British Journal of Radiology (2006) 79, 437-440
© 2006 British Institute of Radiology
doi: 10.1259/bjr/13489819
An investigation of search pattern extent in the threshold contrast detection task
C J Kotre, PhD
Regional Medical Physics Department, Newcastle General Hospital, Newcastle-upon-Tyne NE4 6BE, UK
 |
Abstract
|
|---|
The threshold contrast-detail diameter test is used as a semi-quantitative measure of image quality in radiology. This observation task is called "signal known exactly/background known exactly" because the signals are usually low contrast disks in known positions, and the background is uniform except for noise fluctuations. The performance of the observer undertaking this task can to some extent be predicted from knowledge of the noise power in the image background, and adoption of the assumption that the noise is sampled through an aperture of the same area as the test feature being observed. In order to extend this approach to optimization of clinical images, the effect of the cluttered anatomical background on the detection task must be quantified. To study the effect on detection of nearby structure, a series of contrast-detail tests was carried out using a progressively restricted background area of Gaussian noise, and a range of object diameters. It was found that the observer's ability to detect low contrast objects is progressively reduced as the area of the search area is reduced, the difficulty of the task increasing rapidly as the diameter of the restricted search area falls to less than twice that of the target disk. The results suggest the presence of a search pattern that scales in proportion with the size of the test feature.
|
Introduction
|
|---|
The threshold contrast-detail diameter test has been used for many years as a standard semi-quantitative test of image quality especially in fluoroscopy, radiography and mammography [1–3]. In this technique, the observer is shown a set of disks imaged against a uniform background. The disks vary both in contrast and diameter and are laid out in a fixed pattern within the test object. This type of stimulus is known as a "signal known exactly/background known exactly" (SKE/BKE) task. For each disk diameter the observer works along a sequence of disks of decreasing contrast and decides which is the last visible disk. The contrast of the last visible disk is defined as the threshold contrast at that disk diameter. The process is repeated over a range of disk diameters and the results can be plotted as a contrast-detail curve, which marks the transition from what combination of size and contrast can be seen in the image and what cannot be seen because it is masked by noise.
The most widely accepted model of the behaviour of the human visual system in undertaking the SKE/BKE contrast-detail task is that the threshold contrast obtained is equal to the noise observed in the background, multiplied by a constant, the threshold signal-to-noise ratio, the value of which is thought to be around 2.5 [4]. The noise perceived by the observer is taken to be the standard deviation of a number of luminance samples of the background, where each sample is the mean luminance within an aperture of the same area as the disk under observation [5, 6]. This model broadly fits experimental results when modified to take account of further limitations to detection at small disk sizes due to the finite point spread function of the eye, and at large disk sizes due to the maximum extent of the photo-receptor field over which summation of the noise sample may occur [4, 6].
Although the SKE/BKE contrast-detail task has been shown to be very useful in comparing the performance of radiological image receptors, where the noise in the uniform background is principally that from X-ray quantum fluctuations plus components of system noise, such as electronic amplification noise and the fixed noise pattern of phosphor screens, there are difficulties in extending the results to the prediction of detectability in clinical images. The background in clinical images consists of a wide variety of object shapes, sizes and contrasts all of which tend to inhibit the detection of the diagnostic signal required in any given examination. The presence of the background can be thought of as a noise source, sometimes termed structure noise or clutter, and it is the removal of this noise component that allows small contrast signals to be detected in digital subtraction angiography. Although meaningful results can be obtained in observational studies of contrast-detail test objects overlying a complex anatomical type background [7], it is difficult to extend this to the general problem of dose optimization of clinical images. In order to undertake optimization it will be necessary to be able to quantify background structure as a noise source, and to be able to predict its effect on observer performance. One means of achieving this may be to design algorithms to sample the radiological image in a similar way to a human observer, so that the detectability of a low contrast object can be analysed on a point-by-point basis against the background of real clinical images. Although the present work is based on the contrast-detail test, other approaches to quantifying signal detection in realistic radiological backgrounds have been explored, notably alternative forced choice methods [8–10] and receiver operating characteristic studies [11, 12].
