Merlin AI Model Outperforms Single-task Systems in CT Scan Diagnosis

 A groundbreaking study from Stanford University's Department of Radiology presents Merlin, an artificial intelligence (AI) foundation model designed to interpret 3D CT scans and predict a wide array of medical diagnoses. Published in Nature in March 2026, Merlin analyzes full 3D CT volumes rather than individual slices, allowing it to outperform single-task AI systems.

Developed by Louis Blankemeier and Akshay Chaudhari with backing from the U.S. National Institutes of Health (NIH), Merlin can handle 752 diagnostic tasks simultaneously. In an external validation involving 44,000 scans across four hospitals, it achieved an average accuracy of 81% for classifying 692 medical phenotypes. This model demonstrates the potential for foundation models to advance radiology beyond the current single-task focus.

Unlike conventional AI systems, which analyze individual 2D slices, Merlin processes comprehensive 3D volumetric data, maintaining critical spatial relationships necessary for accurate diagnosis. The model's development involved training on a dataset of over 6 million CT images from 15,331 scans, as well as over 1.8 million diagnosis codes and 6 million text tokens from radiology reports.

Merlin's versatility was tested on 752 diagnostic tasks across six categories, including zero-shot classification of 31 abdominal findings and 3D segmentation of 20 organs. Notably, Merlin demonstrated significant capability in predicting five-year risk for chronic conditions such as diabetes and cardiovascular diseases, achieving an area under the receiver operating characteristic (AUROC) of 0.757.

This development comes as the field of AI in radiology continues to expand, with over 1,000 AI systems approved by the FDA for specific imaging tasks by 2025. Companies like GE HealthCare, Siemens Healthineers, and Philips have led in obtaining these approvals, typically focusing on single-task applications. However, Merlin’s broad capacity for multiple diagnostic tasks marks a significant departure from this norm.

Despite its promising performance, Merlin is not yet ready for clinical deployment. Three key steps are required: prospective clinical trials to validate its efficacy with current patient data, regulatory approval through established pathways, and integration into existing radiology workflows. Standardized interfaces for such complex foundation models are not currently in place.

To encourage further research and validation, the team has made the model, code, and a dataset of 25,494 CT scan-radiology report pairs openly available. The pace at which Merlin enters clinical practice will depend largely on the engagement of additional research groups in conducting necessary prospective studies.