Skip to content Skip to footer

In biomedical science, the process of annotating pixels from crucial elements within a medical image, such as a cell or organ, is called segmentation. This task can be aided by artificial intelligence (AI), which highlights pixels that might indicate the existence of a certain disease or anomaly. However, segmentation is seldom clear-cut, as a group of experts may produce differing results when attempting to outline and define the structures within an image. Considering these disagreements, researchers are working on new tools to illustrate the uncertainty inherent to medical image segmentation.

A team of researchers from MIT, the Broad Institute of MIT and Harvard, and Massachusetts General Hospital have developed a novel AI tool named Tyche, capable of capturing image uncertainty. Tyche — named after the Greek goddess of fortune — can offer multiple valid segmentations, each focusing on slightly differing areas of an image. The user determines how many options Tyche offers, choosing the most accurate selection for their intentions.

Notably, Tyche does not require retraining, unlike other AI tools in the field. This makes it accessible to clinicians and biomedical researchers, who could apply the tool to a range of tasks, such as detecting lesions in a lung X-ray or highlighting irregularities in a brain MRI. The tool is capable of alerting its users to critical information that might not be identified by other models.

Adrian Dalca, assistant professor at Harvard Medical School and part of the research team, said that if the tool fails to identify a nodule that three experts agree is present while two disagree, this should be scrutinized. By considering these uncertainties, the tool can provide a holistic and considered perspective that would be useful in medical diagnosis and research.

Tyche was developed by adjusting a simple neural network design. It makes predictions based on a compilation of 16 example images that outline the task, showing the assignment’s ambiguity and therefore predicting possible variations. By following this method, Tyche can tackle new tasks without the need for retraining.

Concluding its testing stage, Tyche has been found to perform faster, and with more diversely accurate predictions than similar models. For further development, researchers aim to make Tyche’s context set more flexible, enhancing the tool to recommend the optimal segmentation candidates. This work was co-funded by the National Institutes of Health, the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard, and Quanta Computer.

Leave a comment

0.0/5