Recent investigations of human signal detection performance for noise limited tasks have used statistically defined signal or image parameters. The Bayesian ideal observer procedure is then nonlinear and analysis becomes mathematically intractable. Linear, but suboptimal, observer models have been proposed for mathematical convenience. Experiments by Rolland and Barrett involving detection of completely defined signals in white noise superimposed on statistically defined (Lumpy) backgrounds showed that the Fisher-Hotelling model gave a good fit while the simple nonprewhitening (NPW) matched filter gave a poor fit. Burgess showed that the NPW model can be modified to fit their data by adding a spatial frequency filter with response similar to the human contrast sensitivity function. New experimental results will be presented demonstrating that neither model is satisfactory. The results of our experiments done with a variety of spectral densities for the background can be described by a Fisher-Hotelling model modified to include simple circularly symmetric spatial frequency channels as proposed by Myers and Barrett. However, results of our variable viewing distance experiments do not agree with predictions of this simple channelized model. It will be necessary to use a more complex F model with physiologically reasonable spatial frequency channels.

The effect of target size and size uncertainty on human observer ability to see disk targets in uncorrelated noise was measured. disk targets were centered in 64-pixel-radius areas of uncorrelated Gaussian noise. Human observers rated the likelihood that a target was present Size uncertainty was introduced in the target-present stimuli by using disk targets with radii ranging 2 to 33.2 pixels. A constant matched-filter signal-to-noise ration was maintained across the range of disk sizes by adjusting the disk contrast. For this mixed size experiment the observer ratings were analyzed using a multiple-distribution extension of the binormal ROC curve fitting procedure. A control experiment measured observer performance in conditions with target-present stimuli of known-size disks. A third experiment evaluated the influence of noise-area size on performance with known-size disks. An observer detection efficiency index, the square of the ration of d' to SNR, decreased at small and large disk radii. The efficiency index decrease for small disks was less in the control experiment (size- known). Observer efficiency indexes for medium and large disks were not significantly difference for the mixed size experiment and the control experiment. Reducing the noise-area size increased the efficiency for small disks and produced an approximately constant efficiency for the small to medium sized disks. Size uncertainty decreased observer detection performance relative to known-size performance for small disk targets. the observer efficiency index for the small targets was increased when small noise areas are used. This finding suggests that the decreased efficiency index for small targets on large noise areas was caused by increased observer uncertainty of target location.

This paper presents as new approach to image recognition based on a general attraction principle. A cognitive recognition is governed by a 'focus on attention' process that concentrates on the visual data subset of task- relevant type only. Our model-based approach combines it with another process, focus on attraction, which concentrates on the transformations of visual data having relevance for the matching. The recognition process is characterized by an intentional evolution of the visual data. This chain of image transformations is viewed as driven by an attraction field that attempts to reduce the distance between the image-point and the model-point in the feature space. The field sources are determined during a learning phase, by supplying the system with a training set. The paper describes a medical interpretation case in the feature space, concerning human skin lesions. The samples of the training set, supplied by the dermatologists, allow the system to learn models of lesions in terms of features such as hue factor, asymmetry factor, and asperity factor. The comparison of the visual data with the model derives the trend of image transformations, allowing a better definition of the given image and its classification. The algorithms are implemented in C language on a PC equipped with Matrox Image Series IM-1280 acquisition and processing boards. The work is now in progress.

Detection of low-contrast discs in computed radiography (CR) radiographs of a phantom was investigated as a function of film density. A Leeds Test Object 10 phantom was used that contains discs 0.25 to 11.1 mm in diameter. A series of radiographs was obtained using identical Cr imaging plates exposed to approximately 1 mR. The imaging plates were processed in a CR system using a linear characteristic curve sequentially adjusted to modify the density of the output film while maintaining a constant image contrast. Twelve radiologists viewed the resultant films in a darkened room, indicating the number of discs visible at each hole size. The mean number of discs seen at a constant film density of 1.11 was 58.7 +/- 2.8 out of the total of 108 discs present. Variation in disc visibility between radiologists ranged +/- 30% about the mean value. Optimal imaging performance occurred at a density range of between 0.8 and 1.0. Disc detectability was reduced to 90% of this optimal value at lower and upper film densities of 0.55 and 1.50, respectively.

