Eye-tracking methods are fundamental for the study of medical image perception and central to discoveries about why radiologist interpretative failures occur. However, analysis of eye-tracking data has been rather primitive and ignores potentially valuable information. We propose a new approach to the analysis of eye-tracking data based on a recently developed model for the analysis of receiver operating characteristic (ROC) data, free-response operating characteristic (FROC) mark-rating data, and location- specific ROC (LROC) data. Based on Chakraborty's model, we propose a method that integrates the analysis of true and false positive characterizations of eye position data and uses the degree of suspicion of radiologist indicated suspicious regions. The inclusion of the FROC performance data is expected to yield better understanding of image perception than is possible via eye-tracking alone. The integrated analysis is enabled by two recent developments: the Chakraborty search model and a method for estimating its parameters. The parameters of the model correspond to physical quantities that are measured in eye-tracking studies. The project consists of quantifying these correspondences, providing a demonstration of integrated analysis, and showing its advantages over eye-tracking alone. Eye-tracking and FROC studies have so far proceeded along independent tracks, one to understand image perception and the other to measure performance. This project shows how a combined approach can yield a more powerful tool for analyzing eye-tracking data and understanding image perception. With better understanding of image perception based on improved analysis, we will be better able to improve diagnostic performance. Applications of this method include radiologist training and improved CAD algorithms. The rich dataset of simultaneously acquired FROC and eye-position data, and analysis software will be made freely available at the close of our project.
Eye-tracking apparatus measures where radiologists look. This information is fundamental to understand medical image perception and central to learning why radiologist interpretative failures occur. However, analysis of such data has been primitive and ignores valuable information. We propose a new approach to the analysis of eye-tracking data. With better understanding of image perception we will be better able to improve radiologist performance and reduce interpretive errors.
Showing the most recent 10 out of 25 publications
