Deep Dive — AI co-reader caught 15 liver cancers radiologists missed

Alibaba's research arm, DAMO Academy, has published a medical-imaging AI model in Nature Medicine that reads the contrast-enhanced CT scans radiologists already order and flags liver cancers the human eye missed. In a two-month deployment inside a routine clinical workflow, the model helped clinicians find 15 malignant tumors — most of them around a centimeter in diameter — that had been overlooked, and changed treatment plans for patients as a result.

We should be careful about what this is and what it isn't. This is not another demo of an AI that outperforms doctors on a clean benchmark. It is the quieter, arguably more consequential pattern that has defined medical AI's real wins over the past few years: a model acting as a second reader — a "safety net" that sits alongside the radiologist rather than replacing her — and catching the small, faint, off-center lesions that slip through even experienced eyes in a busy clinic. The model is called LiON, short for Liver DiagnOsis Network, and it is the strongest evidence yet that this particular deployment model works in live hospital conditions.

From screening to diagnosis: DAMO's long bet on reading images

DAMO Academy launched its medical AI program in 2017, and its approach has always been conspicuously different from the tech superstar who wants to become your doctor. Instead of building a generalist chatbot, its medical lab focused on a narrow, hard problem: detecting tiny cancers in CT scans that humans tend to miss. The strategy produced a now-familiar family of models, all named after animals and all repeatedly published in Nature Medicine — PANDA for pancreatic cancer screening on non-contrast CT, GRAPE for gastric cancer, COCA for colorectal cancer. LiON is the latest entry, and it moves the program in a new direction: away from screening asymptomatic people and toward diagnosis in routine hospital workflows.

The distinction matters. DAMO's earlier models were screening tools, aimed at catching disease in people who feel healthy, often using a cheap non-contrast CT as part of a regular health checkup. LiON works the other way — it reads the contrast-enhanced CT (CE-CT) scans already being performed on patients who are sick, chronically ill, or being followed up for another cancer, and tries to make sure the report is correct. That is diagnosis-support work, and clinically it is where the demand is highest. Liver malignancies are a leading cause of cancer death globally, and the difficulty is compounded by anatomy: primary liver tumors sit inside an organ that is frequently inflamed or scarred (cirrhosis, fatty liver), while liver metastases — cancer that has spread from the colon, pancreas, or other sites — are actually more common than primary liver cancer and are neurologically easy to overlook because a doctor's attention is drawn to the original tumor. Lesions are often small, low-contrast against surrounding tissue, or tucked into uncommon anatomical positions.

That is exactly the failure LiON targets. In its paper, the authors describe the lesions it rescued as "small, faint, off-center" — average diameter around a centimeter, low contrast, awkward locations. These are precisely the findings that vanish in a high-volume radiology queue.

What LiON actually is, technically

On the engineering side, LiON is a break from the single-network pattern. Rather than one model glancing at one slice, it couples an improved architecture that captures both the relationship between a lesion and the whole liver and the local texture and edges that let a radiologist actually see a lesion as a lesion. That combination is what makes it robust in hard cases — fatty liver, cirrhosis, post-surgical livers — where the surrounding tissue noise hides the tumor. It also fuses image data across multiple phases of the CE-CT scan, aligning pixel-level differences between phases to catch micro-lesions that appear in only a single "flash" of the scan. The multiphase processing is a deliberate design trait: it means LiON can work with the flexible, somewhat messy imaging protocols real hospitals actually produce, rather than requiring a rigid study.

The numbers behind it are substantial. The study team trained LiON on 6,443 patients and validated it retrospectively across 22,251 patients drawn from multicenter and real-world cohorts, reporting an area under the receiver operating characteristic curve (AUC) of 0.975 for malignancy diagnosis. Performance held up in difficult subgroups — an AUC of 0.971 among patients with hepatic steatosis (fatty liver) and 0.924 among those with cirrhosis, which is precisely the population where radiologists struggle most.

The headline result is a single-arm prospective trial in 10,333 patients treated in routine clinical practice, in which LiON ran as an additional AI reader inside the normal workflow. It met its pre-specified primary endpoint — an AUC whose 95 percent confidence lower bound exceeded 0.900 — landing at an AUC of 0.952 (95 percent CI 0.942–0.961). More meaningful than the metric is what happened in the clinic: over the trial, AI–human collaboration surfaced 51 previously overlooked lesions, 15 of them malignant, prompted 37 amended radiology reports, and triggered 22 multidisciplinary team escalations that changed how patients were managed. In one case reported by QbitAI, a 65-year-old follow-up patient with a bladder cancer history had normal liver function tests and tumor markers; the initial read flagged only a calcification, but LiON's judgment pushed clinicians to amend the report and recommend chemotherapy for a metastasis they had missed.

