How to Run Alibaba's RADAR Medical AI Model Locally for CT Scans
Alibaba DAMO's open-weight RADAR model flags 146 conditions from abdominal CT scans. Here's how it works and how to install it on your own GPU.
What is RADAR and why does it matter?
RADAR is an open-weight AI model from Alibaba’s DAMO Academy that reads abdominal CT scans and screens for nearly 150 conditions, including liver, pancreas, stomach, and colorectal cancers. It was trained on more than 400,000 CT exams and 15 million image-report pairs without manual annotation, and a study published in Science tested it against 40,000 real-world exams, reporting an AUC of 0.913. In a reader study, RADAR outperformed 23 of 26 radiologists while lifting their sensitivity by about 10%. Because the weights are public, anyone with a capable GPU can run it, not just institutions with API budgets.
That last point is the real story here. Cancer screening tools have historically lived behind hospital contracts, licensing fees, and per-scan API costs. A model that a district hospital or rural clinic can download and run on local hardware changes who gets access to a “second opinion” style diagnostic tool, not just how fast it runs.
TL;DR
- RADAR is Alibaba DAMO Academy’s open-weight model for screening abdominal CT scans across roughly 146 conditions spanning organs like the liver, pancreas, stomach, and colon.
- The model was trained on 400,000+ CT exams and 15 million image-report pairs, and Alibaba avoided manual annotation by aligning scans directly with existing radiology reports.
- A peer-reviewed Science study across 40,000 exams reported an AUC of 0.913, and RADAR beat 23 of 26 radiologists in a reader study while improving their sensitivity by around 10%.
- Because it’s open weight, hospitals and clinics without budget for per-scan diagnostic APIs can self-host it on a single decent GPU.
- Inference relies on organ segmentation templates that tell the model where the liver, pancreas, stomach, and other organs sit in a given scan before it starts flagging conditions.
- Output comes as a CSV file with one row per case and 146 columns, each representing a confidence score for a specific organ-condition pair.
- The model is genuinely lightweight to run: memory use starts under 1GB and settles around 16GB of VRAM during inference, which fits on commodity hardware, and the script exposes parameters like KV cache and context length to reduce that further.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
How does RADAR actually work?
RADAR is built around a vision-language alignment process. During training, thousands of CT scans were paired with the radiology reports doctors had already written for them, and the model learned to connect visual patterns in the scan with the clinical language used to describe them. This is why DAMO didn’t need manual annotation at scale: the reports themselves acted as labels.
Once trained, RADAR takes two inputs at inference time: the CT scan itself and an organ segmentation mask, essentially a pre-made outline showing where organs like the liver, pancreas, and stomach sit within that particular scan. The mask tells the model where to look before it starts scoring for disease, which narrows the problem from “scan an entire 3D volume for anything unusual” to “check this organ region for these specific known conditions.”
The output is a confidence score for each of the 146 organ-condition pairs, e.g. aorta calcification or aortic sclerosis, packaged into a single CSV row per case. Higher scores indicate stronger evidence that the condition is present in the scan. In effect, RADAR functions as a triage layer: it doesn’t replace a radiologist’s read, but it flags which regions and conditions deserve a closer look, functioning as what DAMO describes as a second pair of eyes.
What hardware do you need to run RADAR locally?
One of the more notable practical details is how modest RADAR’s hardware footprint is. During a local test run, the model initialized using under 1GB of VRAM, then climbed to roughly 16GB once fully loaded and processing a scan. That’s well within range of a single consumer or prosumer GPU, not a multi-GPU server rack. The inference script also exposes tunable parameters like KV cache size and context length, which can shrink memory usage even further if needed.
This matters because the target users DAMO and outside testers have in mind aren’t research hospitals with dedicated AI infrastructure. They’re smaller clinics, rural facilities, and district hospitals that may have one GPU-equipped workstation at most. A model that needs a data-center-grade setup to run isn’t realistic for that audience. RADAR’s relatively light footprint is what makes the “run it on a commodity GPU” pitch plausible rather than aspirational.
How do you install and run RADAR yourself?
The basic setup follows a familiar open-source ML workflow. The model is distributed through a GitHub repository (a fairly large one) alongside a Hugging Face model card containing the weights, datasets, and supporting files. The general steps are:
- Clone the GitHub repository and install the dependencies listed at the root of the repo.
- Run the provided download script to pull both the pretrained model checkpoint and the accompanying data, which includes pre-made organ segmentation templates for a sample CT case.
- Navigate into the inference directory and run the provided inference script, which loads the checkpoint along with a BERT-based text model used for the vision-language alignment.
- Feed in a real abdominal CT scan along with its organ mask. The script processes the scan and outputs a CSV file scoring all 146 conditions.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
One friction point worth flagging: downloading from a free Hugging Face account can hit rate limits partway through. The workaround is simply to rerun the download script, which resumes and completes the transfer without needing a paid account. Total setup time is mostly bottlenecked by download speed given the size of the model and dataset, not by any complex configuration.
Is RADAR worth running for non-specialists?
If you’re not a radiologist or medical professional, the raw output (a spreadsheet of confidence scores per organ-condition pair) isn’t something you can meaningfully interpret. A score of 0.70 for aortic calcification or 0.58 for sclerosis means something specific to a trained clinician reviewing a real patient case, and reading too much into these numbers without that background risks misunderstanding what the model is actually claiming.
Where RADAR is genuinely valuable is exactly the use case DAMO designed it for: supporting radiologists and physicians who are already qualified to interpret CT findings, particularly in settings where a second opinion or an initial triage pass isn’t otherwise available. That includes remote clinics in lower-resource regions, but also underserved areas within wealthier countries. Remote communities in places like outback Australia, for example, often face long travel distances just to access CT imaging and specialist review, so a locally-run screening layer can meaningfully shorten that gap even before a radiologist gets involved.
For developers and technical teams, RADAR is also interesting as a case study: it’s a concrete example of a state-of-the-art, peer-reviewed medical imaging model released with an open and comparatively liberal license, full training data references, and a working inference pipeline, rather than being locked behind a paid API.
Frequently Asked Questions
What conditions can RADAR detect?
RADAR screens for close to 146 conditions across abdominal organs, including cancers of the liver, pancreas, stomach, and colorectum, along with other structural and pathological findings like vascular calcification.
How accurate is RADAR compared to human radiologists?
In a reader study cited in the Science publication, RADAR outperformed 23 of 26 participating radiologists and increased their diagnostic sensitivity by around 10% when used as a support tool. Across 40,000 real-world exams, it achieved an AUC of 0.913.
Can RADAR replace a radiologist?
No. It’s designed to function as a supporting tool, flagging likely areas of concern for a qualified physician or radiologist to review, not as a standalone diagnostic authority.
What GPU do I need to run RADAR locally?
Based on hands-on testing, RADAR runs with VRAM usage starting under 1GB and rising to around 16GB during inference, which fits on a single commodity GPU rather than requiring specialized data-center hardware.
Where can I get the RADAR model and code?
The model, its training data references, and inference scripts are published on GitHub and Hugging Face by Alibaba’s DAMO Academy under an open license.