Quantitative assessment of computed radiography ... - Semantic Scholar

TB, RJ, PMB/206520, 3/02/2006 INSTITUTE OF PHYSICS PUBLISHING

PHYSICS IN MEDICINE AND BIOLOGY

Phys. Med. Biol. 51 (2006) 1–17

UNCORRECTED PROOF

Quantitative assessment of computed radiography quality control parameters O Rampado, P Isoardi and R Ropolo Struttura Complessa Fisica Sanitaria, Azienda Ospedaliera San Giovanni Battista, Corso Bramante 88, 10126 Torino E-mail: [email protected]

Received 24 August 2005, in final form 12 January 2006 Published DD MMM 2006 Online at stacks.iop.org/PMB/51/1 Abstract Quality controls for testing the performance of computed radiography (CR) systems have been recommended by manufacturers and medical physicists’ organizations. The purpose of this work was to develop a set of image processing tools for quantitative assessment of computed radiography quality control parameters. Automatic image analysis consisted in detecting phantom details, defining regions of interest and acquiring measurements. The tested performance characteristics included dark noise, uniformity, exposure calibration, linearity, low-contrast and spatial resolution, spatial accuracy, laser beam function and erasure thoroughness. CR devices from two major manufacturers were evaluated. We investigated several approaches to quantify the detector response uniformity. We developed methods to characterize the spatial accuracy and resolution properties across the entire image area, based on the Fourier analysis of the image of a fine wire mesh. The implemented methods were sensitive to local blurring and allowed to detect a local distortion of 4% or greater in any part of an imaging plate. The obtained results showed that the developed image processing tools allow us to implement a quality control program for CR with short processing time and with absence of subjectivity in the evaluation of the parameters.

1. Introduction Computed radiography (CR) is at this moment the most common digital radiography modality in radiology departments, in place of conventional screen film systems. There are important differences in the quality control approach between the traditional film screen radiology and CR. In conventional radiography, the radiation detector and the display device of the radiograph are the same object. With the CR technology, the detector is a photostimulable phosphor screen, usually indicated as an imaging plate (IP). Digital image data are extracted 0031-9155/06/000001+17$30.00 © 2006 IOP Publishing Ltd Printed in the UK

1

2

O Rampado et al Table 1. CR systems and IP evaluated in this study. Manufacturer

CR devices

Phosphor screen

Kodak Philips (Fuji)

CR 800, 900 & 950 AC-PCR 5000

GP 25 & HR ST-VN

from the exposed IP and used to provide the image on a monitor or film. Therefore, CR and generally digital radiography allow us to distinguish between the quality analysis of detectors and the analysis of the display system. A quality control program to inspect CR performance includes a set of exposures with test objects and an analysis of the obtained digital image data. Several manufacturers established guidelines for acceptance testing (Kodak 2001, AGFA 1999), but there are no industry standards for specifying the performance of these devices and this causes a lack of uniformity in measurement procedures among different manufacturers. Preliminary guidelines of the Task Group No. 10 of the American Association of Physicists in Medicine (Seibert et al, Samei et al 2001) provide a first comprehensive standardized testing protocol and recommend uniform quantitative criteria for the satisfactory performance of CR devices. A limitation of these guidelines lies on the fact that many of the evaluation procedures are not fully quantitative or can be influenced by the subjectivity of the examiner, such as the evaluations of limiting resolution and noise performance. In a typical radiology department with a CR system there may be hundreds of IPs. Some quality tests have to be made on every IP and this may cause a considerable increase in the quality control workload for medical imaging physicists. The aim of this work was to develop a set of software processing tools in order to make a complete quantitative assessment of computed radiography quality control parameters. These tools should reduce the time needed to perform the CR quality tests and avoid any subjective influence in quality parameters evaluation. 2. Materials and methods As listed in table 1, four CR devices (one for each model) in use at two different radiology departments from two different CR manufacturers were evaluated. Test images in DICOM format were sent from the CR devices to our workstation (HP d530 2.8 GHz Pentium IV) through the local network. Table 2 shows the definitions of the quantities of interest for CR characteristics evaluated in this work. The image processing tools were developed as plugins (Java classes) of the public domain Java program ImageJ (Wayne Rusband, National Institute of Health, USA, http://rsb.info.nih.gov/ij/). The freeware java editor Jcreator (Xinox Software, Delft, Netherlands) was used to edit the plugins, and they were compiled using ImageJ on a windows 98 platform but, because of Java properties, they are also usable on other platforms. All image processing procedures were automatically performed on groups of images, opening every image sequentially and writing an output table containing the measurements. Every plate was identified by the PLATE ID number extracted from the DICOM header. The performance characteristics indicated in table 2 correspond to those defined in the summary presentation of AAPM TG10 tests proposed by Samei et al 2001, and we used the same exposure conditions. Differences from the TG10 tests only concerned the test image processing and the objective indicators. The dark noise and the uniformity tests were applied to all the screens, whereas the other tests were applied to one screen for each type and size. The dosimeter used to verify exposure was a solid state detector Unfors Mult-O-Meter (Unfors Instruments, Billdal, Sweden).

