Segmentation in biomedicine involves annotating pixels from medical images to highlight structures such as organs or cells. Current artificial intelligence models often provide only one answer, while the process is inherently uncertain. To address this, a team of researchers introduced a new AI tool named Tyche that can provide multiple plausible segmentations for a medical image. Tyche can be used for various tasks, from identifying lesions in lung X-rays to anomalies in brain MRIs, without the need for retraining, which makes it easy for clinicians and researchers to use.
Tyche aims to capture uncertainty in medical images by providing different segmentation options for the user to choose from. This can help in decision-making and highlight potential areas of concern that other AI tools might miss. The system was developed by modifying a neural network architecture to output multiple predictions for a single medical image input and a context set of example images. By adjusting the network’s layers and training process, the model can ensure that the candidate segmentations are different but still accurate.
The research team focused on addressing the limitations of current AI systems for medical image segmentation, which struggle to capture uncertainty and require retraining for new tasks. Tyche offers a solution that is simple and quick to use, without the need for extensive machine-learning experience. By allowing users to see multiple plausible segmentations, Tyche can help in improving diagnoses and biomedical research outcomes.
In testing with annotated medical image datasets, Tyche demonstrated the ability to capture the diversity of human annotators and outperformed baseline models. The system also performed faster than most models without sacrificing accuracy. By providing multiple candidate segmentations, Tyche can give users an edge in medical image analysis and decision-making.
Future work for the researchers includes exploring more flexible context sets and methods to enhance Tyche’s predictions. They also aim to improve the system so it can suggest the best segmentation candidates for a given medical image. This research received funding from various sources, including the National Institutes of Health, the Eric and Wendy Schmidt Center at the Broad Institute, and Quanta Computer.
Overall, Tyche offers a promising tool for addressing ambiguity in medical image segmentation and providing clinicians and researchers with valuable information for diagnostic and research purposes. By enabling the capture of uncertainty and offering multiple segmentation options, Tyche has the potential to enhance medical image analysis and decision-making processes in the field of biomedicine.
