An expert-level generalist AI for abdominal CT diagnosis: RADAR
Synopsis
This work developed RADAR, a generalist vision-language model trained on more than 400,000 contrast-enhanced abdominal CT examinations and 15 million anatomy-wise image-text pairs, learning directly from clinical reports without manual annotation, achieving high diagnostic performance and robust generalization across internal and external evaluations for 18 anatomical structures and 146 imaging findings, and increasing the diagnostic sensitivity of 26 radiologists by ~10% in a reader study.
Interpretation
RADAR is a generalist vision-language model that covers multiple diagnostic tasks in abdominal CT rather than being limited to a single task. Compared with existing supervised strategies that remain limited in scope, the model addresses 18 anatomical structures and 146 imaging findings, reflecting generalist coverage. Trained on more than 400,000 contrast-enhanced abdominal CT examinations and 15 million anatomy-wise image-text pairs, and evaluated internally and externally across multiple centers and varied clinical scenarios.
RADAR learns directly from clinical reports without manual annotation. This learning approach reduces reliance on manual annotation and enables the model to leverage large-scale clinical text data. The paper explicitly states that training data are clinical reports and that no manual annotation was used.
RADAR achieved high diagnostic performance and robust generalization in internal and external evaluations. Evaluation across multiple centers and varied clinical scenarios supports the model's applicability in different settings. Evaluation covers 18 anatomical structures and 146 imaging findings and includes internal and external validation.
In a reader study, RADAR assistance increased the diagnostic sensitivity of 26 radiologists by ~10%. This result suggests that generalist AI can serve as an assistive tool to improve the diagnostic performance of human experts. The reader study included 26 radiologists and reported an approximately 10% increase in sensitivity.
Perspective
This work targets contrast-enhanced abdominal CT diagnosis, covering 18 anatomical structures and 146 imaging findings, and is intended for multi-center and varied clinical scenarios; its scalable, versatile, and interpretable approach provides a basis for extension to broader radiology tasks.
The current text is abstract-level information and does not provide specific performance metrics, reader study design details, the number of external validation centers, or the concrete form of model interpretability; these require the full paper for further understanding.