Researchers from MIT, in collaboration with the Broad Institute of MIT and Harvard and Massachusetts General Hospital, have introduced a new artificial intelligence (AI) tool known as Tyche, which can provide multiple, plausible image segmentation possibilities for a given medical image. Unlike conventional AI tools, which typically offer a single definitive interpretation, Tyche generates a variety of segmented images, highlighting slightly different areas that could potentially indicate a disease or anomaly.
The AI, named after the Greek divinity of chance, has showcased in initial testing that it doesn’t need to be retrained for different segmentation tasks. This characteristic makes it a more accessible tool for clinicians and biomedical researchers as it can be applied “out of the box” for multiple tasks ranging from identifying lesions in a lung X-ray to pinpointing anomalies in a brain MRI. Theoretically, Tyche could lead to improved diagnoses as well as advancing biomedical research by drawing attention to critical information other AI tools might overlook.
The model’s ability to present multiple segmentation possibilities encourages a broader view of the image rather than a singular, definitive interpretation. This fosters an increased awareness of ambiguity, a critical aspect that has often been overlooked in traditional AI models. Addressing ambiguity opens up the opportunity for improved diagnostics, particularly in uncertain or borderline cases, and could potentially call attention to clinically significant findings that might else be ignored.
The researchers designed Tyche by modifying a simple neural network model. A fundamental step in the method includes feeding the system a few examples, also known as a ‘context set’, to present the segmentation task in question. The researchers discovered that using just 16 example images was enough for Tyche to make robust predictions, however, there’s no limit to the number of examples one can provide. The model was further modified to ensure it yields various predictions based on the same medical image and context, thus capturing the element of uncertainty inherent in the process.
Preliminary tests using Tyche on annotated medical images demonstrated that it was capable of capturing the diversity of human experts, whose interpretations often vary due to subjective factors. The AI system also outperformed its counterparts in terms of speed and the quality of its best predictions. For future development, the researchers envisage a more flexible model, incorporating multiple types of images and text inputs, and are exploring ways to enhance the system’s ability to recommend the best segmentation candidates. This research has been funded, in part, by the National Institutes of Health, the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard, and Quanta Computer.