Merlin analyses 3D abdominal CT scans for research. Choose report generation, image–text comparison, predicted phenotype findings, or five-year disease-risk estimates.
Available tasks (4)
Task options define what the model does. Choose one when you submit a scan or configure generation.
Radiology report (text)
Generate an organ-by-organ text report from a 3D abdominal CT scan for exploratory research. This task uses the RadLLaMA report decoder.
Input
3D abdominal CT scan. No text input needed.
Output
Generated research report, organised by organ system (.txt and JSON).
Image–text similarity (512-d)
Compare an abdominal CT scan with a description you provide. The similarity score measures relative image–text alignment; it is not a calibrated probability.
Input
3D abdominal CT scan and a text query.
Output
Image–text similarity score and 512-dimensional vector (.npy and JSON).
Phenotype findings (text)
Default
Estimate probabilities across 1,692 EHR phecodes and return the top-ranked predicted findings for exploratory research.
Input
3D abdominal CT scan. No text input needed.
Output
Ranked predicted phenotype findings (.txt and JSON).
5-year disease risk (6 conditions)
Estimate five-year incidence risk for cardiovascular disease, ischemic heart disease, hypertension, diabetes mellitus, chronic kidney disease, and osteoporosis. Research use only.
Input
3D abdominal CT scan. No text input needed.
Output
Estimated five-year risks for six conditions (.txt and JSON).
Research details
Training data
Merlin was trained on a large Stanford cohort of paired abdominal/chest CT — 6+ million CT images across 15,331 studies, 1.8+ million EHR phecode labels, and 6+ million radiology-report tokens. Validated on an internal test set (5,137 CTs) and externally on 44,098 CTs from three external sites plus public datasets. Adult human anatomy.
Intended use
Exploratory research using abdominal CT: generated reports, image–text comparison, predicted phenotype findings, and five-year disease-risk estimates. Not for clinical triage, primary diagnosis, or treatment decisions.
Known limitations
CT only (chest/abdomen distribution). Embeddings and phenotype probabilities reflect the training population and can be biased or wrong on out-of-distribution anatomy, contrast phases, paediatric scans, severe pathology, or non-standard fields of view. Image–text similarity is a relative score, not a calibrated probability. The model resamples every scan to 1.5×1.5×3 mm at 224×224×160, so very large or off-centre fields of view may be cropped.