| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Thoracic Imaging |
1 From the Department of Diagnostic Radiology, Henry Ford Health System, Detroit, Mich. Received November 23, 1998; revision requested February 11, 1999; revision received March 31; accepted July 1. Address reprint requests to M.J.F., Radiology Research (2F), Henry Ford Health System, 1 Ford Pl, Detroit, MI 48202 (e-mail: mikef@rad.hfh.edu).
 |
Abstract |
|---|
 TOP Abstract IntroductionMATERIALS AND METHODSRESULTSDISCUSSIONAppendix 1References |
|---|
MATERIALS AND METHODS: Sixty 8 x 8-cm lung pattern images were extracted from digital chest radiographs in healthy individuals. Sixty quantum mottle images of the same size and quantum noise level were extracted from uniformly exposed digital radiographs. Simulated nodules with various peak contrast-diameter products (CD) that emulated subtle tissue-equivalent lung nodules were numerically superimposed at the center on three-fourths of the images. Printouts were independently viewed and scored by five experienced radiologists. The area under the receiver operating characteristic curve (Az) was estimated as a measure of the detectability of the nodules.
RESULTS: At a fixed observer performance level (eg, Az = 0.8), much smaller and lower-contrast nodules were detected on quantum mottle images (1-mm diameter, CD = 0.01 mm), compared with those on anatomic images (4.5-mm diameter, CD = 0.20 mm). The findings generally agreed with the signal-to-noise ratio calculations based on statistical observer models.
CONCLUSION: The detection of subtle lung nodules on chest radiographs is limited by anatomic noise.
Index terms: Diagnostic radiology, observer performance • Lung neoplasms, diagnosis, 60.1215, 60.31 • Lung, nodule, 60.1215, 60.281 • Radiography, digital, 60.1215 • Receiver operating characteristic (ROC) curve • Thorax, radiography, 60.1215
|
Introduction |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONAppendix 1References |
|---|
The detection of abnormalities on medical images is generally understood to be limited by the amount of noise on the image. In the context of this article, noise is defined as the ensemble of all the variations (fluctuating intensities) present on the image that interfere with the detection of the "true" signal that is sought (nodule). (The term "noise" is used to describe "relative noise," or variations signal divided by the mean.)
On chest radiographs, there are two major sources of such variations: quantum noise (mottle), which reflects the variations due to the finite number of x-ray quanta that form the image, and anatomic noise, which reflects the highly "correlated" variations formed by the projection of anatomic features in the thorax, such as ribs, pulmonary vessels, and lung tissue.
The influence of quantum noise on the detection of low-contrast lesions, such as subtle lung nodules, is well understood (6,7). This influence can be minimized either by increasing the patient dose, an action that is generally discouraged, or by improving the detective quantum efficiency of the imaging system. However, substantial improvement in the detective quantum efficiency of imaging systems in the past few decades, which has reduced the level of quantum noise on chest radiographs, has led to no or little improvement in the detection of subtle lung nodules. The influence of anatomic noise on the detection of subtle lung nodules is well acknowledged in the clinical literature (8–12), but the processes by which this influence takes place are not well understood (13,14).
The goal of this study was to substantiate the relative influence of quantum and anatomic noise in the detection of subtle lung nodules on clinical chest radiographs. The objective was somewhat similar to that of a previous study (15) in which the detection of a single-size Gaussian nodule was assessed in the presence or absence of the anatomic pattern associated with a single chest radiograph with various levels of added noise. Noise on the radiograph was measured in the rib interspaces, and the nodule was similarly placed.
In the present study, in contrast, we used a larger number of images with different anatomic and quantum mottle backgrounds. We also used tissue-equivalent lesions of various sizes, which were placed at random locations in a highly controlled observer performance experiment. The aims were to establish the magnitude of difference in the detection thresholds for images with quantum noise and images with anatomic and quantum noise, at a clinically relevant level of quantum noise (exposure), and to compare the observer results with the theoretic expectations based on the noise characteristics of the images.
|
MATERIALS AND METHODS |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONAppendix 1References |
|---|
Acquisition of Anatomic Images
Twenty posteroanterior chest radiographs in 20 patients (10 men, 10 women; age range, 28–77 years; mean age, 41 years) were selected from our clinical database. According to the radiology reports, the patients were free from any acute pulmonary abnormalities. The images were acquired by using a digital storage phosphor radiographic system (FCR-9501-HQ; Fuji Medical Systems, Tokyo, Japan) and 35 x 43-cm standard-resolution phosphor screens (ST-Va; Fuji Medical Systems). A 115-kVp x-ray beam from a high-frequency generator (1050 HF; Acoma, Wheeling, Ill) and a 12:1, 59-lines-per-centimeter antiscatter grid (Gilardoni, Milan, Italy) were used for all acquisitions. All exposures were made by using an automatic exposure control that was calibrated to operate the system with a 22.7-mR entrance exposure for a standard 22-cm-thick Lucite slab. The phosphor screens were processed with the automatic Exposure Data Recognizer, or EDR, mode for the "Chest, General" protocol of the system.
