An Eye-Tracking Experiment Investigating Diagnostic Regions of Interest
Introduction
Large images or video sequences require considerable storage capacity and transmission capacity. Standards for image compression that have been established can be classified into two categories:
i. lossless schemes, in which the reconstructed image is actually identical to the original picture on a pixel-by-pixel basis,
ii. lossy schemes, in which the compression procedure has discarded redundant information resulting in the reconstructed picture differ from the original one.The advantage of the lossless compression method lies in the ability to produce compressed images without losing information, however, it is not possible to reach very high compression rates.
Alternatively, lossy compression schemes offer the opportunity to reach high compression rates but this is achieved at the expense of information loss. This loss might not be so important to certain types of images, however, for medical images, it is imperative that methods ensure a combination of high quality as well as efficient compression.
In order to facilitate the reception of a high quality reconstruction, we take into account the fact that not all of the regions may be of equal importance. Of course, it is essential to identify regions that are vital, in which all information is diagnostically important.
The objective of the following work is associated to the discovery of regions of diagnostic interest in medical images by means of an eye-tracking procedure.
The eye-tracking apparatus
This experiment was conducted in the School of Psychology at the University of Birmingham. We are indebted to Professor Antje S. Meyer of The Behavioral Brain Science Centre, for very kindly making the eye-tracking equipment available and for assisting us in it's use. Our first 'experimental subject' was Dr John Townend, our research collaborator, from the Department of Cardiology in the Queen Elizabeth Hospital. The experiment lasted approximately twenty minutes. Throughout the experiment the cameras of the device were tracking the position of his eyes.
The apparatus we used to track the motion of the eye consists of a helmet to which two mini-cameras are suspended and aligned to point at the pupil of each eye. Each of the two cameras monitors the motion of the corresponding pupil. Before an experiment takes place, calibration and fixation procedures are performed by tracking the eyes as they follow objects across the computer monitor. The motion of both eyes are recorded in a file containing a history of the pupil positions in the 2D space. The path defined by these numbers is not represented by a continuous line, demonstrating in this way the discrete character of the eye movement.
Application to medical images
After calibration and fixation, a series of original medical images were displayed at 10 second intervals and our subject (senior conulstant cardilogist Dr John Townend) was required to focus on interesting features of the picture (for instance, to look at the arteries of an angiogram). Meanwhile, the cameras record all pupil movements to file. Our objective in this expoeriment was the identification of certain features of the image that will result in defining a region of interest.
In the next experiment, pairs of images were shown to the subject at 10 second intervals. Each pair consisted of an original uncompressed image and a decompressed image. The subject was required to focus on features of interest again and to compare the quality of the images.
Analysis of the experiment
As the subject looks at specific areas of interest in the angiogram, the position of the eyes is represented by a series of red dots in the image (movie no.1). The set of the dots define a region of interest as shown below.
Audio 1 includes the comments of the subject on the reconstructed image and the corresponding eye tracking ( Movie 2), the formation of a region of interest ( ROI2) and the associated comments ( Audio 2) on a second image.
A region of interest (shaded in red) as defined eye tracking.
Last updated 19 November 2000