TL;DR

NIH-funded researchers have developed Merlin, a general-purpose machine learning model for analysing 3D abdominal CT scans. Trained on more than 15,000 scans from Stanford, the model predicted diagnosis codes correctly over 81% of the time and forecast chronic disease onset five years in advance with 75% accuracy. The results, published in Nature, show a single general-purpose model outperforming specialist models built for individual conditions.

What Happened

A team of NIH-funded researchers has built and validated Merlin, a machine learning model designed to analyse three-dimensional abdominal CT scans. Unlike most medical AI systems, which are trained to detect a single condition or abnormality, Merlin operates as a general-purpose diagnostic tool capable of identifying multiple conditions from a single scan.

The model was trained on more than 15,000 CT scans sourced from Stanford University medical records. In testing, it matched scans to correct diagnosis codes with accuracy exceeding 81%. More notably, it predicted the onset of chronic diseases — conditions not yet diagnosed at the time of scanning — up to five years in advance, achieving 75% accuracy on that task.

The findings were published in Nature, marking one of the first peer-reviewed demonstrations that a general-purpose radiological AI model can outperform specialist models that were purpose-built for narrower diagnostic tasks.

Why General-Purpose Matters

Most medical AI development has followed a specialist approach: one model per condition, each requiring its own training data, validation, and regulatory approval. This creates practical problems. A hospital deploying AI-assisted radiology might need dozens of separate models, each with its own maintenance requirements and failure modes.

Merlin’s results suggest an alternative path. A single model capable of screening for multiple conditions simultaneously could streamline radiology workflows significantly. Radiologists reviewing abdominal CT scans — one of the most common imaging procedures — would receive a comprehensive AI-generated assessment rather than outputs from multiple disconnected tools.

The five-year predictive capability adds another dimension. If validated in broader clinical settings, the ability to flag patients at elevated risk of developing chronic conditions before symptoms appear could shift abdominal CT scanning from a diagnostic procedure to a screening tool.

Clinical Implications

Abdominal CT scans are ordered frequently across hospital settings, generating large volumes of imaging data that radiologists must review under time pressure. An AI assistant that reliably identifies both current conditions and future risks could reduce diagnostic delays and support earlier intervention for chronic diseases.

Looking Forward

The Merlin model will require further validation across diverse patient populations and hospital systems before clinical deployment. Stanford training data may not represent the full range of demographic and clinical variation seen in broader healthcare systems, including the NHS. Regulatory approval pathways for general-purpose diagnostic AI remain less established than those for single-condition models, which could affect the timeline for real-world adoption.