Radiologist St. James's Hospital Dublin Dublin, Dublin, Ireland
Purpose: - To illustrate the aspects of an effective visual search technique in radiologic studies and identify the characteristics that distinguish novice radiologists from experts including scan path, visual dwell time, and classification of interpretative errors. - To provide an overview of eye tracking as a data tool and its principles. - To discuss the theory, benefits and limitations of employing AI decision support algorithms based on eye tracking data in the context of the current call for explainable AI models.
Methods/Materials: Informational processing in imaging involves consolidating salient patterns, illustrated in the radiologist scan path. Scan path is the temporal sequence of saccades and fixations produced in the assessment of an image, which can be plotted by eye tracking. The current state of eye-tracking technologies in radiological imaging and their use in AI datasets is evaluated.
Results: Novice and expert radiologists differ meaningfully in visual search patterns and their search errors can be stratified into a few key types, illustrated here. These aspects should be considered when actioning eye tracking data. In a 2020 study, eye tracking with speech recognition was 92% accurate in labeling lesion locations for a neuroradiology AI training dataset, thereby demonstrating that fully simulated interpretation can yield reliable lesion labels for training datasets. This is a potential avenue for large scale labeling of medical images on which to train algorithms, which fully embraces the explainable AI model and potentially has time saving qualities.
Conclusions: AI decision support tools are increasingly being investigated as radiological adjuncts. Eye tracking can be employed with high accuracy to label medical images. When employed to train AI, these data have potential to answer the strong call for explainable AI in healthcare through generating verified activation maps. This area would benefit from further study and multicenter evaluation both from a practical and efficacy perspective.