The quantification of object saliency in 2D grayscale images is of primary importance in medical imaging. Saliency, in this paper, refers to the detectability of objects of unknown location, an object being a feature of interest in an image. Object saliency is commonly measured by the speed of an observer at performing a detection task. The more salient an object is, the more quickly it can be detected. The questions raised in this paper are (1) whether the degree of saliency of an object can be solely predicted from its detectability as measured in a location know-exactly task, or whether factors such as geometry and context contribute to saliency in a more complex fashion and (2) whether the contribution of geometry and background complexity to saliency can be quantified. This paper focuses on the problem of detection of stenoses in simulated angiograms. Results from a first such study are presented. From those results a new general methodology to measure saliency in 2D grayscale images was inferred and is presented.

To determine if presently used exposure levels in pediatric imaging can be reduced without loss of information or decrease in diagnostic accuracy, a method for multiple (stacked) image detectors and filters using a computed radiography system was used to obtain identical images of different exposure levels of neonates with either no active lung disease or hyaline membrane disease. Physical characteristics of the images were measured. A contrast-detail study and an ROC study were conducted to measure observer performance. Physical measurements and the contrast detail study revealed that all images were essentially x-ray photon noise limited. The ROC study indicated that diagnostic accuracy did not decrease significantly up to about 75% exposure reduction levels, although image quality rating data decreased with each exposure reduction. Decreasing exposure levels to about 75% of current levels may be acceptable in some clinical situations where dose is a concern, such as in pediatric imaging.

Using free-response methodology, we measured the comparative performance of trained observers for the task of detecting simulated pulmonary nodules in computed radiographic chest images that were compressed using the full-frame discrete cosine transform algorithm. six observers read fifty-one images containing a total of 372 simulated lesions of size ranging from 8 mm to 12 mm and with six different contrasts. The images were compressed to an average of 15:1 with the same parameters that were used in an earlier two-alternative forced- choice analysis. The results showed this level of compression did not increase the number of false=positive calls per image. also, observers tended to ignore the lowest contrast nodules in all images. At low contrast we expect compression to have the greatest effect. Therefore, overall performance was not degraded by the compression process, although performance was compromised at very low contrast.

In order to subjectively determine acceptable dose levels for portable CR, two blind experiments were performed while maintaining conventional screen-film radiography as a reference quality. In the first experiment, a team of two technologists and two radiologists were trained to expose and to quality-assure portable CR images obtained with standard screens at conventional exposures (400 speed). After providing them with a new set of cassettes and informing them that these were 'better and faster,: they were allowed to practice clinically, using the system for bedside AP chest radiography. After four weeks, exposure factors used in those cases were reviewed, and the experiment was repeated with two different teams and two CR systems. In the second experiment, portable CR was used for a large number of cases. The number of radiologists who complained about CR image quality was monitored for cased were alternating technologies (CR and conventional) were used and routinely viewed side by side. during a two-month period, radiation dose gradually increased to a level where radiologists' complaints were significantly reduced. In both experiments, exposure levels gradually migrated to and stabilized at 40-60% higher levels than that routinely used with conventional 400-speed film screen techniques. The perceived need for high exposure ratios between Cr and conventional radiography was related to body size. When the reference quality 'gold standard' remains unchanged, Cr requires higher exposures to yield acceptable image quality, particularly in large patients.

