Segment organs and read findings in abdominal CT

What this involves

RADAR is a vision-language model developed by Alibaba DAMO Academy, trained on approximately 400,000 contrast-enhanced abdominal CT scans paired with anatomy-level report sentences. When supplied with a suitable axial scan in Hounsfield units, it produces two distinct research outputs: a 36-structure voxel segmentation yielding organ volumes in millilitres, and structure-specific finding scores covering both RADAR's 145-finding exploratory vocabulary and 20 findings evaluated against Stanford's Merlin benchmark.

The model requires contrast-enhanced abdominal CT acquisitions where the whole abdomen is included within the field of view. The underlying platform agent converts DICOM series using dcm2niix or ingests NIfTI volumes directly, launches the segmentation and inference job on a dedicated GPU environment, and returns tabular data along with an interactive viewer displaying structure labels overlaid on the original scan.

Proper interpretation requires contextualising finding scores against reference benchmarks rather than reading them in isolation. Scores represent a softmax output between contrasting text prompts rather than absolute diagnostic probabilities. For the 20 benchmarked findings, results must be evaluated using their external area under the curve (AUC) and their percentile ranking across the 5,124 Merlin test scans.

How it goes

1. 1Ingestion and format conversionProvide a NIfTI volume, a URL, or an uploaded DICOM series or archive, which is converted to NIfTI format using dcm2niix prior to analysis.
2. 2Model executionThe agent initiates RADAR using analyze_ct on a dedicated GPU machine and monitors progress via check_job until the scan is processed.
3. 3Inspection of segmentations and volumesInspect visual overlays of the 36 segmented structures and extract calculated organ volumes resampled at 1 × 1 × 5 mm using ct_results.
4. 4Benchmark contextualisation and tabular exportEvaluate the 20 validated findings against Merlin test set percentiles and published AUCs, saving full results to result.json and findings.csv for cohort analysis.

Start it here

Ask in plain language. The agent works on a Linux machine of its own — it installs and runs what the job needs, keeps every file, and reads the results back with the numbers explained. Nothing to set up, and a long run continues after you close the page.

What to watch for

• RADAR scores are softmax values between presence and absence prompts rather than calibrated probabilities, meaning raw scores cannot be compared across different findings.
• Non-contrast scans fall outside RADAR's training distribution, and non-abdominal scans, MRIs, or 0–255 rescaled images will be rejected immediately.
• Partial inclusions, such as the upper abdomen in chest CT scans, lead to cut-off structures whose calculated volumes and finding scores are unreliable.
• A high percentile score on findings with lower validation performance, such as fractures (AUC 0.68) or atelectasis (AUC 0.71), provides weak evidence.
• The 145 exploratory findings lack external validation on the Merlin test set and must only be treated as hypotheses to inspect, never as confirmed results.
• RADAR is not a medical device and is strictly intended for research cohorts, public datasets, method development, and teaching rather than clinical diagnosis.

What it runs

Questions