AI Collaboration With Physicians Raises Abdominal CT Reading Sensitivity by About 10%

A study found that artificial intelligence (AI), working with physicians, increased sensitivity in reading abdominal CT scans by about 10%.

According to BioOn on the 29th (local time), researchers developed RADAR, a general-purpose abdominal CT diagnostic AI trained solely on large-scale clinical reports, without manual human annotation.

The model is based on a scalable vision-language framework and was trained by dividing CT images into anatomical units and aligning them with diagnostic text reports.

Researchers built a large dataset, RAD-CT, comprising 424,911 examinations and 15 million anatomical pairs, to train the model.

In an evaluation of 18 anatomical structures and 146 imaging findings, RADAR recorded an area under the receiver operating characteristic curve (AUC) of 0.913 for diagnosing the 146 findings in a real-world internal cohort of 39,160 cases.

At eight external institutions, AUCs ranged from 0.874 to 0.912. In pathology-confirmed assessments of four cancers—liver, pancreatic, gastric and colorectal cancer—AUCs ranged from 0.891 to 0.984.

RADAR maintained an AUC of 0.904 for acute abdominal conditions excluded from initial training, demonstrating generalization to unfamiliar emergency situations.

In a reading study involving 26 radiologists from 14 institutions, RADAR outperformed most participants and increased sensitivity by about 10% when used collaboratively.

RADAR also provides attention maps that highlight diagnostic visual cues to help interpret readings.

The researchers said, “This study shows that large-scale vision-language learning could offer a feasible path to developing general-purpose AI systems for complex and diverse radiology readings.”

The findings were published in the international journal Science.