In the realm of biomedicine, segmentation is a process where certain areas or pixels within a medical image, such as an organ or cell, are annotated or highlighted. This primarily assists clinicians in pinpointing areas showing signs of diseases or abnormalities. However, there is often a gray area since different experts can have differing interpretations on the same medical image. Recognizing the need for a more versatile tool that can accommodate these variances while providing multiple credible results, a team led by Marianne Rakic, an MIT computer science PhD candidate, has developed a new artificial intelligence (AI) tool known as Tyche.
Named after the Greek divinity of chance, Tyche is designed to offer multiple plausible segmentations, each focusing on slightly different regions of a medical image. The user can specify the number of options to be generated and select the most fitting one. Unlike other AI models, Tyche does not need to be retrained, making it more accessible to clinicians and researchers who can apply it directly for various tasks such as identifying lesions in a lung X-ray or pinpointing anomalies in a brain MRI.
A key feature of Tyche is its ability to tackle the underexplored aspect of ambiguity in medical image segmentation. It is designed to capture and represent the uncertainty in medical images, which is often overlooked by other AI tools. By doing so, the new system aims to highlight potentially crucial information that may otherwise be missed.
To build Tyche, the researchers modified a straightforward neural network architecture. Users start by feeding Tyche a small number of specific examples, following which it can make predictions without retraining. The program also has a unique feature of fostering interaction among various candidate segmentations and examples during the data processing journey from layer to layer within the neural network architecture. This enables the production of differing segmentations while ensuring they all still solve the task. An additional modification was brought in the training process, rewarding the network for maximizing the quality of its best prediction.
In the tests conducted, it was found that Tyche could produce predictions that echoed the diversity of human annotators faster than most models. It could also outperform more complex models trained using a large, specialized dataset. For future developments, the team plans to test a more flexible context set, explore improvement mechanisms for its worst predictions and enhance the system to recommend the best segmentation candidates. This research was funded by the National Institutes of Health, the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard, and Quanta Computer.