Published in Nature Medicine on Aug. 19, a prospective trial on 10,333 patients ran a liver sensing model alongside radiologists, in the hospital's actual workflow.
The system is called LiON, for Liver DiagnOsis Network. It reads the injected CT scans, the reference examination when a liver tumor is suspected, and it was connected not to an archive database but to the current activity of a facility, Shengjing Hospital. The results were published in Nature Medicine on August 19, 2026.
The construction of the model followed three stages: training on 6,443 patients, a multicenter retrospective validation on 22,251 files, and then the decisive test, a prospective trial registered with 10,333 patients with a performance objective set in advance. The target was an area under the curve greater than 0.900; the system reached 0.952.
A safety net, not a replacement
The device places LiON as a second reader, next to the radiologist and not in his place: the model signals, the doctor referees and possibly revises his report. Over the duration of the trial, 51 lesions absent from the initial reports were reported, 15 of which were found to be cancerous. Thirty-seven reports were corrected and 22 files were referred to a multidisciplinary consultation meeting. This is one of the few studies to measure what the AI actually catches up in clinical conditions, rather than its performance on a fixed set of tests.
Read the full article on the original source
Go to the source →