Sample size in ROC studies may be significantly reduced by including only difficult cases in the image set, but variability across studies may be a possible obstacle in the development of such a methodology due to case selection. To assess this situation, 300 cases used in a previous large ROC study, which included nine observers, were independently classified as subtle or typical by two experienced readers. Data from the previous study were reanalyzed using data sets consisting only subtle or typical images as designated by each classifier. Results showed a marked decrease in observer performance from the original study's results when only the subtle cases of either classifier were included in the analyses. For 12 of 15 possible comparisons (3 imaging modes and 3 diseases for subtle cases of either classifier were included in the analyses. For 12 of 15 possible comparisons (3 imaging modes and 3 diseases for subtle cases, 3 modes and 2 diseases for typical cases), the Spearman rho correlation coefficient between the performance indices computed for each reader for the subsets classified as subtle and typical was high and significant (P < 0.05). The results obtained in this preliminary study are encouraging and point out areas that warrant further investigation.

Image quality associated with image compression has been either arbitrarily evaluated through visual inspection, loosely defined in terms of some subjective criteria such as image sharpness or blockiness, or measured by arbitrary measures such as the mean square error between the uncompressed and compressed image. The present paper psychophysically evaluated the effect of three different compression algorithms (JPEG, full-frame, and wavelet) on human visual detection of computer-simulated low-contrast lesions embedded in real medical image noise from patient coronary angiogram. Performance identifying the signal present location as measure by d' index of detectability decreased for all three algorithms by approximately 30% and 62% for the 16:1 and 30:1 compression rations respectively. We evaluated the ability of two previously proposed measures of image quality, mean square error (MSE) and normalized nearest neighbor difference (NNND), to determine the best compression algorithm. The MSE predicted significantly higher image quality for the JPEG algorithm in the 16:1 compression ratio and for both JPEG and full-frame for the 30:1 compression ratio. The NNND predicted significantly high image quality for the full-frame algorithm for both compassion rations. These findings suggest that these two measures of image quality may lead to erroneous conclusions in evaluations and/or optimizations if image compression algorithms.

Receiver operating characteristic analysis has evolved as a useful method for evaluating the discriminatory capability and efficacy of visualization. The ability of such analysis to account for the variance in decision criteria of multiple observers, multiple reading, and a wide range of difficulty in detection among case studies makes ROC especially useful for interpreting the results of a viewing experiment. We are currently using ROC analysis to evaluate the effectiveness of using fused multispectral, or complementary multimodality imaging data in the diagnostic process. The use of multispectral image recordings, gathered from multiple imaging modalities, to provide advanced image visualization and quantization capabilities in evaluating medical images is an important challenge facing medical imaging scientists. Such capabilities would potentially significantly enhance the ability of clinicians to extract scientific and diagnostic information from images. a first step in the effective use of multispectral information is the spatial registration of complementary image datasets so that a point-to-point correspondence exists between them. We are developing a paradigm of measuring the accuracy of existing image registration techniques which includes the ability to relate quantitative measurements, taken from the images themselves, to the decisions made by observers about the state of registration (SOR) of the 3D images. We have used ROC analysis to evaluate the ability of observers to discriminate between correctly registered and incorrectly registered multimodality fused images. We believe this experience is original and represents the first time that ROC analysis has been used to evaluate registered/fused images. We have simulated low-resolution and high-resolution images from real patient MR images of the brain, and fused them with the original MR to produce colorwash superposition images whose exact SOR is known. We have also attempted to extend this analysis to real patient data, using magnetic resonance and single photon emission computed tomography 3D images whose SOR is estimated, but not known exactly. Results suggest that ROC analysis is useful for evaluating observer performance in detecting misregistration when viewing three orthogonal phases of colorwash superposition images, and that these results can be related to image measurements.

Previously an outline model of the radiological diagnostic process has been proposed which posits the importance of the initial glance at a medical image in helping to establish an appropriate diagnosis. As part of a long tern study of knowledge elicitation in mammography we examine the amount of information available to breast screening radiologists within the initial 'glance' at a mammogram. These data are compared to those from examining the same images normally. Overall, performance in a brief presentation was poorer than in normal viewing, as expected, but was also worse than found in comparable brief presentation studies using the chest radiograph. These results are discussed with regard to the inferences which can be made about the nature of mammographic knowledge which is utilized in the diagnostic process and how it is organized within the framework of a conceptual model.

