RADAR Abdominal CT Demo

RADAR Abdominal CT Demo

Reads an abdominal CT volume and returns the probability of each of 146 findings across 18 structures, using the RADAR model. A research and teaching demo, not for clinical decisions.

Made by junma on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This Space is a demo of RADAR, a vision-language model for abdominal CT. You upload a CT volume as a .nii.gz file, and the page returns a probability for each of 146 findings, grouped over 18 anatomical structures. The source describes a model trained on more than 400,000 contrast-enhanced examinations and 15 million anatomy-aware image and text pairs.

This is a third-party demo that wraps the paper authors' weights for research and teaching. It is not a medical device, and the source says not to use it for clinical decisions.

Who it suits: It suits researchers and students who want to see what a generalist CT model reports for a scan.

Industry | Who uses it | How they use it | Example
Medical imaging research | Researchers | Try RADAR on an abdominal CT volume and read the per-finding probabilities | A researcher uploads a volume from a research dataset
Machine learning | ML engineers | Test the released weights through a ready demo | An engineer checks how the model handles a sample volume
Education | Students | See how a vision-language model describes an abdominal scan | A student reads which findings score high on a sample case
Software | Developers | Study the sliding-window inference code in the repository | A developer reads inference_demo.py

You can skip it if you need a tool for patient care or for commercial work, since the weights are non-commercial.

Where to use it: You use it in these places.

• A web browser, on the Space's page on Hugging Face, which runs on a GPU
• A local copy on Windows, macOS or Linux, which needs Python 3.10 and a GPU

Use case | Who | What they get back
Score a CT volume | Researchers | A probability for each of 146 findings across 18 structures
Check coverage | Students | N/A for findings whose organ is not found or is cut off at the volume edge
Study the demo | Developers | The app.py and inference_demo.py files to read

How to set it up: Online use needs only a browser. A local copy needs Python 3.10, a GPU and room for the weights.

• A web browser, for the online page
• Python 3.10 with pip, for a local copy
• A GPU, since the source says it needs a GPU Space and recommends an A10G small
• Room for about 1.5 GB of checkpoint and 0.4 GB of BERT tokenizer files, downloaded on first launch
• Python packages from requirements.txt

1. Open the Space's page on Hugging Face to use it online, with nothing to install.
2. For a local copy, install the Hugging Face Hub package.
3. Download the Space's files into a folder and move into it.
4. Install Gradio 6.17.3 and the packages in requirements.txt.
5. Start the app.

pip install -U huggingface_hub
hf download junma/RADAR-demo --repo-type space --local-dir RADAR-demo
cd RADAR-demo
pip install gradio==6.17.3
pip install -r requirements.txt
python app.py

How to use it: Open the page and upload a CT volume as a .nii.gz file. The page runs the model and returns a probability for each finding. Findings whose organ is not found in the scan, or is cropped at the edge, show as N/A.

python app.py

Example: You upload an abdominal CT volume in .nii.gz format. The page returns a probability for each finding, and shows N/A for any finding whose organ is missing from the scan or cut off at its boundary.

For newcomers and experts: A CT scan is a stack of images of the body, and this model looks at a whole abdominal stack and estimates how likely each listed finding is. Start by uploading one sample volume. Treat the numbers as a research output, not a diagnosis.

Under the hood, the app uses the checkpoint named checkpoint_radar_pretrain.pth, described as RAD-CT pre-trained, with sliding-window inference over a 96 by 256 by 384 region. Weights come from radar-generalist/RADAR on first launch. The record also names three BERT models for text and a dataset called FLARE-MedFM/PancancerCTSeg.

You upload a .nii.gz volume
Sliding-window inference runs on a GPU
The model scores each of 146 findings
Undetected organs show as N/A

Good to know: Medical use is out of scope, and the weights carry a non-commercial license.

• License: cc-by-nc-sa-4, which the source describes as non-commercial and share-alike.
• Cost to run: none named. The source recommends an A10G small GPU Space.
• Data it reaches: the CT volume you upload, and the weights it downloads on first launch from radar-generalist/RADAR.
• Not for: clinical decisions, since the source says this is not a medical device.

On Hugging Face, 13 November 2026

• Likes: 4
• Built with: gradio
• License: cc-by-nc-sa-4.0

Links

• Space: https://huggingface.co/spaces/junma/RADAR-demo
• Files: https://huggingface.co/spaces/junma/RADAR-demo/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.