Alibaba's AI Model Detects 15 Missed Liver Cancers in Nature Medicine Study

Alibaba's research laboratory, working alongside physicians at Shengjing Hospital of China Medical University, built an artificial intelligence system called DAMO LiON designed to spot liver cancer in medical scans. When they tested it in a real hospital setting over two months, analyzing contrast-enhanced CT images from more than 10,000 patients, the model found something striking: it caught 15 cases of liver metastasis that the radiologists had missed. Those 15 patients, as a result, were able to adjust their treatment plans and access surgical or pharmaceutical options they might otherwise have gone without. The findings appeared in Nature Medicine, a signal that AI's role in cancer detection is moving from laboratory promise into clinical reality.

The lesions that slip past human eyes tend to share a profile. They are small. They are faint. They sit at the periphery of the liver, where attention naturally wanes. Yan Ke, an algorithm specialist at DAMO Academy, explained that the team rebuilt the model's architecture to handle exactly this problem. The approach involved taking imaging data from different time points and fusing them together, then training the system to spot pixel-level differences that signal the presence of a lesion. The result was a model sensitive enough to anchor onto the minute abnormalities that human radiologists, working under the ordinary pressures of clinical volume, tend to overlook.

Beyond the raw count of missed cases, the trial revealed how AI changes the work itself. When radiologists used the system as a second reader—a redundant check running alongside their own assessment—their reading time fell by 27 percent. Their sensitivity for detecting malignant tumors rose by 11.5 percent. Perhaps most tellingly, junior physicians supported by the AI performed at the level of senior radiologists. The technology did not replace expertise; it redistributed it, allowing less experienced doctors to work at the standard of their seniors.

The same laboratory developed a separate AI model for colorectal cancer called DAMO COCA. In testing, it achieved 86.6 percent sensitivity for detecting lesions and 99.8 percent specificity. When compared directly against ten radiologists of varying experience, the model's sensitivity exceeded theirs by more than 20 percentage points. Colorectal cancer screening has long struggled with patient compliance. The standard fecal occult blood test requires people to collect their own samples. Colonoscopy demands bowel preparation and the insertion of an endoscope—a combination of physical and psychological discomfort that keeps many people away. The DAMO team proposed an alternative: non-contrast CT imaging paired with AI analysis. By training the model on different intestinal segments separately and leveraging imaging data from the same patient across different time periods, they created a screening tool that requires no bowel preparation and no invasive procedure.

While Alibaba's work focused on imaging, other institutions were advancing AI's role across the full arc of cancer treatment. Moderna and Merck jointly developed intismeran autogene, a personalized mRNA cancer vaccine that succeeded in Phase 3 trials—the first personalized mRNA cancer vaccine to meet its primary endpoint at that stage. The process begins after tumor removal. Doctors sequence the tumor's DNA and RNA, then use AI to identify which mutations are present and which ones the immune system is most likely to recognize. That information gets encoded into mRNA molecules wrapped in lipid nanoparticles, creating a vaccine tailored to one patient's specific cancer. Kim Hee-soo, Vice President of Moderna Korea, noted that even two patients with the same melanoma diagnosis carry mutations as distinct as fingerprints. The vaccine essentially hands the immune system a target list—the specific cancer cell signatures it should attack. The challenge is speed. From tissue collection to first dose takes six to eight weeks, a timeline that encompasses analysis, design, manufacturing, and quality testing. Moderna is working to compress that window while also exploring expansion applications, including a technology that delivers mRNA directly into the body to reprogram immune cells in place, potentially eliminating the complex workflow of traditional CAR-T therapy.

In the United Kingdom, a patient named Reece Hibbert underwent brain surgery for a tumor in his pituitary gland, a lesion just one millimeter across that overlapped with his optic nerve and carotid artery. The surgical risk was severe—preoperative estimates put the chance of residual tumor at 25 to 50 percent and the risk of major vascular injury at 0.5 to 2 percent. The surgical team used an AI system that analyzed real-time video from an endoscope threaded through the nasal cavity to the skull base. The system tracked instrument positions and highlighted the probable locations of blood vessels, nerves, and critical tissues, as well as the safer margins for tumor removal. Eight weeks after surgery, Hibbert had recovered nearly full vision. The BBC reported the case on August 27.

Taken together, these developments sketch a landscape where artificial intelligence is no longer a peripheral tool but infrastructure spanning the entire medical journey. DAMO Academy's imaging systems handle screening. Moderna's personalized vaccines customize treatment. AI-assisted surgery guides the surgeon's hand. China's advantage in this space is substantial: a massive population has generated vast repositories of imaging data from both healthy individuals and patients with specific diseases, providing the training material that medical AI systems require. As more prospective clinical trials accumulate real-world evidence, the transition from academic validation to large-scale deployment is accelerating.