A stochastic model, based on a Markov process, is developed for predicting visual search patterns of mammographers. Data for the model comes from tests where mammographic experts searched conventional screen film mammograms for microcalcifications. A three state model is proposed: (1) long dwells, (2) short dwells, and (3) no dwells, or periods of rapid eye movement. Steady state transition probabilities are reached within fourteen iterations.

The purpose of this study was to investigate the relationship between physical performance characteristics (such as signal-to-noise ratio and Detective Quantum Efficiency (DQE)) and psycho-physical performance (probability of detection), when aperiodic objects on a uniform background are imaged using two digital mammographic systems. The task simulated the detection of microcalcifications. A contrast detail study was performed using the Dutch CDMAM contrast-detail phantom. This phantom uses objects of different diameter and thickness. X-ray images of this phantom were generated by two digital x-ray imaging systems, one using a fiber optic taper to couple the light from a Min-R type phosphor to a CCD, the other one using a lens to couple the light from a Lanex phosphor to a CCD. Images were presented to human observers on the CRTs of the imaging systems in the context of a target detection task. Signal-to-noise ratio, MTF and DQE of both imaging systems were determined using standard image evaluation techniques. The lens coupled system had the highest DQE at low spatial frequencies, but a low MTF and DQE at high spatial frequencies. It yielded the highest detection probability overall in the observer performance study. The fiber optic system on the other hand had a significantly lower DQE at low spatial frequencies, but at high spatial frequencies it had significantly higher DQE and MTF than the lens coupled system. Its probability of detection throughout the performance studies was significantly lower than that of the lens coupled system. Furthermore, the probability of detection of the fiber optic system for small objects did not reflect its superior performance with respect to DQE and MTF at higher spatial frequencies. Presenting the DQE as function of object diameter rather than as function of spatial frequency permitted calculating the detection probability and fitting the Rose Model of Vision. The results serve as a reminder, that the detection of small aperiodic objects, even down to a diameter of about 100 micrometers is not only determined by DQE and MTF at high spatial frequencies but also by DQE and MTF at low spatial frequencies.

Breast cancer screening in the UK has so far been a consultancy-led service. Concerns over future staffing levels and increasing evidence that dual film reading improves cancer detection have led to an investigation of the feasibility of training lay film readers. This research reports the findings of a longitudinal study of seven radiographers reading mammograms for breast cancer screening. The study was designed so that screening mammograms were read independently by two radiographers and a radiologist. Film reading performances was also assessed in terms of screening outcome measures, such as recall, assessment outcome and biopsy outcome. The effect od double reading by radiologist/radiographer pairs was also examined. It was found that radiographers yield high agreement with radiologists and with screening outcomes. Dual reading resulted in increased sensitivity in line with expectations from previous studies on radiologist double reading. It was concluded that the role of the radiographer needs to be defined more precisely in relation to the dual reading model adopted in order to assess training needs.

The perceptual skills of the radiological expert are acquired through cumulative exposure to large numbers of radiographs. This extensive practice probably engenders two kinds of perceptual learning: 'low- level' feature learning, and 'search-learning'. We report here two experiments aimed at establishing the relative importance of these two mechanisms. The first experiment investigated detection of high spatial frequency - low contrast 'targets' in radiographs, and sought to locate the level at which learning occurred by using transfer of training across eyes and across stimuli as indicators of the locus of learning. Detection times improved after extensive practice (200 trials a day for 8 days) with no loss in accuracy. This improvement transferred across eyes, but transferred just partially to targets at new levels of contrast. This pattern of results suggests that both low-level feature learning and some more general, higher level learning, had occurred. Experiment two investigated this possibility in more detail using a visual search paradigm. The possibility that perceptual learning included a change from serial to parallel processing was explored, and measures of transfer of learning to new distractor sets and to novel stimulus dimensions were used to explore the likely locus of learning. Reaction times to the presence of absence of targets among distractors on a random noise background improved as the result of practice (1440 training trials). This learning transferred across eyes, to a new stimulus dimensions set and to different distractor sets. Further, true positive detection rates showed an increase for stimuli with a homogenous distractor set whilst there was no change in false positive rates. The implications of the results for training programs are discussed.

