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

RADAR is a generalist vision-language model trained on over 400,000 contrast-enhanced abdominal CT examinations with 15 million anatomy-aware image–text pairs, learning directly from clinical reports without manual annotation. RADAR provides a scalable and versatile framework for radiology AI, demonstrating expert-level performance across both routine and complex clinical tasks.

Model Preparation

Pre-trained checkpoints are available on HuggingFace.

File	Description	Destination
checkpoint_radar_pretrain.pth	RADAR pre-trained on RAD-CT	radar/ckpt/checkpoint_radar_pretrain.pth
bert-base-chinese	BERT tokenizer and model (Chinese)	radar/ckpt/bert-base-chinese/
bert-base-uncased	BERT tokenizer and model (English)	radar/ckpt/bert-base-uncased/
checkpoint_unet.pth	Pretrained VisionBranch (UNet) checkpoint	radar/ckpt/checkpoint_unet.pth
checkpoint_radar_plus.pth	RADAR+ checkpoint trained from scratch on Merlin-CT-Train	radar/ckpt/checkpoint_radar_plus.pth
checkpoint_radar_plus_finetuned_on_merlin.pth	RADAR+ checkpoint pretrained on RAD-CT and finetuned on Merlin-CT-Train	radar/ckpt/checkpoint_radar_plus_finetuned_on_merlin.pth