| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
Commentary |
School of Medical Imaging Sciences, St Martin's College, Bowerham Road, Lancaster LA1 3JD, UK
Correspondence: Mr Tim Donovan, Senior Lecturer, Medical Imaging Sciences, St Martin's College, Bowerham Road, Lancaster LA1 3JD, UK. E-mail: t.donovan{at}ucsm.ac.uk
 |
Abstract |
|---|
 Top Abstract Visual search and eye...The Bayesian frameworkThe ideal observerRelevance for medical image...References |
|---|
|
Visual search and eye movements |
|---|
|
TopAbstractVisual search and eye...The Bayesian frameworkThe ideal observerRelevance for medical image...References |
|---|
In this short commentary, we reflect on the visual task and decision-making process in medical image interpretation and comment on how radiology acts as a source of endless fascination for vision scientists.
The interpretation of medical images and the development of expertise require both perceptual and cognitive skills. Visual search is a fundamental perceptual skill, whereas diagnostic reasoning and decision making are cognitive skills acquired over time through education and experience. At its most basic, the information processing flow to achieve expert interpretation of medical images is: visual search – object recognition – decision making [1]. During the search process, such as looking for a face in a crowd, the visual system has to solve some difficult problems. This is no different for the radiologist when presented with a medical image. How does the radiologist make sense of the different ambiguous shapes, shades and contours when searching for pathology? The use of eye-tracking as a research methodology has produced some interesting data about this and there are, for example, different patterns of eye movements between novices and experts, i.e. experts will exclude large areas of the image during a search for lung nodules and they concentrate on regions where, in their experience, nodules are more likely to occur [2]. Eye movements are not generally something we are aware of, but the types of eye movement in visual search are not involuntary. Rather, they can be described in terms of target selection, which in turn is related to the motivational state of the radiologist and to higher cognitive processes [3]. Early vision research in medical imaging found that although most radiologists advocate some form of systematic search, eye-tracking studies do not reflect that radiologists actually do this. It was suggested that the radiologist who has to use a deliberate search strategy is paying attention to search and not to the task of image interpretation [4].
The reason why visual search is necessary relates to the variable spatial resolution of the retina; when presented with an image, the visual system will very quickly perform a global analysis before using high-speed eye movements called saccades to direct the highest resolution region of the retina, in order to fixate potential target locations in the visual scenes. It is during these fixations, which average 200–300 ms, that detection and identification processes are applied across the visual field, and when eye movements to subsequent fixation applications are planned and programmed [5]. Understanding search performance and eye movements is important because they relate directly to determining how errors are made in the detection and location of pathology, and to the understanding of how developing expertise changes the way neural resources are used. There is currently no formal theory of optimal eye movement strategies in conducting visual search [6], but recently there have been a number of important research articles advocating a Bayesian approach to perception, which is directly applicable to medical image perception. The Bayesian approach is attractive because it aims to formalize these ideas mathematically by modelling the ideal searcher and comparing human and ideal search quantitatively [7].
|
The Bayesian framework |
|---|
|
TopAbstractVisual search and eye...The Bayesian frameworkThe ideal observerRelevance for medical image...References |
|---|
Regarding perception, Bayesian theory takes account of the trade-off between feature reliability and priors [8]. The visual system is forced to guess if it has to make sense of an ambiguous image/scene where several different objects could have produced the same image. The visual system can, however, make intelligent guesses by biasing its guesses towards typical objects or interpretations. Bayes' formula implies that these guesses, and hence perception, are a trade off between image feature reliability as embodied by the likelihood and the prior probability. Some perception may be more prior driven and others more data driven. The less reliable the image features, the more the perception is influenced by the prior, such as geometry, shape and lighting.
How might the Bayesian framework apply to the radiological task? A radiologist is presented with a chest radiograph from a test bank which may or may not have a lung nodule. Depending on experience and expertise, the perceptual system will be primed towards a particular strategy as not all locations on a chest radiograph are equally probable in possible nodule locations. These are the prior probabilities. Initially, radiologists gain a global impression during a first fixation that uses peripheral vision. Experts have an advantage because they will have a strong prototypical idea of normal anatomy, so perturbed regions of anatomy are identified as possible nodule locations. The posterior probabilities of these locations are calculated and an eye movement (saccade) is made to a fixation location that will maximally increase the likelihood of identifying a nodule. This is repeated until search is completed. The Bayesian framework takes account of the criterion used to indicate whether an area of interest is actually a nodule or not.
