Chinese Researchers Develop AI to Read Abdominal CT Scans, Reportedly Outperforming Many Physicians

Chinese researchers have developed a general-purpose artificial intelligence (AI) model to read abdominal CT scans and found that its diagnostic performance exceeded that of many radiologists.

A team led by Professor Liang Tingbo of Zhejiang University published a study titled “Expert-Level Generalist AI for Abdominal CT Diagnosis” on the 17th in the international journal Science.

The model, called RADAR, differs from conventional AI diagnostic approaches in that it uses already-written structured radiology reports as training signals rather than requiring lesions to be individually annotated.

The training data consisted of 15 million “anatomical structure–image–text” combinations extracted from more than 400,000 contrast-enhanced abdominal CT examinations.

The model was trained using a contrastive learning approach that brought CT image patches and diagnostic statements from the same patient’s report closer together while placing information from different patients farther apart.

By separately identifying and precisely aligning 18 anatomical structures, including the liver, pancreas, stomach, small intestine and kidneys, the model learned detailed associations that would be difficult to capture using only whole images and reports.

In performance evaluations, RADAR recorded a mean area under the curve (AUC) of 0.913 across 146 imaging findings in 18 anatomical regions in about 40,000 internal cases, a dataset close to real-world clinical practice.

AUCs ranged from 0.874 to 0.912 in data from eight external medical institutions, indicating that performance did not decline substantially across hospitals and scanners.

For liver, pancreatic, gastric and colorectal cancers confirmed by pathology, AUCs ranged from 0.891 to 0.984.

The model also recorded an AUC of 0.904 for diseases associated with acute abdominal pain that had been deliberately excluded from training, demonstrating “zero-shot” generalization to conditions it had not encountered before.

In a reading test involving 26 radiologists from 14 institutions, RADAR alone performed better than most of the physicians.

When physicians received assistance from the AI, sensitivity increased by about 10 percentage points, producing the best results through human-machine collaboration.

RADAR also generated an attention map highlighting the image regions supporting its interpretation, allowing physicians to examine the reasons for the AI’s judgment.

However, the researchers said errors could occur when bleeding was limited, tissue layers were indistinct or reports contained systematic bias.

They added that further training was needed for contrast agent allergies, varied imaging phases, rare diseases and noncontrast CT scans.

The researchers said RADAR was closer to “a highly sensitive second set of eyes” than a replacement for radiologists.