NIH Funded AI Merlin Predicts Future Chronic Disease Risk From 3D Medical Scans

A Powerful New Tool for Radiology

Researchers supported by the National Institutes of Health have introduced Merlin, an artificial intelligence system capable of analyzing 3D abdominal CT scans with high precision. This foundation model does more than identify anatomical structures; it can predict chronic disease onset and interpret data from body regions it was never trained to see. The research, published in Nature, indicates that this technology could soon accelerate clinical decision making by providing faster and more detailed insights.

Training on Massive Data

Developed by a team at Stanford University, the model was trained on the largest collection of abdominal CT data to date. This dataset included over 15,000 3D scans paired with radiology reports and nearly one million diagnostic codes. By utilizing a vision-language training method, Merlin learned to understand the nuanced relationships between visual imaging patterns and the clinical descriptions written by doctors during evaluations.

Superior Diagnostic Accuracy

To ensure the model functions reliably in real world settings, it was tested against 50,000 previously unseen scans from four different hospitals. Merlin correctly matched scans to diagnostic labels 81% of the time, and its performance jumped to 90% for a core group of over 100 common conditions. Its ability to process 3D volumetric data allows it to grasp organ relationships and pathology better than many standard 2D analysis tools used in modern clinics.

Predicting Future Health Outcomes

A significant feature of the system is its prognostic capability, which allows it to function as an early warning system for patients. In comparative tests, Merlin identified individuals at high risk for developing diabetes or heart disease within the next five years with 75% accuracy. This suggests the AI can detect subtle biomarkers of tissue change that are often missed by human eyes during standard reviews or preliminary screenings.

Versatility Across Body Systems

The researchers also tested the flexibility of the model by feeding it chest scans, even though its training was strictly focused on the abdomen. Merlin performed as well as or better than AI tools built specifically for chest imaging. This demonstrates that the system has developed a generalized understanding of human anatomy and disease markers rather than simply memorizing specific abdominal features from its training set.

Supporting Clinical Teams

As the United States faces a growing shortage of radiologists, tools like Merlin offer a way to streamline cumbersome clinical workflows. The researchers suggest the model could eventually automate the drafting of medical reports and triage high priority cases for immediate physician review. The team plans to refine the model further while encouraging other health systems to customize the tool with their own local patient data for specialized use.