Quantitative assessment of computed radiography quality control parameters

3

Table 2. Objective indicators of CR performance characteristics. Performance Characteristic

Objective Indicators

Dark noise

Average signal and its standard deviation within 80% of the image areaa Exposure index valueb Signal standard deviation within 80% of the image areaa Maximum difference between quadrants average pixel valuesb Differential and integral uniformity evaluated on a 1 cm × 1 cm ROI matrixc Exposure indicator response normalized to a 1 mR entrance exposurea Slope of the system response (expressed in terms of logarithm of exposure) versus logarithm of actual exposurea Correlation coefficient of the linear fit to logarithm of pixel value standard deviation versus logarithm of actual exposurea Number of phantom details with a contrast noise ratio above a specified thresholdc Modulation transfer function valuesa Difference between measured and actual distances in the orthogonal directionsa Jitter dimension in pixelsa Differences between subarrays Fourier spectrum peak amplitudec Differences between subarrays Fourier spectrum peak positionc Average signal and standard deviation within 80% of the reread unexposed imagea Contrast-noise-ratio of the supposed ghost imagec

Uniformity

Exposure calibration Linearity and autoranging Noise and low-contrast resolution Limiting resolution Spatial accuracy Laser beam function Resolution Uniformity Spatial accuracy uniformity Erasure thoroughness

a

Definitions according to AAPM TG10 (Samei et al 2001). Definitions according to Kodak Guidelines (Kodak 2001). c Definitions introduced in this study. b

2.1. Dark noise, exposure calibration and linearity CR systems have a wide latitude response and the ability of these devices to accommodate a wide exposure range provides a relatively consistent appearance, even in the case of under or over exposed images. In order to avoid image repetitions, technologists may tend towards overexposure (Willis and Slovis 2005), so a dose feedback is absolutely necessary for dose management in CR. An indicator of the average incident exposure on the imaging plate is provided by CR systems, in order to verify proper radiological techniques. Each manufacturer has a specific method for providing this exposure indicator, but the relationship between its value and the exposure value is known. For Kodak IP GP25, the exposure index is proportional to the logarithm of the exposure E in mR (Goldman 2004): EIGP = 1000 log(E) + 2000

(1)

and the same relationship exists between the pixel value of general purpose IP and the exposure. For Kodak IP HR the relationship is the (1), but the constant is 1700 rather than 2000. For the Fuji IP used in this study, the exposure indicator is called sensitivity (S) and it is inversely proportional to the exposure E in mR:S = 200/E (Fuji 1993). The relationship between the pixel value and the exposure depends on the exposure data recognizer (EDR) mode and processing menu selections. In the Semi EDR this relationship is (Tucker and Rezentes 1997): P V = 512 − (1024/L) × [log E + log(S/200)]

