A statistical evaluation of eye-tracking data of screening mammography: Effects of expertise and experience on image reading

Abstract

Screening mammography has been widely used over the last few decades to detect breast cancers at an early stage using low-dose X-ray imaging. Being able to detect breast lesions early is highly beneficial for the patients as it makes treatment more successful and increases the chance of recovery. However, lesion detection is prone to false negative and false positive results due to the subtlety or low prevalence of the lesions, which compromises patient safety. It is critical to understand human perception and interpretation of images and use such understanding to develop ways to help increase the accuracy of cancer detection. In this paper, we present a large-scale eye-tracking study where 1,568 eye movement trials of expert breast radiologists, trainee radiologists, and physicists were recorded while assessing a large number of mammogram images. First, we statistically analysed viewers’ gaze behaviour based on their mean fixation duration. Results demonstrated that there was no difference amongst experts, and that trainees’ mean fixation duration was significantly shorter than experts and physicists’ fixation duration was significantly longer than that of the experts. Second, we evaluated spatial gaze deployment of the three viewer groups and quantified the difference between groups. Results showed that trainees and physicists deviated from where expert radiologists looked in images, and trainees performed better than physicists in terms of focusing on regions of interest. Finally, we investigated the saccadic behaviour of viewers and illustrated the differences between viewer groups in terms of saccade amplitudes and orientations.