The FDA has approved over 1,000 AI systems for radiology. Nearly all can do one thing precisely: identify one organ, one type of finding, one disease. Stanford's foundation model Merlin demonstrates a different approach: it analyzes complete 3D CT volumes, handles 752 diagnostic tasks simultaneously, and outperformed specialized single-system AI in external validation on 44,000 scans from four hospitals. Results appeared in Nature in March 2026.
What makes Merlin different from existing radiology AI
Merlin is a 3D vision-language foundation model: an AI system that processes CT image data, diagnosis codes from electronic patient records, and text from radiology reports simultaneously. It was developed by Louis Blankemeier and Akshay Chaudhari at Stanford University's Department of Radiology, supported by the U.S. National Institutes of Health (NIH).
The decisive technical distinction from previous radiology AI: Merlin analyzes not individual 2D CT slices but the complete 3D volume dataset at once. A typical abdominal CT contains over 300 slices. Existing systems worked slice-by-slice, losing spatial relationships critical to diagnosis. Merlin processes the full volume, recognizing structures and patterns invisible in 2D cross-sections.
The model was trained on Stanford University School of Medicine data: over 6 million CT images from 15,331 scans, combined with more than 1.8 million diagnosis codes from patient records and over 6 million text tokens from actual radiology reports.
What 752 tests revealed
Researchers evaluated Merlin on 752 individual diagnostic tasks across six categories: classifying 31 abdominal findings without specific training (zero-shot), predicting 692 medical phenotypes based on ICD codes, automatic radiology report generation, 3D segmentation of 20 organs, and predicting five-year risk for six chronic diseases including diabetes, hypertension, and cardiovascular disease.
For phenotype classification, Merlin achieved 81% accuracy on average. On a subset of 102 common diagnosis codes, this rose to 90%. For five-year risk prediction, the model reached an AUROC of 0.757 compared to 0.68 for a baseline: a gain of more than seven percentage points. On internal zero-shot classification, Merlin scored F1 0.741, surpassing specialized systems like OpenCLIP and BiomedCLIP. Notably, the model worked on chest CT scans despite training exclusively on abdominal scans.
Context: AI in radiology
Merlin enters a field that has grown rapidly. The FDA approved over 1,000 AI-driven radiology medical devices by end of 2025. GE HealthCare alone holds 115 of these approvals, followed by Siemens Healthineers and Philips. The pattern across all: each system specializes in a single task or anatomical region.
A concrete example: Aidoc received FDA clearance in January 2026 for an abdominal CT tool detecting 14 findings simultaneously, including liver lesions, splenic injury, and appendicitis. This was considered a breakthrough because earlier approved tools covered only one finding category each. By comparison, Merlin was tested on 752 tasks and 692 phenotypes, many times what Aidoc covers.
A second benchmark is Comp2Comp, an open-source project from the same Stanford group, holding multiple FDA clearances for specific CT measurement tasks. Comp2Comp focuses on narrowly defined measurement objectives. Merlin is a generalist foundation model applicable to new tasks without retraining. This is a fundamentally different approach.
Three conditions for clinical deployment
A foundation model like Merlin is not a medical device ready for immediate hospital use. Three concrete requirements must be met.
First: prospective clinical trials. All published Merlin data come from retrospective analyses of historical datasets. FDA or EMA approval requires new studies validating the model under real clinical conditions with current patients.
Second: regulation. The FDA has developed a dedicated approval pathway for AI medical devices. The path from initial clinical data to clearance typically takes two to five years. For a system of Merlin's complexity, covering 752 tasks, the timeline will likely extend toward the upper end.
Third: integration into existing radiology workflows. Scanners, imaging management systems, and reporting software span different generations and protocols. A standardized interface for a foundation model of this complexity does not yet exist in clinical practice.
The researchers released code, model, and a dataset of 25,494 CT scan-radiology report pairs as open source, accelerating external validation by other research groups. Whether this shortens the clinical pathway depends on how many groups undertake prospective studies in coming years.