The 10-bit, 4k (3,520 x 4,280) digital image data stored on a clinical workstation (HIC-654, Fuji Medical Systems) were transferred as 2k (1,760 x 2,140), 10 bits in logarithmic scale, to a research work-station (Sparc 2; Sun Microsystems, Palo Alto, Calif) through a small computer system interface, or SCSI, connection (DASM-FDLR, Analogic, Peabody, Mass). From each radiograph, three 400 x 400-pixel (8 x 8-cm) images were extracted at random locations within three general zones: upper left lung, upper right lung, and lower right lung. Thus, a total of 60 anatomic images were generated.
Acquisition of Quantum Mottle Images
Four storage phosphor radiographs of a 7.62-cm-thick Lucite slab placed at the detector were acquired by using the same x-ray tube, generator, antiscatter grid, and kVp as those used for the anatomic images. The semiautomatic Exposure Data Recognizer mode for the "Test, Contrast" protocol of the system with a fixed latitude of 2.0 was used for all four uniformly exposed radiographs. The acquired radiographs were transferred to our research workstation in a manner similar to that stated previously. Fifteen 400 x 400-pixel (8 x 8-cm) images were extracted from each radiograph, which generated a total of 60 quantum mottle images.
It was important to ensure that the level of quantum noise on the quantum mottle images was the same as that of the anatomic images. To do that, the average screen exposure for each anatomic image was determined from the average pixel value within a 128 x 128-pixel central area of the image by using the following equation:
Figure 1 illustrates the distribution of the average exposure values that were calculated for the anatomic images. The results show a broad distribution (mean, 1.02 mR; SD, 0.4 mR). In accordance with these results, the quantum mottle images were acquired with a screen exposure of 0.8 mR; this value was slightly lower than the mean value given previously. The quantum mottle images were acquired at 0.8 mR prior to the full verification of the calculated results. However, after verification, the level of exposure for the quantum mottle images was found to be sufficiently within the wide range of exposure levels on the anatomic images.
|
r 0.6D; C is the peak contrast of the nodule (C = 
E/E0 = ln[E0/E]); and D is the diameter of the nodule at the imaging plane, which is specified as the full-width-at-fifth-maximum of the profile (Fig 2).
|
n is the density of the nodule (assumed to be 1.0 g/mL); l is the density of the lung (assumed to be 0.3 g/mL); R is the ratio of the scatter to the primary signal intensity; and M is the contrast produced per unit thickness of nodular material, when no scattered radiation is present. For this study, R was assumed to be 0.68 on the basis of our measurements in a chest phantom. M was assumed to be 0.0235 mm-1 on the basis of findings from our prior work (table 3 from reference 20). The values for peak contrast–diameter products (CDs) for the nodules were chosen to produce a range of observer responses from "probably not present" to "probably present," which bracketed the detection thresholds for each type of background. This was accomplished by performing two pilot studies in which the appearances of nodules of varying sizes and contrast were examined against anatomic and quantum mottle backgrounds by two experienced radiologists (including W.R.E.). From the results of the pilot studies, peak contrast and diameter for the actual study were chosen to bracket the detection thresholds. The images used in the pilot studies were not used in the actual study. For each type of background, three nodular diameters were identified, with three distinct CDs. Table 1 shows the CD values of the nodules used in the observer performance experiments.
|
For this study, we constructed and used a Hurter and Driffield curve similar to that of a conventional screen-film system for chest radiography (Ortho-C/Lanex; Eastman Kodak, Rochester, NY), as illustrated in Figure 3. The 
values for this curve within the optical density range applicable to this study were similar to those of currently used screen-film systems (eg, InSight, Eastman Kodak). For each image, the average pixel value within a 128 x 128-pixel central area was used as the mean value for the main gradient component. Thus, the average level of contrast and optical density at the center of all images where nodules were superimposed were constant.
