March 17, 2026
At a glance
- Researchers have developed a machine learning model that can interpret abdominal CT scans to make diagnoses and predict risk for some chronic diseases.
- This model has the potential to reduce the time it takes to analyze and interpret CT scans.
Scientists have developed an AI-powered tool that can interpret 3D images from CT scans and diagnose certain abdominal diseases.
Yakovchuk Vyacheslav / Shutterstock
Computed tomography (CT) scans are a common type of 3D medical imaging that can be used to identify and diagnose tumors, infections, and other health problems. However, radiologists must interpret the scans, which can take up to 20 minutes per scan. As the use of CT scans increases, the number of radiologists is not keeping up. Therefore, the radiologist has to interpret more scans and it can be a longer process to get a diagnosis.
Machine learning models that can analyze scan images have the potential to accurately interpret CT scans. However, current models that can interpret medical images are primarily limited to 2D images such as X-rays. An NIH-funded research team led by Stanford University’s Dr. Akshay Chaudhary has developed an AI-powered tool that can interpret 3D images. A description of this tool, called Merlin, and a demonstration of its functionality will be available on March 4, 2026. nature.
Researchers trained Merlin using more than 15,000 abdominal CT scans, radiology reports, and nearly 1 million associated diagnostic codes. We then tested our tool using over 50,000 abdominal CT scans and publicly available datasets from various hospitals. Merlin’s performance was evaluated on six different types of tasks.
For diagnosis, the team checked to see if Merlin could reproduce the radiologist’s results. On average over hundreds of diagnosis codes, Merlin predicted the correct diagnosis with over 81% accuracy. For 102 codes, Merlin’s accuracy was above 90%.
The research team also tested Merlin’s ability to predict whether a healthy patient would develop a chronic disease within five years based on CT scans. Merlin predicted with 75% accuracy which patients would develop the disease over the next five years for six chronic diseases. This suggests that Merlin can detect features in CT scans that cannot be detected by a trained radiologist’s eye.
To test Merlin’s generalizability, the researchers used Merlin to analyze chest CT scans, which were not used to train the model. Still, Merlin performed as well or better than existing models trained specifically for chest scans.
The research team found that Merlin struggled with some of the more complex tasks. When creating radiology reports based on CT scans, Merlin had a tendency to underreport findings. Merlin also had trouble identifying and outlining organs in 3D space.
The results suggest that Merlin can assist in the interpretation of CT scans and reduce the burden on radiologists. The researchers hope to gain approval to use Merlin for simpler tasks in the clinic. We also plan to improve the model to handle more complex tasks, such as creating radiology reports. In the meantime, they have made their model, code, and dataset available to other researchers.
“Our models and data provide the community with a robust backbone on which to build,” Chaudhari says. “The sky is the limit from here.”
Related links
References
Merlin: A visual language-based model and dataset for computed tomography. Blankemeyer L, Kumar A, Cohen JP, Liu J, Liu L, Van Veen D, Gardesi SJS, Yu H, Paschali M, Chen Z, Delbruck JB, Reis E, Holland R, Truitz C, Brusgen C, Wu Y, Lian L, Jensen MEK, Ostmeyer S, Varma M, Varanarasu JMJ, Huang Z, Huo Z, Nabulsi Z, Ardila D, Wen WH, Jr. EA, Ahuja N, Freese J, Shah NH, Zakarchuk G, Willis M, Yara A, Johnston A, Boutin RD, Wendland A, Langlotz CP, Hom J, Gatidis S, Chaudhary AS. nature. March 4, 2026. doi: 10.1038/s41586-026-10181-8. Epub ahead of print. PMID: 41781626.
funding
NIH’s National Institute of Biomedical Imaging and Bioengineering (NIBIB), National Heart, Lung, and Blood Institute (NHLBI), and National Institute of Arthritis and Musculoskeletal and Skin Diseases (NIAMS). Advanced Research Projects Agency for Health (ARPA-H); Medical Imaging and Data Resource Center; ProMedica Foundation.
