← Back to Home

A Second Look at Liver Imaging: LiON-Assisted Reading Finds 15 Previously Missed Malignant Lesions

A prospective trial involving more than 10,000 people brought liver imaging AI into routine image interpretation. The study showed that human–AI collaboration could identify previously overlooked lesions, but whether it can improve patients’ long-term outcomes still requires comparative trials.

By SURL BioNews

A small, inconspicuous lesion on a liver scan can influence subsequent cancer diagnosis and treatment. As radiology departments handle large volumes of computed tomography images every day, adding another check for easily missed details is a concrete test for medical AI. In a prospective trial involving 10,333 people, human–AI collaboration using the liver diagnostic system LiON identified 51 previously overlooked lesions, 15 of which were malignant.

The system was jointly developed by Alibaba DAMO Academy, Shengjing Hospital of China Medical University, and other institutions. The study was published in *Nature Medicine* on August 19, 2026. It processes contrast-enhanced computed tomography images to help identify primary liver malignancies and lesions from cancers in other organs that have metastasized to the liver. Its use involves patients who have already undergone imaging, so the findings cannot be directly extrapolated to cancer screening effectiveness in the general population.

Tiny lesions are challenging because they may be clearer only during specific contrast phases, while fatty liver, cirrhosis, and complex anatomy can also increase the burden of interpretation. According to Alibaba, LiON integrates images from different phases and analyzes both the entire liver and local textures, using subtle differences across phases as clues for identification.

The researchers first trained the model using data from 6,443 patients, then retrospectively validated it using multicenter and real-world data covering 22,251 people. The subsequent prospective trial used LiON as an additional image-reading tool within the existing workflow, achieving an area under the curve (AUC) of 0.952 for identifying malignant lesions. This metric reflects the system’s ability to distinguish malignant from nonmalignant cases; it cannot be interpreted as meaning that “95.2% of patients receive a correct diagnosis.”

What happens after lesions are found is more closely tied to clinical value than a single performance figure. The paper’s abstract reported that human–AI collaboration led to revisions of 37 imaging reports, 22 referrals for multidisciplinary team discussion, and changes in the management of some patients. An Alibaba Cloud article separately described 15 of the lesions as small metastatic lesions, while the paper’s abstract used the broader classification of “malignant lesions.” It is therefore inappropriate to describe all 51 lesions as missed cancers.

Alibaba also reported that AI assistance reduced image interpretation time by 27%, increased sensitivity for detecting malignant lesions by 11.5%, and enabled less experienced physicians to approach the diagnostic performance of senior experts. However, the company’s account did not clearly state whether the sensitivity increase was a relative percentage or a percentage-point change, and the improvements seen in image-reading tests should not be directly applied to all hospitals.

This prospective trial used a single-arm design and is not yet sufficient to establish how much the system can reduce delayed diagnoses or improve survival compared with usual care. The burden of additional examinations triggered by false positives, applicability across different healthcare systems, and conditions for deployment according to approved uses also require clarification. At present, the most concrete evidence is that AI can prompt physicians to reexamine suspicious findings within routine imaging workflows. Whether these prompts can consistently translate into patient benefits remains to be verified in subsequent comparative studies.

References

  1. finance.biggo.com
  2. Alibaba Group
  3. Alibaba Cloud Community