AI System Detects Subtle Changes in Series of Medical Images Over Time

Traditional approaches for analyzing longitudinal image datasets typically require significant customization and extensive pre-processing. For instance, in studies of the brain, researchers often begin with raw brain MRI data, focusing on a specific brain area, correcting for variations in view angles, adjusting for size discrepancies, and eliminating artifacts—before proceeding with the main analysis. Now, a new AI-driven system that can effectively detect changes in medical images over time and predict outcomes offers enhanced sensitivity and adaptability, making it applicable to a wide variety of medical and scientific contexts.

This innovative system, called LILAC (Learning-based Inference of Longitudinal imAge Changes), leverages machine learning techniques and was developed by researchers at Weill Cornell Medicine (New York City, NY, USA). In a study published in the Proceedings of the National Academy of Sciences, the team demonstrated how LILAC could analyze diverse time-series images, or longitudinal data, which included developing IVF embryos, healing tissues after injuries, and aging brains. The researchers found that LILAC can detect even minute differences between images taken over time and predict related outcomes, such as cognitive scores from brain scans. The system is designed to work with much greater flexibility by automatically handling corrections and identifying key changes.

In a proof-of-concept experiment, the researchers trained LILAC on hundreds of image sequences from in-vitro fertilized embryos as they developed. The system was then tested on new sequences to determine which image in each pair was taken first—a task that is difficult to perform without a clear time-related signal in the images. LILAC accurately completed this task with around 99% accuracy, with only a few errors in image pairs with short time intervals. Additionally, LILAC proved highly effective in ordering images of healing tissue from the same sequences and detecting group-level differences in healing rates between untreated tissue and tissue that had undergone an experimental treatment.

Similarly, LILAC was able to predict time intervals between MRI scans of healthy older adults' brains, as well as estimate individual cognitive scores from MRIs of patients with mild cognitive impairment, performing with much less error compared to conventional methods. In all these cases, the researchers showed that LILAC could be easily adapted to emphasize the most relevant image features for detecting changes in individual subjects or distinguishing between groups, potentially offering new clinical and scientific insights. The next phase of research will involve testing LILAC in real-world settings, such as predicting treatment responses from MRI scans of prostate cancer patients.

“This new tool will allow us to detect and quantify clinically relevant changes over time in ways that weren't possible before, and its flexibility means that it can be applied off-the-shelf to virtually any longitudinal imaging dataset,” said study senior author Dr. Mert Sabuncu. “We expect this tool to be useful especially in cases where we lack knowledge about the process being studied, and where there is a lot of variability across individuals.”