Artificial intelligence (AI) tools have great potential in the field of biomedicine, particularly in the process of segmentation or annotating the pixels of an important structure in a medical image. Segmentation is critical for the identification of possible diseases or anomalies in body organs or cells. However, the challenge lies in the variability of the results, as different human annotators might provide differing interpretations of the same image.
To address this issue, a team from MIT, the Broad Institute of MIT and Harvard, and Massachusetts General Hospital, led by MIT computer science PhD candidate Marianne Rakic, has introduced a new AI model named Tyche. Named after the Greek divinity of chance, Tyche provides multiple plausible segmentations that highlight slightly different areas of a medical image, creating a range of options that a user can select from, thus addressing the issue of uncertainty.
An additional innovation of Tyche is that it can tackle new segmentation tasks without requiring retraining. Retraining AI models is a data-intensive and time-consuming process that requires significant machine learning experience. However, the use of Tyche bypasses this, making it more user-friendly for clinicians and biomedical researchers, and it could potentially be applied indiscriminately across the spectrum of segmentation tasks.
The simplicity and efficiency of the Tyche model may significantly improve diagnoses and aid biomedical research by highlighting potentially crucial information that other AI tools might overlook. Marianne Rakic emphasizes the importance of the system’s capability to encompass ambiguity, which has often been overlooked despite being a valuable operator in decision-making processes.
The architecture of Tyche is a modification of a straightforward neural network, which processes data through a structure of interconnected layers of nodes or neurons, loosely based on the human brain. The researchers addressed the issue of retraining requirements and the failure to capture uncertainty by allowing Tyche to output multiple predictions based on one medical image input and a set of examples provided by the user, and by adjusting the layers of the network to allow communication between the candidate segmentations.
A unique feature of Tyche is its capacity to learn from a small “context set” of 16 example images to make good predictions, with virtually no limit to the number of examples that can be used. This makes it considerably more efficient than most models, allowing it to be applied to a variety of segmentation tasks without requiring retraining.
In addition, Tyche has the capability to be used with an existing, pretrained model for medical image segmentation. In this context, Tyche allows the model to produce multiple candidates by making slight adjustments to the images.
The researchers discovered that Tyche’s predictions accurately represented the diversity of human annotators and outperformed most baseline models in terms of speed and convenience, while also capturing more diverse results than more complex models trained on large, specialized data sets. Future developments of Tyche might include enhancing the system to enable it to recommend the best segmentation candidates and exploring the possibility of including text or multiple types of images in the context set.