|
|
|
At the beginning of each session, the reading protocol was explained to the observer: The observer task was to discern whether a nodule was present at the center of the image and to indicate his confidence level by using a five-point grading scale (Table 2). Before the study began, the observer viewed two graded example hard-copy films; each contained six images of a background with six superimposed simulated nodules with increasing values. This was followed with the reading of two practice hard-copy films that were similar to those used for the actual study. The observer was subsequently provided with the expected answers so he could assess the visibility level of the nodules and adjust the grading. The observer was told that there was a 75% probability that a nodule was present on each image.
|
Data Analysis
A standard receiver operating characteristic analysis (22–25) was used to process the results of the observer performance experiments. For each observer, the scores ascribed to the images that depicted nodules with a particular value and the scores ascribed to the control images that depicted no nodules were counted in separate data sets. These data were then processed by using ROCFIT (June 1993 version), a program developed by C. E. Metz of the University of Chicago, Ill, to estimate the maximum likelihood of a binomial receiver operating characteristic curve and its associated parameters from a set of categorical rating-scale data (26,27). The area under the receiver operating characteristic curve (Az) was obtained for each observer as a function of the CD of the nodule for each type of background. The data were averaged across observers. The accuracy of each averaged Az value was assessed by estimating the standard error (SE), as described by Swets and Pickett (25), by using the following equation:
To relate the results of the observer performance experiments to the noise characteristics on the images, Az values were used to estimate the signal-to-noise ratios (SNRs) for the nodules by using the following equation (14,28):
The experimentally observed SNRs were then compared with the SNRs predicted with two observer models, the Hotelling model (which was used to perform the detection task by cross-correlating the image data with a template at the location of the nodule, taking into account the background variations), and a non-prewhitening model with a visual-response function (29) (which was used to perform the detection task by cross-correlating the image data with a modified version of the signal intensity as a template, not taking into account the background variation). The details of these calculations are provided in the Appendix.
|
RESULTS |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONAppendix 1References |
|---|
|
|
|
Figure 6 illustrates the measured SNR as a function of the calculated SNR for anatomic and quantum mottle images from the non-prewhitening and the Hotelling observer models. The two observer models yielded similar results with both backgrounds. However, in both cases, model observer values were higher than those inferred by the Az values. The observer results for both backgrounds were in general agreement with those obtained with the two observer models; the results did not indicate which model was best.
|
|
|
DISCUSSION |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONAppendix 1References |
|---|
In this study, an attempt was made to quantitatively compare the relative influence of quantum and anatomic noise in the detection of low-contrast subtle lung nodules on chest radiographs. The results suggest that this diagnostic task is not quantum-limited. The detection of lung nodules is affected much more by the anatomic structured pattern of the thorax than by quantum noise.
As shown in Figure 5, for both types of backgrounds, the Az value increased with the CD of the nodule. The observed behavior can be explained in the context of signal detection theory and in terms of the underlying probability distribution functions for noise and for signal and noise (25,30). If these distribution functions are assumed to be Gaussian, the Az value is expected to increase with signal as the integral of a Gaussian distribution function (14), which creates a sigmoidal curve that becomes saturated at an Az of 1.0. The rate of the increase is dependent on the width of the Gaussian distribution function; a narrow width distribution function leads to a sharp saturation, and vice versa.
In our results, the uniform increase in the Az value with the CD of the nodule for both quantum mottle images and anatomic images seems to fit this description. For anatomic images, however, the slope was about half that of the quantum mottle images, which suggests a wider underlying probability distribution function for the backgrounds of anatomic patterns.
In this study, the position and characteristics of the target were clearly explained to the observers; thus, the observer experiments were of the type known as signal-known-exactly (31,32). In such studies, the detection performance is expected to be better than that of clinical studies in which the observers must find nodules at unknown locations. The use of a signal-known-exactly protocol was based on two rationales: (a) The statistical characteristics of signal-known-exactly protocols are well understood, and (b) the goal of the study was not to investigate the detection in absolute terms, but rather, to understand the relative differences associated with the backgrounds of quantum noise and quantum and anatomic noise under ideal circumstances.
To demonstrate the influence of knowing the possible location of a nodule, a comparison can be made between the findings in this study and those of an earlier study, in which observers were to detect similar kinds of simulated nodules that were randomly positioned on full-chest radiographs (19). The investigators found that a CD of at least 0.8 mm (9-mm diameter for a tissue-equivalent lesion) was necessary if a nodule was to be detected, whereas investigators in this study found a corresponding value (for Az = 0.8) of about 0.2 mm (4.55-mm diameter). This comparison clearly demonstrates the importance of localization in the detection of lung nodules.
