DeepFAN Wins Class III Certification and Improves Lung Cancer Diagnosis

An international team of clinicians and developers has published the evaluation results for DeepFAN in Nature Cancer. DeepFAN is an AI system for analyzing pulmonary nodules on CT scans. The algorithm went through a multicenter MRMC (multiple readers, multiple cases) trial with a paired comparison, and it markedly improved lung cancer diagnosis among junior radiologists. The system is already certified as a Class III medical device, so it has moved beyond the research prototype stage.

A test on those who need the hint most

The researchers chose a difficult but practical audience: radiologists with little experience. These are the doctors who most often read chest CT scans first in district and city hospitals. Their report determines whether a patient goes on to follow-up monitoring or to further examination.

MRMC is considered one of the most rigorous ways to evaluate an AI assistant in radiology. Several doctors assess the same set of cases in two modes: on their own and with the algorithm's suggestions. The paired design removes the main source of noise, the differences between doctors and between patients. The same people are compared on the same images. Involving multiple centers also reduces the risk that the model performs well only on one clinic's data.

Who is behind the work

The project brought together academic, clinical and commercial organizations:

• University of Hong Kong — the research side and the authors of the Nature Cancer paper;
• Peking Union Medical College Hospital (北京协和医院) — one of China's leading clinical centers;
• Peking University People's Hospital;
• Wuhan Third Hospital;
• Huangshi Central Hospital;
• Deepwise (深睿医疗) — a company that develops AI for medical imaging and handled the product side.

This lineup reflects a setup typical of Chinese medicine. Large hospitals in the capital contribute expertise and data, regional hospitals provide realistic workloads, and an industry partner turns the algorithm into a product that can be certified.

What Class III certification means

Under China's regulatory system, Class III is the highest risk category for medical devices. Software that influences diagnosis must undergo clinical evaluation and demonstrate safety and effectiveness. For AI tools this is a serious barrier: many models remain papers and demos, and only a handful reach registration.

DeepFAN combines three elements that are rarely found together:

• a publication in a leading specialty journal, Nature Cancer;
• multicenter validation using the MRMC design;
• registration as a device in the highest risk class.

Where it could help in the clinic

Lung cancer screening with low-dose CT generates huge volumes of images, while the number of experienced radiologists grows more slowly. If AI really does lift the reading quality of junior specialists, it works as an equalizer between central and peripheral hospitals. A novice doctor gets a second opinion at the moment of reading, not after the report is written.

The precedent matters for the market as well. Medical AI is increasingly judged by how it changes a doctor's work, not by an algorithm's metrics on a closed test set. For hospitals, this is the language in which purchasing decisions are made. For developers, it is a template: a clinical design with real readers, multiple sites and regulatory status. DeepFAN's next step is to confirm the results in everyday practice and to integrate into radiologists' workstations, where the suggestions must not slow down reading.