Researchers from the Massachusetts Institute of Technology (MIT), the Broad Institute of MIT and Harvard, and Massachusetts General Hospital have introduced a new artificial intelligence (AI) tool to capture the uncertainty in a medical image.
Funded by the National Institute of Health, the Eric and Wendy Schmidt Center and Quanta Computer, the Tyche machine-learning model could help clinicians and researchers capture crucial information.
In biomedicine, AI models help clinicians by highlighting pixels that show signs of a certain disease or anomaly. However, these types of models usually only provide one answer.
“Having options can help in decision-making” and “so it is important to take this uncertainty into account,” said MIT computer science PhD candidate, Marianne Rakic.
Researchers developed Tyche after modifying a straightforward neural network architecture. After feeding the tool a few examples of segmentation tasks, such as images of lesions in a heart MRI segmented by different human experts, the model learned the tasks and found that 16 example images were enough for the model to make good predictions without retraining.
The team modified the network to output several predictions based on one medical image input and context set, adjusting the network’s layer so candidate segmentations produced could interact with each other and the examples in the context set.
Furthermore, researchers modified the training process to maximise the quality of its best prediction, allowing Tyche to ensure that candidate segmentations are slightly different while still solving the task.
The team also saw that the tool was able to outperform more complex models trained using a large, specialised dataset and performed faster compared to most models.
Researchers believe that Tyche could benefit clinicians and biomedical researchers more than other methods due to its lack of need for retraining, speed and ability to be applied to a variety of tasks.
The team plans to use a more flexible context set on Tyche and aim to explore methods to improve its worst predictions and enhance the system to recommend the best segmentation candidates.
