Eye Movements and Visual Search in Dentistry

Dental Radiology as a Model for Perception Studies

¹ Academic Center for Dentistry, The Netherlands
² Rutgers University,
³ Saint Peter's College,
* University of Medicine and Dentistry of New Jersey,

There are a number of significant reasons for studying perception problems in radiology. In this abstract we consider four, namely, optimizing imaging chains; learning about image interpretation and the behavior of radiologists; the necessity of using other information for diagnosis; and automating vision and perception tasks. The purpose of this presentation is to introduce the idea of using dental radiology as a model for perception studies. While all of the traditional problems are still present, there are many factors that make dental radiology a unique model of perception.

In dental radiology, the most common film format is an order of magnitude smaller (3 x 4 cm) than whole body films that are typically used in eye movement studies. These intraoral films are used for a variety of diagnostic tests, but in each case only 2 to 3 teeth are imaged per film. A radiographic examination of the entire mouth requires 18-20 film exposures. There is a great need to understand the process as the dental radiograph is the most common radiograph examination.

Imaging chains are not necessarily fixed as a number of electronic recording devices are available to replace film. Films are not always observed on a masked light box, indeed, in practice a light box may not be used at all. Furthermore, the majority of film interpretations are done by general practitioners, not radiologists. There is a wide variety of structure in a single film, and there can be both embedded and nonembedded targets present at the same time. Film display and context also vary, as films are rarely presented by themselves, but in sets. Because of the frequency of these examination and the variability of the interpretations, there is a great deal of interest in automating interpretation tasks in order to reduce observer variability.

In this talk we shall present some initial background on dental radiology and illustrate how perception problems can be modeled in dental radiology. Some initial experimental paradigms and a summary of the problems that are to be studied will be presented.

Eye Movements on Dental Images

Studies of eye movements can help to provide some understanding of visual perception tasks. In this particular investigation, expert and novice observers were used to ascertain information about perception in dental radiology. Various types of dental diseases located in the hard tissues (enamel, dentine and bone) were used for the study. Since this perception task requires the observer to access information at various locations in the image and merge the information together to make a diagnosis, the eye movement study can provide insight into how the data are obtained and utilized. To compare the performance of novices and of experts, 16 radiographs with and without dental pathologies were presented to three groups of rive observers each a group of dental students before the start of their first year (no expertise); a group of dental students at the end of the first year (low level of expertise); and a group of dental/general practitioners (expertise present). The observers were asked to point out the conspicuous areas and, if possible, to label them. These areas were categorized by a dental radiologist into nine categories of dental radiographic entities. Some of these entities refer to normal anatomical structures, others to dental pathology. Another feature space was created by categorizing these areas as visually conspicuous, cognitively conspicuous or a combination of both Differences were analyzed using MANOVA and contrast analysis. In a subsequent part of this study, the eye movement patterns of a group of eight novices with low level expertise and of eight experts were registered during observation. For this part of the study, only those radiographs were used that had been categorized during the first part of the study as having the same set of conspicuous areas for novices and experts. Differences were tested with a t-test.

The results show that experts mention less visually conspicuous areas; no differences were found between the two groups of less experiences observers in this respect. Significant expertise effects were found for cognitively conspicuous areas. The results show, furthermore, that experts have more fixations in cognitively conspicuous areas. Novices were more distracted by visually conspicuous areas.

Expert Eye Movements as a Sample Space for Image Processing

Dennis Carmody, Ph.D.,³ Stanley M. Dunn, Ph.D.,² Mel Kantor. D.D.S.,* Elly van der Stelt-Schouten, MS.¹ and Paul F. van der Stelt, D.D.S., Ph.D.¹

The recording of eye movements during visual tasks can provide insight into many aspects of visual perception. In this work, the efforts of experts in dental radiology were explored. The specific domain studied consisted of various stages of periapical disease. A unique aspect of this domain is that diagnosis requires that the observer access information at various locations in the image and merge the information together to make a decision. Thus, the eye movement study can provide insight into how the data are gathered and processed. This approach is markedly different than other studies in which the images each contain a specific target for which the observer is searching.

The image data set included eight of each of three stages of periapical disease and eight normal controls. For each observer, recordings were made over two sessions and included calibration procedures for each image. The observers were asked to provide a diagnosis after viewing each image. The diagnosis was made without the use of additional clinical data and the tooth crown was excluded from the image in an effort to prevent the possibility of secondary information influencing the diagnosis. Data analysis was performed using the scan path information, fixation locations and the diagnosis provided by each subject.

In this talk the results which will be emphasized are those which relate to the nature of the eye movements, such as scanning patterns and particular regions of interest. Although not studied independently, there are several factors which may have had an influence on the results, including the effect of enlarging the original 3 x 4 cm images, which was required in order to obtain measurable eye movement during the experiment; the accumulation of information at various sites in the image; and the limited amount of information provided to the clinicians for diagnosis during the experiments as compared to a normal clinical setting, as described above.

Image Processing with Gaze Selected Sample Space

Biological visual process emulation has been a long-standing goal of research in image interpretation. In this work we focus on two aspects of biological visual processing for dental image classification; gaze directed feature selection and low-level processing.

The gaze directed feature selection is accomplished using >ecordings of eye movement data collected from experts (dental radiologists) while performing diagnosis on dental images. The raw position data are clustered into fixations using spatial and temporal criteria. Further processing of the image features occurs only in the regions of fixation. Thus, the areas surveyed in detail by the experts are also those considered important for classification by computer. The image analysis of the fixation regions simulates the low-level visual processing steps executed by simple receptive cells. This processing includes multi-scale analysis and edge detection at various orientations. Property lists containing the edge-related features at the fixation sites are presented to an inductive learning system. The inductive classification system is designed to learn the concepts of disease which are present in the data set.

The results of this work will be discussed in terms of classification performance and the effectiveness of the visual model for feature detection and analysis. This study can also provide insight into the manner in which visual perceptual tasks are accomplished and indicate which features are the most important for classification by humans and computers.

Back to Research Projects