Segmentation in biomedicine involves annotating pixels from important structures in medical images, such as organs or cells. Artificial intelligence models can assist clinicians by highlighting pixels that may indicate signs of a certain disease or anomaly.
However, these models typically provide only one answer, while medical image segmentation problems are often more complex. Different human annotators may offer different segmentations, leading to disagreements on the borders of structures in images.
Introducing a new AI tool called Tyche, developed by researchers from MIT, the Broad Institute of MIT and Harvard, and Massachusetts General Hospital. Tyche can capture uncertainty in medical images by providing multiple plausible segmentations, allowing users to choose the most suitable one for their needs.
Unlike other methods, Tyche does not require retraining for new segmentation tasks, making it easier for clinicians and researchers to use. It can be applied for various tasks, from identifying lesions in X-rays to pinpointing anomalies in MRIs.
By addressing ambiguity and providing multiple segmentation options, Tyche could improve diagnoses and aid in biomedical research by highlighting crucial information that other AI tools might overlook.
For more information, the paper on Tyche will be presented at the IEEE Conference on Computer Vision and Pattern Recognition as a highlight.
The researchers developed Tyche by modifying a basic neural network architecture to output multiple predictions based on a single medical image input and a context set of examples. This allows Tyche to solve new tasks without needing to be retrained.
Testing Tyche with annotated medical image datasets showed that its predictions captured the diversity of human annotators and outperformed baseline models in terms of accuracy and speed.
Future work will focus on using a more flexible context set and improving Tyche’s worst predictions, as well as enhancing the system to recommend the best segmentation candidates.
Funding for this research was provided by the National Institutes of Health, the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard, and Quanta Computer.
