The development of a new machine learning model named Merlin could transform the interpretation of medical scans, potentially expediting the diagnostic process. Funded by the National Institutes of Health (NIH), this model utilizes 3D abdominal computed tomography (CT) scans to identify anatomical features and predict disease onset years in advance, outperforming specialized automated tools in various tasks.
Researchers from the Stanford University School of Medicine trained Merlin using an extensive dataset comprising over 15,000 patient CT scans linked with radiology reports and nearly one million diagnostic codes. This dataset is recognized as the largest collection focused on abdominal CT data to date. The study involved evaluating Merlin across six broad task categories, which included over 750 diagnostic and prognostic activities.
Merlin is classified as a foundation model, designed to learn from large-scale unlabeled datasets that include diverse information types. The introduction of this advanced model could simplify the traditionally complex and lengthy interpretation of CT scans, addressing challenges posed by the physician shortage in the United States, according to co-first author Louis Blankemeier, Ph.D.