Alibaba's medical AI catches missed liver tumors

The liver is notoriously quiet when it is failing, and by the time symptoms show up, options are often grim. Even on contrast-enhanced CT scans, things like cirrhosis, fatty liver, and messy vascular anatomy create visual clutter that causes even veteran radiologists to miss tiny early lesions.

Alibaba DAMO Academy, working alongside Shengjing Hospital of China Medical University and other partners, just unveiled DAMO LiON (Liver DiagnOsis Network), a dedicated liver cancer diagnostic AI model published in Nature Medicine.

This is built as a workflow-compatible diagnostic assistant inside real clinical environments. In trials, it cut physician scan interpretation time by 27% and pushed malignancy detection sensitivity up by 11.5%.

Catching 1-centimeter hidden tumors

In a two-month, real-world prospective clinical trial involving over 10,000 patients, DAMO LiON flagged 51 confirmed lesions that human doctors had completely overlooked.

That haul included 15 small metastases originating from outside the liver, such as colorectal or pancreatic cancers. Crucially, most of these missed spots were tiny, averaging just 1 centimeter across. In oncology, catching a lesion at 1 centimeter instead of 3 centimeters is often the difference between a clean resection and palliative care.

The system even matched or occasionally outperformed solo radiologists on complex cases. More practically, junior radiologists using the tool produced diagnostic evaluations comparable to seasoned specialists.

Spotting lesions that vanish in seconds

During contrast-enhanced CT scans, certain aggressive tumors only flare up for a few seconds as the contrast agent rushes through the liver, quickly disappearing back into the background tissue.

To catch these phantoms, DAMO LiON balances global liver anatomy with micro-level boundary textures. This keeps it stable even when dealing with heavy cirrhosis or fatty tissue that typically confuses automated image recognition.

On top of that, the model iteratively fuses multi-phase CT scans, calculating pixel-level variations across different scan phases to pinpoint lesions that show up only momentarily.

This builds directly on DAMO Academy's prior clinical line-up, which includes DAMO PANDA for pancreatic cancer, DAMO GRAPE for gastric cancer, and DAMO COCA for colorectal screening. With LiON, Alibaba is shifting from pure population screening directly into deeper, high-stakes diagnostic decision support.

Why it matters

For patients and junior hospital staff, this tackles radiologist burnout and regional inequality. High-level radiologists take decades to train, and reading hundreds of complex CT scans a day invites fatigue-driven oversights. By lopping off more than a quarter of the reading time while elevating junior doctors to specialist performance, this tool gives community and regional hospitals immediate access to top-tier diagnostic firepower without forcing patients onto months-long waitlists at elite urban centers.

For the health-tech sector, Alibaba is setting a template for how tech giants should tackle medical AI. Instead of pitching generic chat assistants or flashy theoretical models, DAMO Academy is tackling ugly, multi-phase clinical radiology bottlenecks and proving real utility through massive prospective hospital trials.

We think the main hurdle moving forward will be hardware heterogeneity. Pixel-level multi-phase CT fusion is extraordinarily sensitive to variations in hospital scanner calibration, contrast injection timing, and image artifacts. If Alibaba can maintain this level of sensitivity across lower-end scanners in rural clinics, DAMO LiON could genuinely shift clinical standards for early cancer detection.