The simulated nodular patterns used in this study were circular. Findings from studies on the etiologic and computed tomographic characteristics of lung nodules (33,34) showed a great variability in overall shape, outline irregularity, and in-homogeneity of these nodules, especially in cases of primary carcinoma. The perceived limitation of the nodular shapes used in our study, however, proved to be insubstantial for anatomic images. The projected circular images of simulated nodules over the anatomic backgrounds overlapped with the structured pattern of the backgrounds, and the conjoined images of the nodules had irregularities and spiculations that were very similar to those of real lung nodules.
In a previous study (19), we demonstrated that experienced observers were not able to distinguish real and simulated nodules of the type used in this study. It should be noted, however, that the irregularity in the shape of a nodule that is placed on a quantum mottle background may affect its detectability. If irregularity reduces the detectability of a nodule against such a background, this may suggest an overestimation of the influence of anatomic noise in our study; the magnitude remains to be determined in further research.
The results of this study suggest that anatomic noise associated with the background lung structure strongly limits the detection of subtle nodules in the lung. One of the implications of this finding is the potential for a reduction in the dose to the patient; since the relative quantum noise is inversely related to the square root of the dose to the patient and since the detection of lung nodules is not limited by quantum noise, it may be concluded that the exposure to the patient at chest examination can be reduced without a marked loss in the detectability of lung nodules. It should be noted, however, that this may adversely affect the detection of mediastinal lesions or the presentation of other features, such as interstitial lung disease; these tasks were not addressed in this study.
|
|
|
Appendix 1 |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONAppendix 1References |
|---|
The non-prewhitening SNR, or SNRnpw, was calculated by using the following equation:
For the signal term S(f), the contrast profile of the nodules was approximated by means of a two-dimensional Gaussian function that had a full-width-at-fifth-maximum diameter equivalent to that of the actual nodule profiles, as follows:
is a scaling factor equal to 0.148. Figure 2 illustrates the Gaussian approximation of the nodule profile. H(f) and W(f) were measured experimentally. H(f) was measured by using a high-precision–edge testing device. The details of this measurement method are fully disclosed in a previous article (35). W(f) was assessed on the basis of Fourier analysis of the image data by using a method developed previously (21,36). Each image was divided into a 2 x 2 matrix with 128 x 128 subarrays (ie, four 2.56 x 2.56-cm blocks). Very-low-frequency trends within each subarray were removed by subtracting a second-order polynomial fit from the data.
The image data were then scaled to relative values and the power spectrum was computed within each subarray by using a two-dimensional fast Fourier transformational algorithm and a Hamming spectral estimation window. The estimates from all four subarrays were averaged to obtain the power spectrum for each image, and the data in each radial differential circle were averaged to obtain the radial power spectrum for all the images. Figure A1 shows the estimated noise power spectra. The spectral estimates for different images of each type were similar. The data for all 60 anatomic and all 60 quantum mottle images were averaged for the observer model calculations.
|
Acknowledgments |
|---|
|
Footnotes |
|---|
Author contributions: Guarantors of integrity of entire study, E.S., M.J.F., W.R.E.; study concepts and design, E.S., M.J.F.; definition of intellectual content, E.S., M.J.F.; literature research, E.S., W.R.E.; experimental studies, E.S.; data acquisition and analysis, E.S.; statistical analysis, E.S.; manuscript preparation, E.S., W.R.E.; manuscript editing and review, E.S., W.R.E., M.J.F.
