Visual search, perception and cognition when reading stack mode cranial CT

Abstract

Today, radiology is one of the most important diagnostic disciplines within medicine. In the last two decades, the discipline has changed tremendously by going digital and increasing the importance of multi-slice images. At the same time, there is little known about the processes of visual search, perception and cognition involved in their interpretation. Medical image interpretation research mostly concentrates on conventional radiography, possibly a result of the lack of an experimental framework that enables the study of perception processes in multi-slice imaging. The main aim of this thesis therefore is to enable the study of visual search, perception and cognition in multi-slice images. To this end, gaze parameters that have proved useful in the study of gaze behavior in conventional radiography interpretation were adapted to t the challenges of stack mode reading. An important step herein is to calculate xations independent of the stimulus material as they often cover multiple slices. Furthermore, additional parameters were developed for the multi-slice context, which describe the z-component of gaze represented by scrolling. The thesis aimed at testing these parameters by studying two di erent factors that in uence the reading process. One factor is decision outcome. In a rst experiment 16 radiologists were asked to identify hemorrhages on cranial CT cases. Their gaze and scrolling behavior linked to True Positive, True Negative, False Positive and False Negative decisions was examined. It emerged that radiologists dwelled longest on True Positive decision sites, followed by False Positive, False Negative and nally True Negative decision sites. Radiologists took considerably longer to decide on False Negative as compared to True Positive sites and on False Positive as compared to True Positive sites. Di erences between False Positive decisions and other decision outcomes have also been identi ed with regard to the number of xations and reading time. The second and third experiments concentrated on the factor 'image size'. A total of 43 radiologists in two institutions were asked to interpret cranial CT and to identify any intracranial hemorrhages that were present in twenty cases of two di erent sizes (14 x 14 cm versus 28 x 28 cm). Performance, reading time and preference of the radiologists did not di er between the two sizes. The experiments have hence demonstrated that none of the two image sizes is globally preferable to the other. However, perceptual processes appear to di er between the two image sizes: in small images xations were longer and covered more slices. This was associated with the use of motion detection, as scrolling though the stack while resting the gaze in one position is linked to holistic recognition. Furthermore, in both institutions, the time to rst xation of True Positive locations was shortened for small as compared to large images. Detection of targets was hence faster in these images. At the same time, dwell time was prolonged in small images, suggesting that radiologists can better resolve large images. Based on the results of the two experiments, the use of small images can be recommended during the initial phase of interpretation to signal out potentially perturbed structures that are worth examining in detail. For a closer examination, these structures should be enlarged as larger structures appeared to be easier to resolve. The three experiments have yielded valuable insights into the challenges of multi-slice medical image interpretation research and demonstrated that scienti c perception research of interpreting multi-slice images is possible and should be used to improve modern radiology.