In the field of biomedicine, segmentation refers to the process of highlighting important structures in a medical image, from organs to cells. Artificial intelligence (AI) models are starting to play a pivotal role in this task, but there are limitations with most existing models, mainly due to the fact that they are unable to factor in the inherent uncertainty of medical image segmentation. For example, five human experts may each provide different interpretations of a medical image, disagreeing on the borders of a structure in an image. Current AI models traditionally only provide one segmentation interpretation, which may not reflect the range of potential interpretations.
Researchers at MIT, the Broad Institute of MIT and Harvard, and Massachusetts General Hospital have developed a new AI tool called Tyche to address this issue. Tyche, named after the Greek divinity of chance, is able to provide multiple plausible segmentations of a medical image, taking into account the element of uncertainty. The user can specify the number of segmentation interpretations that Tyche outputs and select the most suitable one.
Unlike other models, Tyche does not need to be retrained for each new segmentation task, making it a valuable tool across a variety of medical imaging applications. This is a significant advantage as training AI models usually involves showing the model numerous examples, a process that is both time-consuming and requires extensive machine-learning experience. Therefore, Tyche’s versatile capabilities could enable easier usage for both clinicians and biomedical researchers, improving diagnoses and aiding in biomedical research by potentially uncovering crucial information that other AI models may miss.
To take into account uncertainty, researchers adapted a standard neural network architecture to enable the output of multiple predictions based on single medical image inputs and a set of example images. By implementing changes throughout the network’s layers, Tyche can ensure variation among the candidate segmentations while still fulfilling the task required.
The team discovered that having a “context set” of just 16 example images was sufficient to enable accurate predictions. Additionally, the researchers found that Tyche’s predictions were superior to those from baseline models and could even outperform more complex models that had been trained on larger, specialized datasets. Notably, Tyche also performed faster than most models.
Looking ahead, the team are planning to extend Tyche’s capabilities by working on enhancements such as including text and multiple types of images in the context set and improving the quality of Tyche’s worst predictions. They are also planning to adapt Tyche so it can recommend the best segmentation candidates.