The read-speed findings are the kind that matter in a real department. With the AI as a co-reader, interpretation time dropped by 27 percent and sensitivity for malignancy rose by 11.5 percent — and junior physicians reading with the AI reached the performance level of senior radiologists, flattening the steep learning curve that large teaching hospitals normally trade on.

Why the "second reader" is where medical AI is actually winning

Step back and LiON is a useful lens on a broader truth that often gets lost in the AGI discourse: the medical AI that is demonstrably working in the real world is almost never the AI that "replaces" a clinician. It is the AI that becomes load-bearing infrastructure for one. When a model's only job is to sit in the workflow and flag discrepancies — "the human looks, then the machine looks, and anything they disagree on goes to a senior for arbitration" — the failure mode that has made autonomous medical AI so slow to land becomes something you can engineer around.

Two MIT-style observations underline it. First, a second reader is allowed to be imperfect and still be enormously valuable, because it does not need to be right all the time — it just needs to be right on the cases the human is wrong on, and a human who is told "go check this again" is a far better adjudicator than an algorithm alone. Second, the economics work: you are not paying to put a GPU in place of a radiologist, you are paying a modest inference cost to stop the expensive, emotionally ruinous event of a missed cancer. That is an argument hospitals can actually make. We explored why models can be confidently wrong in ways that matter for exactly this kind of stakes, in our explainer on what an AI hallucination is — a reminder that sensitivity promise is not the whole story.

DAMO has spent eight years turning that thesis into a pipeline rather than a one-off paper. Its models have repeatedly been flagged as breakthrough medical devices by the US FDA, entered China's National Medical Products Administration (NMPA) innovative-device channel, and — according to the company — powered more than six million screenings in China via a "one scan, multiple checks" approach to non-contrast CT. LiON extends the same playbook from broad screening into the higher-value diagnosis segment.

The skeptical read: what the paper still doesn't prove

A paper this clean deserves scrutiny. The most important caveat is baked into the trial design: it is single-arm, with no randomized comparison against a control workflow, and the authors themselves close by saying that "further evidence from prospective comparative studies across diverse healthcare systems is warranted." A single-arm deployment showing that LiON plus an escalation protocol catches extra cancers is compelling evidence of an effect, but it cannot cleanly isolate how much of that effect comes from the model versus from simply adding a second layer of re-review — because the human–AI collaboration numbers, not standalone AI performance, drove the 15-caught headline.

There is also the matter of provenance. The research was largely run and funded by the company that stands to sell the product; Alibaba holds patents on the detection methods, and the competing-interest statement discloses that several authors are Alibaba employees with stock. Nature Medicine peer review is a meaningful filter — three of these models have cleared it — but it does not remove the commercial incentive, and the FDA's own rulemaking around just how carefully generative and autonomous imaging tools should be vetted is still very much in flux, as we covered when the FDA floated doctor-style competency exams for medical AI devices.

And then there is generalizability. LiON was trained on Chinese multicenter data and validated in Chinese hospital workflows; the droop in the cirrhosis subgroup (0.924 vs 0.975 overall) is a reminder that even within the population it was designed for, it is weaker exactly where the need is greatest. Whether it transfers to other countries' imaging protocols, device brands, and patient mixes — where scanner settings and screening norms differ — is an open question that only foreign prospective trials can answer. The honest framing is that LiON is a high-quality, vendor-backed demonstration of a plausible and broadly useful pattern, not yet a proven universal product.

What to watch next

Three threads are worth following. First, whether a prospective, randomized comparative trial follows — ideally in a non-Chinese health system — since that is the only evidence that would convert "caught 15 in one deployment" into a generalizable clinical claim. Second, whether LiON or a successor clears NMPA and pushes the FDA breakthrough designation toward US market entry; the way a diagnostic co-reader navigates the emerging GenAI-competency regulatory framework will set a template for the whole category. Third, watch the product direction: DAMO has spent a decade arguing that "one scan can screen for many cancers," and LiON's success in the diagnosis tier strengthens the case that liver — one of the trickiest organs to image — can become a routine screening target rather than a diagnostic afterthought. That is the long game, and it is bigger than any single AUC.

Catching a cancer that was going to be missed is a small number of people in a trial — 15 patients, most with tumors barely a centimeter wide — but each one is a person who got treatment instead of a delayed diagnosis. The careful, unglamorous work of making the AI second reader both precise enough and cheap enough to run on every enhanced CT scan may end up mattering far more than a flashier model that promises to put a doctor out of a job. The question now is whether the evidence holds up outside one company's own hospitals.

Do you trust an AI second reader that catches what a radiologist missed — or does the single-arm evidence leave you unconvinced? Tell us in the comments.

Sources: Nature Medicine via PubMed · QbitAI · DAMO Med · South China Morning Post · ClinicalTrials.gov