Radiological image quality can be objectively quantified by the statistical decision theory. This theory is commonly applied with the noise of the imaging system alone (quantum, screen and film noises) whereas the actual noise present on the image is the 'anatomical noise' (sum of the system noise and the anatomical texture). This anatomical texture should play a role in the detection task. This paper compares these two kinds of noises by performing 2AFC experiments and computing the area under the ROC-curve. It is shown that the 'anatomical noise' cannot be considered as a noise in the sense of Wiener spectrum approach and that the detectability performance is the same as the one obtained with the system noise alone in the case of a small object to be detected. Furthermore, the statistical decision theory and the non- prewhitening observer does not match the experimental results. This is especially the case in the low contrast values for which the theory predicts an increase of the detectability as soon as the contrast is different from zero whereas the experimental result demonstrates an offset of the contrast value below which the detectability is purely random. The theory therefore needs to be improved in order to take this result into account.

Digital display workstations are now commonly used for cross-sectional image viewing; however, few receiver operating characteristic (ROC) studies have been performed to evaluate the diagnostic efficiency of hard copy versus a workstation display for neuroradiology applications. We have performed an ROC study of film and 1K workstation based on the diagnostic performance of neuroradiology fellows to detect subtle intra- axial (high density (HD) and low density (LD)) and extra-axial (fluid, blood) lesions presented on computed tomographic (CT) images. An ROC analysis of the interpretation of approximately 200 CT images (1/2 normals and 1/2 abnormals) was performed by five experienced observers. The total number of abnormal images were equally divided among the three represented types of lesions (HD, LD, and extra-axial lesions). The images comprising the extra-axial lesion group were further subdivided into the following three distinct types: subdural hemorrhage, subarachnoid hemorrhage, and epidural hemorrhage. A fraction of the abnormal images were represented by more than one type of lesion, e.g., one abnormal image could contain both a HD and LD lesion. The digitized CT images were separated into four groups and read on the standard light box and a 1K workstation monitor equipped with simple image processing functions. Confidence ratings were scaled on a range from 0 (least confident) to 4 (most confident). Reader order sequences were randomized for each reader and for each modality. Each observer read from a total of eight different groups with a four-week intermission following the fourth group. The randomly assigned image number, lesion type and approximate location, incidental findings and comments, and confidence ratings were reported in individual worksheets for each image. ROC curves that were generated and analyzed for the various subgroups are presented in addition to the overall generalized jackknifed estimates of the grouped data. Also, 95% confidence intervals are presented for the differences in the area under the ROC curves. Although there were no statistically significant differences in the diagnostic accuracy between the original CT slice with HD and LD lesions viewed on the light box and on the 1K display workstation, the observers tended to record extra- axial lesions more frequently and with more confidence on the 1K display in comparison to the light box primarily due to the added advantage of adjusting the display window levels.

Screen film radiography often fails to optimally display all regions of anatomy on muskuloskeletal exams due to the wide latitude of tissue densities present. Various techniques of image enhancement have been applied to such exams using computerized radiography but with limited success in improving visualization of structures whose final optical density lies at the extremes of the interpretable range of the film. An existing algorithm for compressing optical density extremes known as dynamic range compression has been used to increase the radiodensity of the retrocardiac region of the chest or to decrease the radiodensity of the edge of the breast in digital mammography. In the skeletal system, there are regions where a single image may contain both areas of decreased exposure that result in light images and areas of higher exposure that result in dark regions of the image. Faced with this problem, the senior author asked Fuji to formulate a modification of the DRC process that incorporates a combination of the curves used for chest and breast images. The newly designed algorithm can thus simultaneously lower the optical density of dark regions of the image and increase the optical density of the less exposed regions. The results of this modification of the DRC algorithm are presented in this paper.