Medical imaging is a sophisticated field within the healthcare industry, with the task of interpreting results posing significant challenges. AI models have been developed to support medical professionals by analyzing images, identifying possible signs of diseases. However, a major drawback of these AI systems is that they typically propose a single solution, even when there are multiple possible interpretations for a medical image.
This problem has prompted researchers from Massachusetts Institute of Technology (MIT), the Broad Institute of MIT Harvard, and Massachusetts General Hospital to develop an AI system called Tyche. This system aims to accommodate the inherent ambiguity in medical image segmentation. Marianne Rakic, an MIT computer science PhD candidate and lead author of the study, sees multiple options as an asset in medical decision-making, with awareness of uncertainties in medical images influencing treatment decisions.
The AI system, named after the Greek goddess of luck, generates multiple possible segmentations for a single medical image, addressing ambiguity. Each segmentation highlights varying regions, allowing doctors to opt for the most applicable for their needs.
In terms of function, Tyche operates via four key steps:
1. Users provide Tyche with a “context set” of example images that illustrate desired segmentation tasks. Human-segmented images offer insight into the task and potential variations.
2. The neural network of Tyche was revamped to handle uncertainty. The system’s layers were rearranged so the generated potential segmentations can interact with one another and the context set examples.
3. The Tyche system is designed to produce multiple predictions based on a single medical image and its context set.
4. The system’s training process was adjusted to reward Tyche for producing the optimum prediction. Users then can view all produced medical image segmentations by Tyche.
A key advantage of Tyche is its adaptability. The system can undertake new segmentation tasks without needing complete retraining. This is a notable shift from traditional AI models that necessitate extensive training on large datasets and expertise in machine learning. With Tyche, tasks range from identifying lung lesions in X-rays to spotting brain anomalies in MRIs.
In looking ahead, the research team plans to investigate the use of a more diverse context set, potentially including text and various types of images. They also aim to devise ways to enhance Tyche’s worst predictions, and develop a function that would enable the system to recommend the best segmentation candidates.