|
References |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONAppendix 1References |
|---|
This article has been cited by other articles:
|
|
 |
|
  J. Vikgren, S. Zachrisson, A. Svalkvist, A. A. Johnsson, M. Boijsen, A. Flinck, S. Kheddache, and M. Bath Comparison of Chest Tomosynthesis and Chest Radiography for Detection of Pulmonary Nodules: Human Observer Study of Clinical Cases Radiology, December 1, 2008; 249(3): 1034 - 1041. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 M. J. Tapiovaara Review of relationships between physical measurements and user evaluation of image quality Radiat Prot Dosimetry, March 1, 2008; 129(1-3): 244 - 248. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 E. Samei, S. A. Stebbins, J. T. Dobbins III, and J. Y. Lo Multiprojection Correlation Imaging for Improved Detection of Pulmonary Nodules Am. J. Roentgenol., May 1, 2007; 188(5): 1239 - 1245. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
E. Samei, R. S. Saunders Jr, J. A. Baker, and D. M. Delong Digital Mammography: Effects of Reduced Radiation Dose on Diagnostic Performance Radiology, May 1, 2007; 243(2): 396 - 404. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
H. P. McAdams, E. Samei, J. Dobbins III, G. D. Tourassi, and C. E. Ravin Recent Advances in Chest Radiography Radiology, December 1, 2006; 241(3): 663 - 683. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
L. J. M. Kroft, W. J. H. Veldkamp, B. J. A. Mertens, J. P. A. van Delft, and J. Geleijns Detection of simulated nodules on clinical radiographs: dose reduction at digital posteroanterior chest radiography. Radiology, November 1, 2006; 241(2): 392 - 398. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
C. L. Hoe, E. Samei, D. P. Frush, and D. M. Delong Simulation of Liver Lesions for Pediatric CT Radiology, December 21, 2005; (2005) 2381050477. [Abstract] [Full Text]
|
|
|
|
|
|
M. Bath, M. Hakansson, S. Borjesson, S. Kheddache, A. Grahn, M. Ruschin, A. Tingberg, S. Mattsson, and L. G. Mansson Nodule detection in digital chest radiography: introduction to the RADIUS chest trial Radiat Prot Dosimetry, May 17, 2005; 114(1-3): 85 - 91. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
M. Hakansson, M. Bath, S. Borjesson, S. Kheddache, A. A. Johnsson, and L. G. Mansson Nodule detection in digital chest radiography: effect of system noise Radiat Prot Dosimetry, May 17, 2005; 114(1-3): 97 - 101. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
M. Bath, M. Hakansson, S. Borjesson, S. Kheddache, A. Grahn, F. O. Bochud, F. R. Verdun, and L. G. Mansson Nodule detection in digital chest radiography: part of image background acting as pure noise Radiat Prot Dosimetry, May 17, 2005; 114(1-3): 102 - 108. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
E. Samei, J. T. Dobbins III, J. Y. Lo, and M. P. Tornai A framework for optimising the radiographic technique in digital X-ray imaging Radiat Prot Dosimetry, May 17, 2005; 114(1-3): 220 - 229. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
M. Bath, M. Hakansson, J. Hansson, and L. G. Mansson A conceptual optimisation strategy for radiography in a digital environment Radiat Prot Dosimetry, May 17, 2005; 114(1-3): 230 - 235. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
M. Bath, M. Hakansson, A. Tingberg, and L. G. Mansson Method of simulating dose reduction for digital radiographic systems Radiat Prot Dosimetry, May 17, 2005; 114(1-3): 253 - 259. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
J. Hansson, M. Bath, M. Hakansson, H. Grundin, E. Bjurklint, P. Orvestad, A. Kjellstrom, H. Bostrom, M. Jonsson, K. Jonsson, et al. An optimisation strategy in a digital environment applied to neonatal chest imaging Radiat Prot Dosimetry, May 17, 2005; 114(1-3): 278 - 285. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
L. G. Mansson, M. Bath, and S. Mattsson Priorities in optimisation of medical X-ray imaging--a contribution to the debate Radiat Prot Dosimetry, May 17, 2005; 114(1-3): 298 - 302. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 D G W Onnasch, A Schemm, and H-H Kramer Optimization of radiographic parameters for paediatric cardiac angiography Br. J. Radiol., June 1, 2004; 77(918): 479 - 487. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 S. R. Turner, E. Samei, B. S. Hertzberg, D. M. DeLong, R. Vargas-Voracek, A. Singer, C. H. Maynor, and M. A. Kliewer Sonography of Fetal Choroid Plexus Cysts: Detection Depends on Cyst Size and Gestational Age J. Ultrasound Med., November 1, 2003; 22(11): 1219 - 1227. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
E. Samei, M. J. Flynn, E. Peterson, and W. R. Eyler Subtle Lung Nodules: Influence of Local Anatomic Variations on Detection Radiology, July 1, 2003; 228(1): 76 - 84. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
J. T. Dobbins III, E. Samei, H. G. Chotas, R. J. Warp, A. H. Baydush, C. E. Floyd Jr, and C. E. Ravin Chest Radiography: Optimization of X-ray Spectrum for Cesium Iodide-Amorphous Silicon Flat-Panel Detector Radiology, January 1, 2003; 226(1): 221 - 230. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
) was generated by using two Gaussian gradient functions (dashed lines). The difference in the screen-film and model transformations is due to the baseline optical density. The gradient functions can be manipulated easily to construct various transformational curves.