As a starting point for the design of algorithms to mimic an observer undertaking a contrast-detail task, it is necessary to extend the knowledge of the human observer noise sampling to investigate not only the area of the sampling aperture, but the spatial distribution of the samples and the effect on observer performance of the proximity of adjacent structures.
|
Experimental observations
|
|---|
A series of static contrast-detail test patterns were presented to six experienced observers on a monochrome TV display monitor. The display was first calibrated in terms of luminance using a Hagner S2 photometer (Hagner AB, Solna, Sweden) whose response is matched to that of the human eye, and the monitor settings were then left fixed for the duration of the experiment. The photometer was calibrated using a source of known luminous intensity in terms of illuminance, from which the luminance calibration of its internal detector was derived using the relationship between luminance of a surface and illuminance at a point for a Lambert source given by Guibelalde et al [13].
Each test pattern consisted of a 4x4 matrix of 16 circular test features decreasing in contrast in steps of 0.84 (reciprocal cube-root of 2). Computer-generated Gaussian noise was added to the whole image sufficient to produce a threshold cut-off within the contrast range displayed. The test images consisted of 512x512 8-bit pixels displayed in a 20 cmx16 cm format at a viewing distance of approximately 50 cm, although the observer was not restricted in viewing distance. The observations were carried out with room lights dimmed, and observers were allowed to score disks as "half seen" where detection was uncertain, following common practice in scoring contrast-detail tests in radiological image assessment.
Four detail diameters were used (32, 16, 8 and 4 pixels, equivalent to 11.3 mm, 5.6 mm, 2.8 mm and 1.4 mm, respectively) and each of these was displayed within a larger circular background area of mid-grey with noise added to both disk and background. Outside the area of mid-grey, the screen was blanked to black with no noise to limit the search area of the observer. The combinations of disk diameter (Dd) and search area diameter (Ds) were varied to give no search restriction, then Ds/Dd = 4, 3, 2, 1.5, 1.25 and 1.125 for each of the four disk diameters. The test patterns were presented in random order with no time restriction on the observation. Three example test patterns are reproduced in Figure 1
. Figure 1a shows the 16 pixel diameter disks with no search restriction (other than the limits of the image frame), Figure 1b,c shows the same diameter with Ds/Dd = 3 and 1.5, respectively.
View larger version (20K):
[in this window]
[in a new window]
|
Figure 1. Examples of test images.(a) The 16 pixel diameter disks with no search restriction (other than the limits of the image frame). (b,c) The same diameter with Ds/Dd = 3 and 1.5, respectively.
|
|
At the conclusion of all observations, the recorded scores were converted to luminance threshold contrasts (CT) and averaged across observers.
|
Results
|
|---|
Figure 2 shows conventional contrast-detail plots of some of the results with quadratic curve fits on logarithmic axes. To avoid obscuring the figure, only plots for the unrestricted search and Ds/Dd = 3, 1.5 and 1.125 are shown. As Ds/Dd is reduced, the threshold contrasts increase as expected. The error bar illustrates a typical±1 standard error of the mean value of threshold contrast for the experiment. Standard errors for each point were calculated, but are not shown to avoid obscuring the figure.
Figure 3 shows the ratio of threshold contrast for the restricted search, normalized to that of the unrestricted search, plotted against Ds/Dd, for all four disk diameters. Although the values are subject to considerable uncertainties, the trend with Ds/Dd appears to be similar for all diameters, suggesting the presence of a search pattern that scales directly with the size of the object being observed.
Taking the similarity of the results of Figure 3
as sufficient to support the hypothesis that the search pattern scales in proportion to the size of the object, Figure 4 shows the data from Figure 3 averaged across all disk diameters. It is notable that the mean ratio of threshold contrasts already has value greater than unity (although unity is within the error bar) at Ds/Dd = 4, indicating that some interference with the observer's search pattern is already present with this background. The ratio increases quickly for values of Ds/Dd < 2, as the visual task changes from finding a disk on a background to detecting the difference between the intensity of the disk and the narrow band of noisy background surrounding it. This finding is in agreement with previous results observed for a design of mammographic contrast-detail phantom, where a marker circle at 2.7 times the diameter of the test feature was found to reduce detection, but marker circles at greater relative diameters (>6) had little effect [14].
