AI and Radiologists

AI and Radiologists

（Translated by Doubao)

 

By Du Ping (杜平）

Professor, Communication University of China, Nanjing

 

September 18, 2026

Toutiao (A ByteDance company)

 

I was chatting with a doctor‑friend, and he put forward a rather surprising viewpoint.

 

Getting an X‑ray at the hospital feels straightforward. My doctor‑friend said that while the scanning itself is quick work done by machines, radiologists behind the scenes perform an enormous amount of work interpreting those scans.

 

He also shared a striking observation: hospitals worldwide face a severe shortage of physicians who read medical images. Statistics project that the global shortfall of radiologists will exceed 19,000 by 2036.

 

An abdominal CT scan consists of hundreds of slices. Physicians must go through each one, taking an average of 20 minutes to finish a single report. Around 300 million such scans are performed across the globe every year, with abdominal scans accounting for one‑quarter of that total.

 

How serious is this challenge? Last March, Nature published a research paper specifically exploring how AI could ease the manpower shortage in radiology. Chinese tech media ran a bold headline declaring, “The global shortage of human radiologists is solved.”

 

Yet insiders know it is never that simple. Everyone has considered using AI to fill staffing gaps, yet existing medical‑imaging AI acts as a specialist: each model is trained for one particular condition. An AI designed to detect lung nodules cannot assess liver conditions. Building such an AI also requires two to three years of manual data annotation. In contrast, a radiologist reviews multiple organs end‑to‑end for every report. For AI to become a genuine assistant to doctors, it must possess that same capability.

 

This long‑standing global challenge has now been addressed by a Chinese research team. On September 18, research outcomes from Alibaba DAMO Academy, in collaboration with the First Affiliated Hospital of Zhejiang University and other institutions, were published in Science. This achievement earned publication in this top‑tier international journal largely due to its outstanding performance.

 

Named DAMO RADAR, this general‑purpose medical‑imaging AI model covers 18 organs on abdominal CT scans and identifies 146 diseases. It achieved an AUC score of 0.913 across nearly 40,000 real‑world examinations, hitting expert‑level radiology performance for the first time. In real‑world practice, deploying this medical‑imaging AI not only cuts down working time but also enables junior physicians to deliver performance comparable to senior specialists.

 

Just how acute is the radiologist shortage? Take the First Affiliated Hospital of Zhejiang University as an example. Its radiology department has over 300 staff yet receives 8,000‑9,000 patients daily, with nearly 4,500 CT scans. The department chair, with more than thirty years of experience in medical imaging, admits that mental alertness declines by afternoon. That is precisely where AI can stand guard. It works regardless of time of day, delivering equally thorough analysis at 3 p.m. as at 3 a.m.

 

While Nature was still debating whether AI could plug workforce shortfalls, China has delivered a general‑purpose model attaining expert‑level performance, with both model and code fully open‑sourced. This breakthrough sets the tone for the next phase of medical‑AI development: competition no longer focuses on sheer numbers of tools, but on solving real‑world clinical problems.

 

As my doctor‑friend put it, China’s medical standards have advanced faster than many overseas counterparts in recent years. With the aid of large‑language AI models, we can expect healthcare for Chinese people to grow more accessible, and treatment outcomes to keep improving in the near future.