RADAR: An Expert-Level Generalist AI for Abdominal CT Diagnosis

Artificial intelligence (AI) in radiology aspires to deliver expert-level diagnosis across diverse clinical tasks, yet existing supervised strategies remain limited in scope. We 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. Across extensive internal and external evaluations spanning multiple centers and varied scenarios, RADAR achieved high diagnostic performance and robust generalization for 18 anatomical structures and 146 imaging findings. In a reader study, AI assistance increased the diagnostic sensitivity of 26 radiologists by approximately 10%, underscoring its clinical utility. In summary, RADAR offers a scalable, versatile, and interpretable solution, demonstrating that generalist AI can match human experts in both routine and complex radiology tasks.

Data Preparation

Pre-trained checkpoints and supporting files are available on HuggingFace.

File	Description	Destination
Pretrained checkpoint	RADAR pre-trained on RAD-CT	radar/ckpt/checkpoint_radar_pretrain.pth
ckpt/bert-base-chinese	BERT tokenizer and model (Chinese)	radar/ckpt/bert-base-chinese/
ckpt/bert-base-uncased	BERT tokenizer and model (English)	radar/ckpt/bert-base-uncased/
ckpt/checkpoint_unet.pth	Pretrained VisionBranch (UNet) checkpoint	radar/checkpoint_unet.pth
ckpt/checkpoint_radar_plus.pth	RADAR+ checkpoint trained from scratch on Merlin-CT-Train	radar/ckpt/checkpoint_radar_plus.pth
(Optional) data/merlin_data_train_full/resized_images/	CT images resized to spacing [1, 1, 5] (demo images provided in this repo)	radar/data/merlin_data_train_demo/resized_images
(Optional) data/merlin_data_train_full/resized_masks/	Anatomical masks generated by TotalSegmentator (104 anatomies → 36 main structures; demo masks provided)	radar/data/merlin_data_train_demo/resized_masks