Alibaba unveils MAOSS, an AI model for early detection of fatty liver disease

Alibaba DAMO Academy has introduced MAOSS, an AI model that looks for signs of fatty liver disease in routine non-contrast CT scans and helps select patients for in-depth diagnostics earlier. For medicine, this is an important case that is not about "AI instead of the doctor" but about getting more value out of examinations that already exist.

Why it matters

Fatty liver disease, which the English-language literature increasingly calls steatotic liver disease, can go without symptoms for a long time. As a result, patients often reach a doctor only at the stage of fibrosis or cirrhosis, when the window for early intervention narrows considerably. According to Alibaba DAMO Academy, the disease currently affects about 30% of the world's population and could rise to 55.7% by 2040.

Standard ultrasound and serum markers are not always sensitive enough, and more accurate methods are not available in every clinic. MAOSS targets exactly this gap.

The model was developed together with Shengjing Hospital of China Medical University and Nanjing Drum Tower Hospital, and the results were published in Nature Communications. The system combines several types of data: non-contrast CT scans, blood values, and computed features such as the texture, density, and shape of the liver. The idea is not to make a hospital buy a new, rare instrument, but to retrain the familiar flow of examinations so that it flags who should not be sent home without further checks.

What MAOSS can do

According to the Nature Communications paper, the model was trained and tested on several datasets, from an internal set of 2071 cases to a real-world pool of 18,504 examinations. A separate cohort of 1192 patients was analyzed, where the task was not simply to find steatosis but to select people at elevated risk of more severe disease: steatohepatitis and significant fibrosis. This design matters: it shows the system was tested not only in laboratory conditions but also closer to the ordinary clinical flow.

The key results are as follows:

• the share of high-risk patients identified rose from 16.6% to 52.4%
• AUC for different stages of steatosis was 0.904–0.917
• the average AUC of radiologists without the model's help was 0.709
• when MAOSS was used as an assistant, physicians' accuracy rose to 0.798

The most interesting point lies elsewhere: the model looks for a signal where the patient may have come in for a different reason entirely. If a person has already had a standard CT, the system can use that scan for additional opportunistic screening, without requiring a separate expensive procedure. For health systems this is a strong argument, because the cost of adoption often depends not only on the algorithm but also on the need to change the patient pathway. Here Alibaba is trying to fit into existing infrastructure.

Not instead of the doctor

As Alibaba presents it, this is not a "digital hepatologist" but a decision-support tool. MAOSS does not replace diagnosis, biopsy, clinical assessment, or the patient's follow-up care; it helps spot earlier those whom the standard pathway may miss. The approach looks pragmatic: the model does not promise autonomous medicine, but reduces the share of missed cases and makes early selection cheaper.

This fits well with DAMO Academy's broader strategy. It already promotes AI cancer screening and reports more than 50 million people covered by medical AI checks in ten countries and regions.

But the result should not be overstated either. This is a retrospective validation and a research publication, not evidence that the model has already become a universal standard for hospitals worldwide. Any such tool requires local validation, tuning to a specific patient flow, and clear physician accountability. Otherwise even a strong metric on paper can run into false positives, specialist overload, or problems interpreting results in real practice.

What it means

The MAOSS story shows where medical AI is moving fastest: not toward conversational assistants, but toward quiet systems that extract additional signal from data that has already been collected. If this approach scales, a routine CT could gradually become not just an image for the current complaint but also an early filter for chronic diseases that are still noticed too late today.