Radiologists in busy hospitals read stacks of liver CT scans every day, and missed or delayed diagnoses remain a real problem when the workload is that high. A small lesion on one slice can go unnoticed, and liver cancer that is caught late is harder to treat.
The clever bit
A team led by researchers in China, working with Alibaba's DAMO Academy and partners in France and the UK, built LiON, an AI system that reads contrast-enhanced liver CT scans and flags likely cancers. It was trained on 6,443 patients and tested on 22,251 more. Then it went into routine practice as an extra reader alongside the radiologists. Across those earlier tests, LiON scored an AUC of 0.975 for spotting cancer, and it held up in patients with fatty liver disease and with cirrhosis, two groups where scans are harder to read.
In a single-arm trial covering 10,333 patients, the AI and clinicians working together identified 51 lesions that had been overlooked, 15 of them cancers. Those finds led to 37 amended radiology reports and 22 referrals to specialist teams, and changed care for some patients. The trial met its main accuracy target. That target was an accuracy score whose lower bound had to clear 0.900. LiON reached 0.952.
Hooray
Fifteen liver cancers that might have slipped by were caught while the system ran quietly in the background of normal hospital work. It's a safety net that fits into the existing workflow instead of replacing anyone.
What comes next
The trial had no comparison group, so it can't show whether patients lived longer or fared better. The authors say comparative studies across different health systems are needed. Some authors work for Alibaba, which has filed related patents. The paper is peer-reviewed in Nature Medicine.