On August 24, Alibaba’s DAMO Academy, in collaboration with Shengjing Hospital of China Medical University and other institutions, developed DAMO LiON, an AI model for diagnosing liver cancer. The model can identify tiny cancerous lesions in the liver from CT scans. During a two-month real-world prospective clinical trial, the AI model detected 15 malignant tumors that had previously been missed. Most of the lesions were approximately 1 centimeter in diameter, enabling patients to receive surgery or medication in a timely manner. The related paper was published in the leading international academic journal Nature Medicine.
The liver is a common site for malignant tumors. These include primary liver cancer, which originates in the liver and is commonly referred to as “liver cancer,” as well as liver metastases that spread from other sites, such as colorectal or pancreatic cancer. The latter are often more common. Regardless of the type, the earlier the disease is detected, the better the treatment outcomes.
However, early detection is not easy. Because lesions are small and can be obscured by cirrhosis, fatty liver, and the liver’s complex anatomy, doctors may miss them even with the help of contrast-enhanced CT. This is especially true in cases of liver metastases, where doctors’ attention may be drawn to the primary tumor, causing them to overlook small lesions in the liver.
To address this challenge, DAMO Academy drew on years of experience in medical imaging AI to develop the DAMO LiON model, which assists doctors in interpreting scans as an “AI safety officer.” The model can accurately identify primary liver cancer and is even more effective at detecting easily overlooked liver metastases.
The experimental results showed that the AI model was more accurate than radiologists in identifying malignant tumors. With the assistance of the AI model, doctors reduced their image interpretation time by 27%, while their sensitivity to malignant tumors increased by 11.5%, effectively reducing missed diagnoses. With AI assistance, junior doctors achieved diagnostic performance comparable to that of senior doctors.
The AI model identifies tiny metastatic tumors
The research team subsequently deployed the AI model in hospitals for use in routine image interpretation. When the AI’s preliminary findings differed from those of a doctor, the case was referred to a senior radiologist for review and, when necessary, escalated to a multidisciplinary team (MDT) for discussion.
Over the course of two months, the AI analyzed contrast-enhanced CT scans from more than 10,000 patients and helped doctors detect 15 previously overlooked cases of liver metastases, prompting adjustments to the patients’ treatment plans. In one case, a 65-year-old patient with a history of bladder cancer returned for a follow-up examination. The patient’s liver function indicators and tumor markers were normal, and the doctor’s initial assessment identified only calcified lesions in the liver. The AI, however, accurately identified liver metastases, prompting the doctor to amend the report and recommend chemotherapy.
Yan Ke, an algorithm expert at DAMO Academy, said that the malignant lesions detected by the AI generally shared three characteristics: they were “small, faint, and atypical.” Their average diameter was approximately 1 centimeter, they had relatively low contrast against the surrounding liver tissue, or they were located in less common anatomical positions.
According to the team, the DAMO LiON model uses an improved network architecture that can both capture the relationship between lesions and the liver as a whole and effectively preserve local textures and boundaries. This improves its performance in challenging cases involving fatty liver, cirrhosis, and postoperative changes in the liver. In addition, the AI can iteratively fuse images from different phase combinations, effectively capturing pixel-level differences between phases and accurately detecting tiny lesions that “flash by” during contrast-enhanced CT scans.
DAMO Academy has been working on medical AI since its establishment in 2017. It was also among the first in the industry to propose using AI to identify subtle lesions in medical images that are difficult for the human eye to detect. DAMO Academy has developed a number of models, including DAMO PANDA, an AI model for pancreatic cancer screening based on non-contrast CT; DAMO GRAPE, an AI model for gastric cancer screening; and DAMO COCA, an AI model for colorectal cancer screening. These achievements have been published in Nature Medicine three times, entered the innovation pathway of the National Medical Products Administration, and received the U.S. Food and Drug Administration’s (FDA) Breakthrough Device designation.
With this development, DAMO Academy has further expanded the capabilities of medical AI from screening to diagnosis, helping more patients receive timely and accurate treatment.
This article was provided by Alibaba and republished by QbitAI with authorization. The views expressed are those of the original authors.