AbdomenAtlas 3.0 is available - 9,262 patients paired with image + mask + report + metadata

[Zongwei Zhou]

Dear colleagues,

 

We have released a fully open dataset for benchmarking vision-language models, including tasks of radiology report generation, visual question answering (VQA), early tumor detection, and many more. The dataset focuses on cancer and includes voxel-wise annotations of liver, kidney, and pancreatic tumors. In addition, all major organs, veins, and arteries are annotated. This level of annotations allows reasoning (e.g., chain of thought), as models can ground their report generation, VQA, and detection in detailed structures and spatial relationships rather than relying on raw CT images alone.

 

0 | Two lines of code to download the entire dataset

 

git clone https://github.com/MrGiovanni/RadGPT.git

cd RadGPT; bash download_atlas_3.sh

 

Moreover, we would be pleased to explore the following collaboration opportunities, which may be helpful for your ongoing projects, grant applications, manuscript development, and commercialization:

 

1 | Medical AI Bootcamp: JHU+JHMI+UCSF+Harvard+Nvidia

We are currently offering visiting research positions at Johns Hopkins University, either in person or remotely, for undergraduate, Master’s, and PhD students. Participating students will be granted a JHU email account and access to our resources, including AbdomenAtlas and A LOT OF GPUs. They will also work very closely with our team on research projects. See more details here: https://www.cs.jhu.edu/~zongwei/advert/Call4Research.pdf

 

2 | External validation datasets

We can provide access to our curated multicenter datasets for external validation. The dataset has been annotated and collected from 145 hospitals worldwide. See more details here: https://www.zongweiz.com/dataset

 

3 | Expert voxel-wise annotations

Our network includes 50 board-certified radiologists who can provide high-quality annotations of pancreatic cancer and surrounding anatomical structures, including tumors and relevant organs, directly on your raw CT scans.

 

4 | Reader studies and silent trials

We have both the radiologist panel (N = 50) and the datasets (N = 400,000) to support multi-reader studies and silent trial evaluations.

 

If any of these opportunities align with your current or planned work, I would be glad to discuss details further.

 

Thanks and regards,

Zongwei Zhou, PhD

Assistant Research Professor

Computer Science, Johns Hopkins University

P: 1-(480)738-2575 | E: zzh...@jh.edu

https://www.zongweiz.com/

Scholar | GitHub | YouTube | Calendar