What Is RADAR? Alibaba's Open Cancer-Screening AI Explained
RADAR is Alibaba DAMO's open medical AI that scans CT images for 146 conditions. Here's how it works and what its clinical results showed.
What is RADAR?
RADAR is an open-weight medical AI model built by 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, and in a clinical study published through Science.org it outperformed the majority of radiologists it was tested against. Because the weights are public, any hospital or clinic with a decent GPU can run it locally, no per-scan API fees required.
TL;DR
- RADAR is Alibaba DAMO Academy’s open-source AI model for screening abdominal CT scans, covering nearly 150 conditions across organs like the liver, pancreas, stomach, and colon.
- The model was trained on over 400,000 CT exams paired with 15 million image-report pairs, and DAMO says it required no manual annotation, since it learned by aligning scans with the radiology reports already written for them.
- In a study across roughly 40,000 real-world exams, RADAR reportedly achieved an AUC of 0.913, a strong score for a diagnostic classifier.
- In a reader study comparing it against human experts, RADAR outperformed 23 of 26 radiologists and boosted overall reader sensitivity by about 10% when used as an assistive tool.
- Because it’s open weight, the model can run on a single commodity GPU, which matters for hospitals and clinics that can’t afford ongoing API costs.
- The model outputs a confidence score for each of 146 conditions per scan, structured as one row per case and one column per organ-condition pair, meant to flag findings for a qualified radiologist to review, not replace one.
- Its biggest practical impact may be in under-resourced settings, rural hospitals or regions without on-staff radiologists, where a free, locally-run second opinion could meaningfully raise the floor on diagnostic quality.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
How does RADAR actually work?
RADAR is built around a vision-language architecture that connects what a CT scan shows to what a radiologist would write about it. During training, the model ingests large volumes of CT scans alongside their corresponding radiology reports. Rather than requiring humans to manually label every scan with the specific disease present, RADAR learns the association between visual patterns and the language radiologists already use to describe them. This is why DAMO describes the training pipeline as not needing manual annotation: the reports themselves act as the labels.
Before RADAR analyzes a scan, it uses pre-made organ outlines, essentially templates showing where organs like the liver, pancreas, and stomach typically sit within an abdominal CT. These masks help the model narrow its attention to the right anatomical regions before it starts scoring for disease, rather than scanning the image blindly.
Once trained, the model takes in a CT scan and its associated organ mask and produces a confidence score for each of the 146 conditions it was trained to detect. The output format is a CSV file, one row per case, with 146 columns, each labeled by organ and condition. Every score represents the model’s confidence that a specific issue, calcification in the aorta or a lesion in the pancreas, for instance, is present in that scan. Higher scores mean stronger evidence, and clinicians use those scores to prioritize which regions of a scan warrant closer review.
What did the clinical study show?
The headline numbers come from a study published via Science.org, run across roughly 40,000 real-world CT exams. RADAR reportedly hit an AUC (area under the curve) of 0.913. AUC is a standard way to measure how well a diagnostic model distinguishes between patients who have a condition and those who don’t, with 1.0 being a perfect score and 0.5 being no better than random guessing. A score above 0.9 generally indicates strong discriminative performance for a screening tool.
The more striking result came from a reader study, where RADAR’s outputs were compared directly against a panel of human radiologists. According to the figures cited in DAMO’s materials, RADAR outperformed 23 of 26 radiologists in the study. Just as notably, when radiologists used RADAR as an assistive second read rather than a replacement, their own diagnostic sensitivity, meaning their ability to correctly catch true cases of disease, improved by about 10%. That combination, a model that beats most individual experts but also makes those experts better when paired with it, is the more clinically meaningful finding than the raw AUC score alone.
Why does it matter that RADAR is open weight?
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
Most AI diagnostic tools in medicine are locked behind commercial APIs, priced per scan, and often out of reach for hospitals in lower-income regions. RADAR’s release changes that math. Because the model weights, training data references, and inference code are published openly on Hugging Face and GitHub, any institution with a capable GPU can run the model without paying per-scan fees or sending patient data to a third-party cloud service.
That has real implications for parts of the world with thin radiology coverage. A rural district hospital in South Asia, a clinic in sub-Saharan Africa, or even remote communities in wealthier countries like Australia’s outback, where patients sometimes travel long distances for imaging review, could run RADAR locally as a triage or second-opinion tool. It doesn’t replace a radiologist’s judgment, but it can flag scans that need urgent attention in settings where no radiologist is immediately available.
DAMO Academy also released the license under fairly permissive terms, though it’s not entirely without restrictions worth checking for institutions considering clinical deployment.
Is RADAR practical to run locally?
Yes, and that’s part of the point. Independent testing of the model’s inference script showed it loading with modest memory requirements, initially under 1GB of VRAM before climbing to around 16GB during full inference, well within range of a single commodity consumer or workstation GPU. The repository includes options to reduce settings like KV cache size and context length to shrink that footprint further if needed.
The setup process involves cloning the model repository, installing dependencies, downloading the pretrained checkpoint plus a companion text encoder, and downloading a dataset of organ segmentation masks used to guide the model’s attention during inference. Downloads come through Hugging Face, and free-tier accounts may hit rate limits when pulling the full dataset, though retrying typically resolves it.
Running inference means feeding RADAR a CT scan and its corresponding organ mask, after which it produces the 146-column confidence score CSV described above. The barrier to entry is low enough that a technically capable IT team at a mid-sized hospital, not just a well-funded research lab, could plausibly stand this up.
What are the limits of a model like RADAR?
RADAR is a screening and flagging tool, not a diagnostic authority. Its output is a set of confidence scores meant to direct a trained radiologist’s attention, not a final medical verdict. The reader study results are encouraging, but they reflect performance in controlled study conditions, and real-world deployment across varied patient populations, scanner hardware, and imaging protocols can behave differently than benchmark datasets suggest.
There’s also the question of oversight. A confidence score of 0.70 for aortic calcification means little without a physician who understands what threshold matters clinically and how to weigh it against the rest of a patient’s history. Open access to the model doesn’t remove the need for qualified medical judgment in the loop, it just makes that judgment more available where staffing is thin.
Frequently Asked Questions
What conditions can RADAR detect?
RADAR screens for nearly 150 conditions across abdominal organs, including cancers of the liver, pancreas, stomach, and colon, plus other findings like vascular calcification, based on abdominal CT scans.
How was RADAR trained?
It was trained on over 400,000 CT exams paired with 15 million image-report pairs, learning to associate visual patterns in scans with the language used in the radiology reports written about them, without requiring manual disease labeling.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
Did RADAR outperform human radiologists?
In a reader study, RADAR reportedly outperformed 23 of 26 radiologists and, when used alongside radiologists as an assistive tool, boosted their diagnostic sensitivity by about 10%.
Can RADAR run on a normal GPU?
Yes. Reported testing showed the model running with VRAM usage in the range of a single commodity GPU, making it accessible to hospitals or clinics without specialized AI infrastructure.
Is RADAR meant to replace radiologists?
No. It’s designed as a screening and second-opinion tool that flags likely findings for review, intended to support radiologists, especially in under-resourced settings, rather than substitute for their diagnosis.