidence will determine whether the system becomes a useful part of everyday liver cancer diagnosis or remains primarily a promising research tool.
For now, the study offers a more grounded picture of AI's potential in medicine.
LiON did not replace the radiologist. It did not independently decide who had cancer. Instead, it provided another layer of scrutiny, and in 51 cases, that extra look uncovered lesions that had previously gone unnoticed.
Among them were 15 malignancies.
Whether catching those cancers changes patients' long-term outcomes remains to be seen. But the finding shows why researchers are increasingly interested in AI not as a replacement for doctors, but as a tool that can help them notice what human eyes may occasionally miss.