(2)

where L is the latitude set. Test images obtained from the reading of unexposed IPs (dark noise evaluation) or for those exposed at different doses (system efficiency and linearity) are used to check whether the exposure indicator value is provided correctly. Its numerical value is present in the DICOM header of the image file, so no post processing is needed to perform these tests. The developed plugin queried the DICOM header by string manipulation instructions. Groups and elements

4

O Rampado et al (a)

(b)

(c)

Figure 1. Schematic view of the three methods used to define uniformity parameters, with indication of the ROI and of the uniformity indices.

for the location of the exposure indicator in the DICOM header were ‘0018,1405 Relative x-ray Exposure:’ for Kodak and ‘0018,6000 Sensitivity:’ for Fuji. An indication of the exposure level may also be obtained by calculating the average pixel value in a central region of interest (ROI) on the 80% of the plate area (PV80%). The dark noise evaluation is only affected by the CR reader and plate characteristics, because no exposure is involved in the test. The exposure indicator value obtained from the reading of the unexposed plate should not exceed a specified threshold: PV80% > 744 for Fuji, EIGP < 80 and EIHR < 380 for Kodak (Samei et al 2001). The exposure calibration and linearity are strongly dependent on the exposure conditions (additional filtration, distance, etc), and therefore it is very important to check the constancy of these parameters, in order to perform the test correctly. The linearity of the exposure indicator value is tested by exposing the same IP to approximately 0.1, 1 and 10 mR (1 mR = 2.58 × 10−7 C kg−1 ) entrance exposures in a sequence of three exposure-reading cycles. 2.2. Uniformity Uniformity of response is a fundamental parameter for detectors in all medical imaging fields. A uniform exposure should result in a uniform response of the CR system. Many definitions (Kodak 2001, Samei et al 2001, Masden 1997, AAPM 2005) have been proposed to establish uniformity indicators. In this study we used three different approaches (figure 1): (1) After the image separation into four discrete quadrants, a uniformity index (Uquad) was evaluated as the difference between the average values (P V i , i = 1, . . . 4) of the two quadrants with the highest (MAX(P V i )) and lowest (MIN(P V i )) average pixel values (Kodak 2001). Only half of the actual quadrant area was used in the calculation of the average pixel value in each quadrant (figure 1(a)). (2) Upvsd: standard deviation of pixel value within 80% of the image area (Samei et al 2001). (figure 1(b)). (3) Uint and Udif: a matrix of adjacent square ROI (1 cm × 1 cm) was used to analyse the image area. Uniformity indicators used in this case were similar to those used in nuclear medicine (Masden 1997) or for the quality assurance of monitors (AAPM 2005). Differential uniformity here refers to the maximum percentage difference between adjacent ROI average pixel values (P V i+1 and P V i in the formula shown in figure 1(c)). Integral uniformity refers to the percentage difference between the maximum (MAX(P V i )) and the minimum (MIN(P V i ) ROI average pixel value. Being the pixel value defined differently by the two systems under test, in order to use these indices for both systems, the pixel value was reassigned in terms of exposure before the analysis, using inverse formulations of the relationship between pixel value and exposure: P Vexp = 10(P V −2000)/1000

(3)


5

for Kodak GP (for Kodak IP HR the relationship is the same, but the constant is 1700 rather than 2000) and P Vexp = 10(512−P V )·(L/1024) + 200/S

(4)

