A 65-year-old man returned for a routine follow-up after bladder cancer treatment. His liver function was normal. His tumor markers were normal. On the contrast-enhanced CT scan, the radiologist on duty saw what looked like a harmless calcium deposit and wrote it up as such. Then a piece of software looking over the same images disagreed. It flagged a small spot in the liver as a metastasis. The report was revised, chemotherapy was recommended, and the treatment plan changed, as described in the study's case report.
That quiet disagreement is the story behind a study published in Nature Medicine in August 2026. The model is called DAMO LiON, short for Liver DiagnOsis Network, and it was built by Alibaba's DAMO Academy (达摩院) together with Shengjing Hospital of China Medical University (中国医科大学附属盛京医院) and other institutions. It is among the earliest AI systems for liver malignancy diagnosis we could identify that was tested not just on stored images, but inside the daily reading workflow of a real hospital.
What LiON actually does
The liver is a favorite place for cancer to hide. Primary liver cancer is one thing, but metastases that spread from the colon, pancreas or bladder are more common, and they are easy to overlook. A lesion can be a centimeter wide, low in contrast against surrounding tissue, and tucked into an awkward corner of the anatomy. When a patient already has a known tumor elsewhere, a radiologist's attention naturally goes there first. The small spot in the liver becomes a "small, faint, off-to-the-side" finding that slips through.
LiON was trained to be a second pair of eyes on contrast-enhanced CT (增强CT) scans — the standard imaging used to stage known cancers. Its architecture does two things at once: it keeps track of how a suspicious spot relates to the whole liver, and it preserves the fine local texture and edges of that spot. It also iteratively fuses the multiple phases of a contrast scan, comparing them at the pixel level to catch lesions that appear only briefly as the dye flows through.
The result, in the experiments reported by the team, is a system that sometimes identifies malignant tumors more accurately than radiologists reading alone — especially in hard cases involving fatty liver, cirrhosis or complex anatomy.
What the Numbers Show
Two numbers from the study carry most of the practical weight, and one finding carries none of the arithmetic.
- Reading time fell 27%. When doctors used LiON as an aid, they spent roughly a quarter less time interpreting each scan. In radiology departments where backlogs are measured in days, that is not a cosmetic gain.
- Sensitivity rose. The model helped doctors catch more of the malignancies that were actually there, reducing false negatives — the misses that matter most to patients. The study reports the improvement as 11.5%; from the materials I could read, I could not establish whether that is 11.5 percentage points or an 11.5% relative gain, so treat the size of the effect as unresolved.
- Junior doctors improved to senior level. With the system's help, less-experienced radiologists reached diagnostic performance comparable to senior specialists.
Those figures come from the controlled part of the research. The more telling test came afterward, when the team deployed the model into routine clinical reading rather than a lab.
Two months in the real world
For two months, LiON ran alongside radiologists on enhanced CT scans from more than 10,000 patients. When the model and the doctor's first read disagreed, the case went to a senior radiologist for review, and if needed to a multidisciplinary team (MDT) discussion.
Over that period, the system surfaced 15 previously overlooked small metastases, averaging about one centimeter across, and changed the treatment plans of the patients involved. The 65-year-old bladder cancer patient was one of them. Alibaba's own English newsroom describes a broader count of 51 confirmed, previously overlooked lesions in that deployment, of which the 15 small metastases were a part; the Chinese-language reports of the same study emphasize the 15 cases. Either way, the pattern is the same: the AI was most useful precisely where human attention is weakest.
Why "second reader" is the phrase to watch
It would be easy to read this as "AI beats doctors." That is not what the study shows, and it is not what the researchers claim. The design puts the model in a supporting role: it does not issue a diagnosis, it raises a flag that a human then verifies.
That distinction matters for how the technology spreads. A system that replaces a radiologist has to be nearly perfect and carries enormous legal and ethical baggage. A system that quietly double-checks every scan — a "safety net" that never gets tired — can be added to existing workflows without asking doctors to change how they work. The same logic sits behind the other oncology models from DAMO Academy: DAMO PANDA for pancreatic cancer, DAMO GRAPE for gastric cancer, DAMO COCA for colorectal cancer. Together they point to what the academy calls a "one scan, multiple checks" (一扫多查) approach, in which a single CT exam is mined for several cancers at once.
For patients, the appeal is concrete. Nobody wants a second scan, a second appointment or a second dose of radiation. If the images already exist, asking a large model (大模型) to look at them again costs the patient nothing but time the machine spends computing.
Honest limitations
- The headline figures — 27%, 11.5%, 15 metastases — come from the developers and their published paper. They have not been independently replicated outside the study team, and the study was led by institutions with a direct role in building the model.
- The count of overlooked lesions differs between the company's English newsroom (51 lesions, including 15 small metastases) and Chinese reports (15 cases). The exact definitions of what was counted are not fully reconciled here.
- A single-arm trial at participating hospitals cannot establish how the model performs across every scanner, clinic or patient population. Community hospitals with different equipment may see different results.
- The study measures detection and reading speed, not patient survival. Whether catching these lesions earlier changes outcomes is plausible but was not the endpoint of this work.
What readers can do now
- If you have had cancer elsewhere in the body, ask your doctor whether your follow-up imaging specifically assessed the liver — metastases there are small and easy to miss.
- When you get a CT report, keep a copy of the images themselves, not just the written conclusion. Findings can be re-examined, and a second read needs the original scans.
- Treat AI "second reader" tools as a prompt for a conversation with a specialist, not a verdict. A flagged spot still needs a clinician to decide what it means for you.
