A new Merlin AI CT scan model has been built to learn directly from 3D abdominal CT imaging at scale, with the goal of supporting both routine scan interpretation and longer-horizon risk signals. The system is designed as a general-purpose foundation model that links CT volumes with radiology reports and diagnosis codes, rather than relying on extra manual labeling.
In testing, Merlin was evaluated across six task types covering diagnostics, prognostics and quality checks, spanning 752 individual tasks. Researchers also assessed how well outputs held up on large sets of previously unseen scans, including external datasets and reported performance gains in areas such as phenotype-linked predictions and multi-year chronic disease risk comparisons.
Key Capabilities Assessed Included
- Zero-shot findings classification and phenotype classification
- Image-to-text retrieval and radiology report generation
- Five-year disease risk prediction from CT alone
- 3D organ segmentation in abdominal imaging
The team says it plans to release the trained models, code and dataset after removing protected health information, while pursuing regulatory pathways for simpler clinical uses. The Merlin AI CT scan model may also be fine-tuned by health systems using local data.
