Alibaba’s DAMO LiON AI Model Helps Detect 15 Missed Liver Tumors

Introduction

In liver imaging, the most difficult lesions are not always the largest ones. A lesion measuring around one centimeter, showing weak contrast with surrounding tissue, or appearing in an unusual anatomical location can be easy to overlook. Alibaba’s DAMO Academy has introduced DAMO LiON, an AI model designed to assist radiologists in finding such malignant liver lesions on contrast-enhanced CT scans. The related research has been published in Nature Medicine.

The diagnostic challenge

Malignant lesions in the liver include primary liver cancer as well as metastases originating from cancers such as colorectal or pancreatic cancer. Metastatic lesions can be particularly difficult to detect because radiologists may focus on the known primary tumor and pay less attention to subtle abnormalities in the liver. Cirrhosis, fatty liver, postoperative changes and complex anatomy can further complicate interpretation.

DAMO LiON is positioned as a second reader rather than a replacement for physicians. According to the source, its architecture captures the relationship between a suspected lesion and the overall liver while preserving local texture and boundary information. It also combines images from multiple contrast phases, allowing the system to identify small pixel-level changes that may appear only briefly during a CT examination. The target is a group of lesions described as “small, faint and oddly located.”

Results from clinical use

The research team deployed the system in a hospital workflow for a two-month prospective real-world trial. DAMO LiON reviewed contrast-enhanced CT scans from more than 10,000 patients and helped identify 15 malignant tumors that had not been recognized in the initial interpretation. Most were lesions measuring approximately one centimeter. The findings prompted timely surgery or drug treatment for some patients and led to changes in treatment plans.

The source also reports that the model performed better than the radiologists included in the comparison when identifying malignant tumors. With AI assistance, doctors spent 27% less time reading the scans, while sensitivity for malignant tumors increased by 11.5%. Junior doctors reached the level of senior doctors with assistance from the system. These figures should, however, be interpreted in the context of the reported study and deployment setting rather than assumed to apply uniformly across hospitals and patient populations.

A cautious human-AI workflow

The deployment model is notable because it keeps clinical responsibility with the medical team. When the AI output differs from the doctor’s initial conclusion, the case is sent to a senior radiologist for review and may be escalated to a multidisciplinary team. In this arrangement, the model acts as an alerting and error-reduction tool, while diagnosis and treatment decisions remain subject to clinical judgment.

DAMO Academy has previously developed imaging systems for pancreatic, gastric and colorectal cancer screening. DAMO LiON represents a move from screening toward diagnostic assistance. More broadly, it reflects a shift in medical AI from demonstrating that a model can detect abnormalities to proving that it can fit into real clinical workflows and deliver measurable benefits.

Important questions remain, including performance across hospitals and scanners, long-term stability, false-positive management and accountability when AI and clinicians disagree. If future multicenter studies reproduce the reported results, DAMO LiON could become a useful second-reading tool for subtle liver lesions. For now, it is better understood as clinical decision support, not an autonomous diagnostician.

Source: QbitAI