The aim of this project is to use artificial intelligence (AI) to extract valuable information from unstructured eye movements of highly-skilled domain experts, in particular those of expert clinicians as they perform complex diagnostic decision-making tasks. Such eye-movement data is rich in patterns that can be deciphered using the power of unsupervised machine learning algorithms (such as k-nearest neighbor/hierarchical clustering and principal components analysis) or unsupervised deep learning algorithms (such as deep generative models, autoencoders, and long short-term memory autoencoders for sequence data). Furthermore, as novices transform into experts, patterns embedded in their eye movements (time spent on regions of interest vs. time spent on surgical equipment) may offer a valuable tool for extracting features that pinpoint the critical mechanisms (’eureka moments’) behind expert decision-making. The primary objectives of this project are (1) to collect eye movements of novice and expert ophthalmologists as they view medical images during eye-disease diagnoses using benchtop-based, head-mounted, or Virtual Reality embedded eye trackers (Eyelink 1000, Pupil Labs Core, or HTC Vive Pro, respectively) and (2) to apply unsupervised machine learning/deep learning approaches to extract meaningful information from this data. Features to be extracted from this data include but are not limited to: fixation duration and fixation count in regions of interest, fixation order, saccade velocity, and pupil diameter. This data collection and data analytics project will enable extraction of the most relevant features for task-oriented training of future AI-based disease diagnosis systems. Capturing eye movements, and thereby the underlying visual decision-making mechanisms behind an expert’s knowledge that are not otherwise quantifiable, will allow us to mimic these mechanisms in AI systems, potentially improving their diagnostic accuracy and interpretability for future clinical applications.
Continue reading
