Artificial intelligence (AI) captures uncertainty in medical scans

Editor's Note

Artificial intelligence (AI) is a useful tool for helping clinicians to determine health problems from medical imaging, but AI often provides just one answer, when there may be a number of possible interpretations. Now, researchers from MIT, the Broad Institute of MIT and Harvard, and Massachusetts General Hospital have introduced a new AI tool that can capture the uncertainty in a medical image. On testing the system with datasets of annotated images, the researchers found that Tyche made better, faster predictions than existing models. It also outperformed more complex models. 

Dubbed Tyche (named for the Greek divinity of chance), the system is detailed in a preprint published January 24 on arXiv. Rather than giving just one solution, Tyche offers multiple plausible segmentations, each of which highlights a slightly different area of a medical image. Users provide the system with examples, such as images of lesions in a heart MRI, to feed predictions. Just 16 images is reportedly enough for the system to understand ambiguity (a “context set”) and make quality predictions, the researchers found. 

Users can specify how many segmentation options they want to receive and review them to select the best segmentation for a particular patient. 

The system is also able to handle additional segmentation tasks without the need for more training, allowing it to be easily adaptable, from finding lesions in a lung X-ray to identifying anomalies in a brain MRI, researchers write.