Why This Study Matters
Liver malignancies are commonly evaluated with contrast-enhanced CT (CE-CT), but missed or delayed diagnoses remain a persistent challenge in busy, real-world radiology workflows. A prior Cochrane meta-analysis of 21 studies (3,101 patients) found that conventional CT alone has a sensitivity of just 77.5% (95% CI: 70.9–82.9%) for hepatocellular carcinoma — meaning it may miss the cancer in roughly 22.5% of affected patients. That gap is what the Liver DiagnOsis Network (LiON), a CE-CT-based AI system, was designed to help close by acting as a second reader inside the existing clinical workflow rather than replacing radiologists outright.
Study Design
LiON was developed and tested across three stages: model development in 6,443 patients, multicenter retrospective validation across 22,251 patients, and a prospective, single-arm real-world clinical trial registered as NCT07153783. In the prospective trial, LiON functioned as an additional AI reader alongside radiologists interpreting CE-CT scans — layered onto the existing workflow rather than tested as a standalone replacement.
| Metric | Detail |
|---|---|
| Patients (prospective trial) | 10,333 |
| Site | 1 — Shengjing Hospital, Shenyang, China |
| Randomization | None — single-arm design |
| Dosing | N/A — diagnostic AI, not a drug |
Site and single-center detail come from the ClinicalTrials.gov registration for NCT07153783 rather than the abstract itself, which does not state where the prospective trial was conducted.
Patient Population
The development and validation cohorts (28,694 patients combined) were drawn from multicenter and real-world cohorts in routine clinical practice. Per the ClinicalTrials.gov registration, the prospective single-arm trial itself was conducted at one site: Shengjing Hospital of China Medical University in Shenyang, Liaoning Province, China. The abstract does not report detailed inclusion or exclusion criteria beyond "routine clinical practice" patients undergoing CE-CT.
Primary Endpoint Results
The trial's primary endpoint was an AUC for malignancy diagnosis, with success pre-defined as the lower bound of the 95% CI exceeding 0.900.
The trial met its primary endpoint, defined as an AUC for malignancy diagnosis with the lower bound of the 95% CI exceeding 0.900, achieving an AUC of 0.952 (95% CI: 0.942-0.961).
— Verbatim from the published abstract, Nature Medicine
For context, the retrospective multicenter validation stage (22,251 patients) produced an even higher AUC, suggesting performance was well-maintained when the system moved from retrospective testing into live clinical use. Per the ClinicalTrials.gov registration, the primary outcome was assessed on a timeframe of up to 90 days from enrollment; the abstract itself does not separately state a data-cutoff or follow-up duration.
Subgroup Analyses
The abstract reports two subgroup results directly, both drawn from the retrospective validation analysis:
| Subgroup | AUC | 95% CI |
|---|---|---|
| Hepatic steatosis (fatty liver) | 0.971 | 0.952–0.985 |
| Cirrhosis | 0.924 | 0.901–0.946 |
No other subgroup-level statistics — by lesion type, tumor size, or patient demographics — were reported in the abstract or located in external sources for this run, so none are presented here.
Safety Profile
Not specified in source. LiON is a diagnostic AI-assistance tool, not a therapeutic intervention — neither the abstract nor the external sources identified for this run report adverse-event or safety data in the sense used for drug or device-treatment trials.
Interpretation and Broader Context
Used as a second reader rather than a replacement, LiON surfaced lesions that radiologists had initially missed — including small metastases — without requiring a change to the underlying CE-CT workflow. Its accuracy held up even in cirrhotic and fatty-liver patients, populations where lesion detection is traditionally harder. Measured against the Cochrane review's cross-study CT sensitivity and specificity benchmark for hepatocellular carcinoma (77.5%/91.3%), LiON's reported AUCs look considerably stronger — though these are not a head-to-head, within-trial comparison, since the Cochrane figures come from a different, unrelated pool of studies and populations.
Limitations
The study authors are explicit about this study's boundaries: the prospective trial was a non-randomized, single-center design with no comparator arm, run only at Shengjing Hospital in Shenyang, China. It did not report survival or other downstream clinical outcomes, so it cannot show whether AI-assisted reading changes how patients ultimately fare — only that it changed what radiologists saw on the scan. The authors call for prospective, comparative studies across more diverse healthcare systems before drawing conclusions about real-world patient-outcome impact.
In a prospective, single-arm, single-center trial of 10,333 patients, an AI second reader (LiON) met its pre-specified accuracy bar for diagnosing liver malignancy on contrast-enhanced CT and surfaced dozens of lesions radiologists had initially missed, without requiring a change to the existing CT workflow. Because the trial was non-randomized, ran at one hospital, and did not track survival or other downstream outcomes, it establishes that the AI reads scans accurately alongside radiologists — not yet that doing so changes how patients ultimately fare.