View larger version (9K):
[in this window]
[in a new window]
|
Figure 4. The ratio of threshold contrast for the limited search, normalized to that of the unlimited search, plotted againstDs/Dd for the data from Figure 3 averaged across all disk diameters. The error bars show±1 standard error on the mean.
|
|
|
Conclusions
|
|---|
It had been shown that for the SKE/BKE contrast-detail detection task undertaken against a background of Gaussian noise, the observer's ability to detect low contrast objects is progressively reduced as the area of the search area around the object is reduced, the difficulty of the task increasing rapidly as the diameter of the restricted search area falls to less than twice that of the target disk. The results further suggest the presence of a search pattern that scales in proportion with the size of the test feature. These results are relevant to the design of contrast-detail test phantoms for radiological imaging systems, and also to the broader investigation of the effect of background anatomical structure on the detection of diagnostic features in radiology.
Received for publication September 7, 2005.
Revision received November 2, 2005.
Accepted for publication November 14, 2005.
|
References
|
|---|
- Hay GA, Clarke OF, Coleman NJ, Cowen AR. A set of X-ray test objects for quality control in television fluoroscopy. Br J Radiol 1985;58:335–44.[Abstract/Free Full Text]
- Bijkerk KR, Lindeijer JM, Thijssen MAO. The CDMAM phantom: a contrast-detail phantom specifically for mammography. Radiology 1993;185(P):395
- Institute of Physics and Engineering in Medicine. Recommended standards for the routine performance testing of diagnostic X-ray systems. (IPEM Report 77) York: IPEM 1997.
- Schnitzler AD. Image-detector model and parameters of the human visual system. J Opt Soc Am 1973;63:1357–68.[CrossRef][Medline]
- Rose A. The sensitivity performance of the human eye on an absolute scale. J Opt Soc Am 1948;38:196–208.[CrossRef][Medline]
- Chesters MS, Hay GA. Quantitative relation between detectability and noise power. Phys Med Biol 1983;28:1113–25.[CrossRef]
- Kotre CJ. The effect of background structure on the detection of low contrast objects in mammography. Br J Radiol 1998;71:1162–7.[Abstract]
- Ohara K, Doi K, Metz CE, Giger ML. Investigation of basic imaging properties in digital radiography. 13. Effect of simple structured noise on the detectability of simulated stenotic lesions. Med Phys 1989;16:14–21.[CrossRef][Medline]
- Bochud F, Valley J-F, Verdun FR. Estimation of the noisy component of anatomical backgrounds. Med Phys 1999;26:1365–70.[CrossRef][Medline]
- Burgess AE, Jacobson FL, Judy PF. Human observer detection experiments with mammograms and power-law noise. Med Phys 2001;28:419–37.[CrossRef][Medline]
- Kundel HL, Nodine CF, Thickman D, Carmody D, Toto L. Nodule detection with and without a chest image. Invest Radiol 1985;20:94–9.[CrossRef][Medline]
- Reverez G. Conspicuity and uncertainty in the radiographic detection of lesions. Radiology 1985;154:625–8.[Abstract/Free Full Text]
- Guibelalde E, Vano E, Llorca AL. Quality assurance of viewing boxes: proposal for establishing minimum requirements and results from a Spanish quality control programme. Br J Radiol 1990;63:564–7.[Abstract/Free Full Text]
- Kotre CJ, Porter DJT. A printed image quality test phantom for mammography. Br J Radiol 2005;78:746–8.[Abstract/Free Full Text]
Top
Abstract
Introduction
), and Ds/Dd = 3 (
), 1.5 (
) and 1.125 (
). The error bar illustrates a typical±1 standard error on the mean value of threshold contrast for the experiment.
), 5.6 mm (
) and 11.3 mm (
).