New AI tool could capture uncertainty in medical images

Credit: Unsplash+.

Researchers at MIT, the Broad Institute of MIT and Harvard, and Massachusetts General Hospital have developed a groundbreaking AI tool named Tyche, which promises to revolutionize the way medical images are interpreted.

This tool, named after the Greek goddess of chance, helps medical professionals by providing multiple interpretations of medical images, thereby capturing the inherent uncertainty in medical diagnosis.

Medical image segmentation is a critical process in diagnosing diseases, where AI helps by identifying and highlighting relevant parts of an image, like organs or tumors.

Traditional AI models, however, tend to provide a single interpretation, which might not always capture the full range of possibilities that multiple expert doctors might see.

Tyche addresses this by offering several possible interpretations of a medical image, allowing doctors to see different potential outcomes and make better-informed decisions.

Marianne Rakic, a PhD candidate at MIT and the lead author of the study introducing Tyche, highlights the importance of acknowledging uncertainty in medical diagnostics, which can significantly influence treatment decisions.

One of the standout features of Tyche is its ability to adapt to new tasks without needing extensive retraining.

This is particularly beneficial in the medical field, where being able to quickly shift between different types of image analyses is crucial.

Tyche can be used right away for various tasks, from spotting lesions in lung X-rays to identifying anomalies in brain MRIs, making it a versatile tool for clinicians and researchers.

The system works by taking a small set of example images—about 16—from a task, such as segmenting heart lesions from an MRI.

These examples, known as a “context set,” show how different experts might interpret the same medical scenario. This helps Tyche understand the task and recognize the range of possible interpretations.

Tyche’s ability to generate multiple outcomes is based on a modification of a standard neural network architecture, which is a type of AI model inspired by the human brain.

This modification allows the AI to produce various plausible segmentations for each image it analyzes, enhancing the richness of the data provided to doctors.

In practice, Tyche can output a specified number of predictions for a single image. For instance, if a doctor wants five different interpretations, Tyche will provide five distinct segmentations, highlighting different aspects and potential concerns in each one.

This method is akin to rolling dice, with the AI offering a range of results, some of which might mirror the diversity of opinions among human experts.

When tested, Tyche not only matched but often exceeded the performance of traditional models, delivering faster and more diverse predictions. It was also capable of improving upon the best predictions of existing baseline models.

The development team believes that Tyche could eventually support even more complex tasks by incorporating different types of data, such as text descriptions or various image formats, into the context set. Future improvements may also enhance the less accurate predictions, ensuring that all possible outcomes are useful.

This innovative tool could significantly impact medical diagnostics and research, providing a more nuanced view of medical images and supporting the detection of subtle, yet critical, health issues.

Funded by several prestigious institutions, including the National Institutes of Health, Tyche represents a significant step forward in the application of AI in medicine.