新的人工智能方法捕捉医学图像中的不确定性

MIT, the Broad Institute of MIT and Harvard, and Massachusetts General Hospital have introduced a new AI tool called Tyche, which addresses the issue of uncertainty in medical image segmentation. Unlike traditional AI models that provide only one answer, Tyche offers multiple plausible segmentations for a single medical image, capturing the diversity of human annotators. Notably, the system can be applied to new segmentation tasks without requiring retraining, making it more accessible to clinicians and biomedical researchers.

Tyche is designed to overcome the limitations of existing neural network-based AI systems by capturing uncertainty and generating multiple predictions without the need for complex models or extensive retraining. The researchers modified the neural network architecture to output diverse candidate segmentations based on a given medical image input and a small set of example images. By ensuring that the candidate segmentations are distinct yet effective, Tyche provides better and faster predictions compared to baseline models, and it has the potential to outperform complex models trained on specialized datasets.

The ability of Tyche to highlight uncertainty in medical images could significantly impact decision-making in clinical settings and biomedical research, ultimately improving diagnoses and calling attention to crucial information that other AI tools might miss. The researchers plan to further enhance Tyche's capabilities and explore methods to improve its worst predictions. This research is supported by funding from the National Institutes of Health, the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard, and Quanta Computer.