The Bayesian framework is complementary to other theories such as signal detection theory. This theory can be explained as an explicit quantitative application of statistical decision theory in perception and cognition, as the theory specifies Bayesian ideal observers and optimal decision rules for simple detection and discrimination tasks [11].
|
The ideal observer |
|---|
|
TopAbstractVisual search and eye...The Bayesian frameworkThe ideal observerRelevance for medical image...References |
|---|
The Bayesian ideal observer is useful as a benchmark against which to compare human performance [5]. This approach also divides the perceptual task up into pieces and then combines them to understand the whole. They are also a useful point for developing realistic models [10]. In Bayesian networks, a generative model specifies the causal relationship between random variables (e.g. objects, lighting and viewpoint); one can also specify the task, i.e. the costs and benefits associated with different possible errors in the perceptual decision. The models can then allow the inference solution, i.e. the Bayesian ideal observer selects the interpretation that has the maximum expected utility.
In the paper by Najemnik and Geisler [6], they developed their "ideal searcher" by first characterizing visibility maps of the visual system under consideration for targets and backgrounds of interest. The ideal searcher works by simply collecting responses from possible target locations, updating the posterior probabilities, and then moving the eyes to maximize the new information. To optimally integrate responses across fixations, the ideal searcher accumulates the weighted responses from each potential target location.
Given the prior probabilities of possible target locations and the visibility maps which specify all relevant values of performance, one can simulate behaviour of the ideal searcher. The ideal searcher will make centre-of-gravity fixations. Saccade lengths of the ideal searcher are moderate in size because the posterior probabilities at nearby locations are suppressed, so that recently fixated locations are not returned to and posterior probabilities at distant locations tend not to be increased. Najemnik and Geisler [6] found that this was reflected in humans. Occasional exclusion saccades are made. The ideal searcher does well with parallel detection, integration of information across fixations and selection of the next fixation. Humans, however, are not very efficient at integration of information across fixations as apparently little can be gained by posterior probability information more than one or two fixations in the past. However, a coarse memory is needed to reduce the likelihood of returning to the same display region, and this explains inhibition of return.
Ideal observers can be quite useful in identifying the constraints in perceptual tasks, such as image noise or neural noise, and much of the measured variation in human perception may be due to pre-neural factors [11]. An example of where human performance does not match the performance of the ideal observer, and where the human may be limited by pre-neural factors, is the decline in the identification of familiar features with eccentricity from the fovea, which drops much more quickly than that predicted by the ideal observer [12].
Perceptual learning has also been studied in humans by comparison with an optimal Bayesian algorithm [13]. This showed that humans learn rapidly, and that this rapid initial learning reflects a property inherent to the visual task and not a property particular to the human perceptual learning mechanism.
|
Relevance for medical image perception |
|---|
|
TopAbstractVisual search and eye...The Bayesian frameworkThe ideal observerRelevance for medical image...References |
|---|
These findings are relevant in the design of viewing and reporting stations, and the use of image processing software/CAD so that it is matched to human abilities and limitations. It is pertinent that medical image perception research should concentrate on the observer as well as the technology as it seems interobserver differences are often greater than differences between imaging techniques, of which an excellent example is the study by Weatherburn et al [16]. This work found that the detection of chest lesions did not vary between conventional film, CR (computed radiography) hard-copy and PACS (picture archiving and communication systems) soft-copy images, but there were statistically significant differences between observers.
Bayesian theory is a useful starting point for trying to explain some of the eye-tracking patterns we see, and to provide insight into understanding image interpretation. This is because it provides a mathematical framework for representing the properties of the image, describing the image interpretation task and taking account of the costs and benefits associated with different perceptual decisions.
Received for publication September 20, 2006. Revision received December 13, 2006. Accepted for publication December 14, 2006.
|
References |
|---|
|
TopAbstractVisual search and eye...The Bayesian frameworkThe ideal observerRelevance for medical image...References |
|---|
This article has been cited by other articles:
 |
|
 |
|
  BJR review of the year -- 2007 Br. J. Radiol., April 1, 2008; 81(964): 265 - 269. [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| BJR | DMFR | IMAGING | ALL BIR JOURNALS |
