Alibaba DAMO Academy, in collaboration with Shengjing Hospital and other institutions, has developed the liver cancer diagnostic AI model DAMO LiON, which can accurately identify small liver lesions through enhanced CT images. In a two-month real-world prospective trial, the model detected 15 malignant tumors that had been previously overlooked, most of which were around 1 cm in size. The related paper was published in Nature Medicine.
The AI Safety Officer, from "missing" to "filling the gaps"
Liver cancer and more common liver metastases are extremely difficult to detect in the early stages. Small lesions are easily disturbed by liver cirrhosis, fatty liver, and complex anatomical structures, and doctors' attention is often drawn away by the primary lesion. DAMO LiON acts as an "AI safety officer" to assist in reading images, with a higher accuracy rate in identifying malignant tumors than radiologists; when used together, it reduces reading time by 27% and increases sensitivity by 11.5%, allowing junior doctors to reach the level of senior doctors. The model uses an improved network architecture, balancing the relationship between lesions and the entire liver while preserving local texture boundaries. It can also integrate images from different phases to capture fleeting small lesions.
In practical deployment, if the AI's initial diagnosis conflicts with the first assessment, it is reviewed by senior doctors, and multidisciplinary discussions are escalated when necessary. Within two months, it analyzed over 10,000 patients, identifying 15 liver metastases that had been overlooked. A 65-year-old bladder cancer patient, whose tumor markers were normal and whose initial diagnosis showed only calcification, was correctly identified by the AI as a liver metastasis, leading to a change in chemotherapy. Since 2017, DAMO Academy has been focusing on medical AI, expanding its capabilities from screening to diagnosis, enabling more patients to receive timely and accurate treatment.
