Liver cancer: used as a second reader on 10,333 CT scans, LiON identified 51 lesions that had previously gone undetected

Trained on 6,443 patients and validated on over 22,000, this AI system met its objective in real-world conditions and resulted in 37 corrected radiology reports.

Published in Nature Medicine, the study Lion (Liver DiagnOsis Network), an AI system that analyses scanners with contrast agent injection to detect malignancies in the liver. It can process several phases of acquisition, integrate clinical data, and integrate into the usual radiology services circuit, where large volumes promote missed or late diagnosis.

From retrospective validation to routine testing

Trained out of 6,443 patients and subsequently validated in 22,251 multicentre and common practice patients, LiON achieved a AUC of 0.975 to distinguish malignancies. It remains effective in patients with hepatic steatosis (0.971) and, to a lesser extent, cirrhosis (0.924).

The team then tested it in a single-arm trial involving 10,333 patients, where it acted as an additional reader alongside the radiologists. The main objective is achieved with a AUC of 0.952. Most importantly, the collaboration between AI and physicians revealed 51 previously unseen injuries, including 15 malignancies, resulting in 37 modified reports and 22 referrals to multidisciplinary meetings. The authors point out that prospective comparative trials in a variety of health systems are still needed to measure the effect on patient fate.