Artificial Intelligence (AI) models are increasingly being employed in the field of biomedicine to assist clinicians with image segmentation, a process that annotates pixels from important structures in a medical image, such as an organ or cell. However, these AI models often offer a singular answer, while image segmentation in the medical field is usually shrouded in a degree of ambiguity.
Marianne Rakic, an MIT computer science PhD candidate, asserts the importance of accommodating for this uncertainty during interpretation, given its significant influence on decision making processes. Rakic, in partnership with a team of researchers from MIT, the Broad Institute of MIT and Harvard, as well as Massachusetts General Hospital, introduces Tyche, an AI tool engineered to represent the varying degrees of uncertainty in medical image segmentation.
Tyche, named after the ancient Greek divinity of fortune, provides multiple plausible segmentations, each outlining slightly different areas of a medical image. Users can dictate the number of alternative outputs from the tool and choose the most appropriate version to apply to their respective field of work, whether it’s organ identification or disease detection. Unlike previous models, Tyche doesn’t require retraining for each new task, making it less labor-intensive and more consistently reliable. This feature enables Tyche to function right out of the box, immediately applicable to a wide array of tasks, which ranges from identifying lung lesions in a lung X-ray to locating anomalies in a brain MRI. By bringing potentially vital information to the user’s attention, Tyche aims to augment diagnostic procedures and aid in biomedical research.
Philip Rakic and his coauthors designed Tyche to address the underexplored issue of ambiguity in medical image interpretation, where the same image can lead to varying conclusions among experts. Tyche is aimed to capture these variances and bring attention to the differences in expert interpretations.
The design of Tyche is based on a simple neural network architecture which is adjusted to produce multiple predictions based on a single image input, making it possible to present a range of possible segmentations. As data move from layer to layer in the neural network, the alterations produced at each step can interact with each other, thus ensuring a spectrum of different, yet viable segmentations. Also, the researchers created another version of Tyche that allows a single-use, pretrained model for medical image segmentation to produce multiple candidates by making minimal modifications to images.
Tyche undoubtedly symbolizes a significant progression in the AI tools available for medical imaging applications. The future development and scale-up of Tyche is estimated to bring about a revolution in the interpretation of medical imaging, by offering an AI tool that echoes the complexity and variability inherent to human interpretation, while also augmenting it with the processing power and speed of machine-based models.