for Fuji. As a consequence, Uint and Udif are expressed as percentage values and Upvsd and Uquad are expressed in mR units, as the exposure value calculated for every pixel. The contribution of the heel effect to the uniformity index values was evaluated comparing several images obtained with two sequential half-exposures between which the orientation of the cassette was reversed, and with two sequential half-exposures without change of orientation. The percentage difference between the images (evaluated using the same ROI of the uniformity analysis) resulted to be less than 2%. Uniform exposures also provide a stringent test for many common image artefacts that can occur in CR systems. Both AAPM and manufacturers guidelines state that all images should be examined for banding, black or white spots and streaks. In computed radiography, as well as in other imaging techniques, image artefacts are a source of ‘noise’ that can degrade the diagnostic quality. To investigate these artefacts we used an algorithm for automatic detection. In the adjacent ROI analysis, this algorithm detected those pixel with values that differed from the ROI mean by more than 3%, being about 1% the relative standard deviation. The image was then segmented into three levels, assigning white, grey or black colour to pixels below, within or above the range of the mean value ±3%. A median 2 × 2 filter was applied to this segmented image in order to remove black or white isolated pixels, which are related to noise peaks rather than to wide defects, and non-uniform regions were thus highlighted. 2.3. Noise and low-contrast resolution Low-contrast resolution characterizes detectability of a low-contrast object, and is influenced by several factors, including the object size, contrast between the object and the background, image noise and the system’s MTF. Image noise is primarily determined by the dose setting of the x-ray tube, the detector efficiency and the reconstruction algorithm. In this study the noise and low-contrast resolution properties of the CR system were tested by acquiring three images of a low-contrast phantom, using 0.1, 1 and 10 mR, at 70 kVp beam with 1 mm of Cu filtration. Test images were acquired turning off any vendor specific nonlinear image processing. The phantom used was the TOR [CDR] (Faxil, Leeds, UK) that consists of a flat Perspex disc carrying four types of test pattern. The so-called low-contrast sensitivity test pattern, comprising 17 circular details (11 mm diameter) with x-ray contrast values specified by vendor in the range 0.075–0.002, was used in this study. The noise is quantified by the standard deviation of the pixel value in a fixed small region of the image (PVSD). The logarithm of noise is linearly dependent on the logarithm of exposure E (Christodoulou et al 2000): log(PVSD) = a + b log E

(5)

with a correlation coefficient >0.95. The evaluation of the minimum discernible contrast to characterize the low-contrast resolution is generally performed in a subjective fashion on a test phantom with a lowcontrast resolution pattern. In order to make a more objective analysis, we developed a specific software tool, with an algorithm that automatically locates the contrast objects and performs measurements. The automatic recognition of the phantom details can be performed in different ways. General registration methods to match the test image (independently of the phantom used) with a pre-processed template have been proposed (Kwan et al 2003), but the computational time necessary to perform them is quite long (several hours for each

6

O Rampado et al

Figure 2. Regions of interest used on the TOR phantom image to calculate the contrast-to-noise ratio of the low-contrast objects.

registration). We developed a specific algorithm that recognizes geometrical reference lines and points present in the phantom, performed in the following steps: – the coordinates of the centre of the phantom were located following the application of an edge detection method; – the relative rotation of the phantom image was defined by the profile plot analysis of a properly positioned circular line; – for every contrast object, measurements of the mean and standard deviation of pixel values were performed in an internal circular ROI (7 mm diameter) and in a surrounding ROI (14 mm internal diameter and 17 mm external diameter), as shown in figure 2; – the contrast-to-noise ratio was calculated for every object using the following relationship: P Vin − P Vout . CNR = 2 σin2 + σout

(6)

The number of details with a CNR value above a specified threshold (for each exposure level) could be used for constancy quality control. 2.4. Spatial resolution Quality control guidelines suggest to qualitatively establish the limiting spatial resolution with an image of a line pair pattern device. The resolution of a CR system should be more objectively evaluated by measuring the frequency-dependent modulation transfer function (MTF). Several publications discuss details of different methods to perform this measure (Fujita et al 1992, Bradford et al 1999, Samei and Reimann 1997, Stierstorfer et al 1999). In October 2003, the International Electrotechnical Commission (IEC) published an international standard (IEC 2003) containing specifications regarding measurements of the detector quantum efficiency (DQE) and of the MTF that should be used when manufacturers describe the imaging


7

performance of an x-ray imaging device. The IEC standard provides many details on the object test and x-ray source, but it could be difficult to apply this exact procedure in the context of a constancy control with limitations in time and instrumentation: we think that one of the other methods proposed and described could be implemented for this purpose. In the software we developed, methods based on the image of a slit camera, an edge and a bar pattern were implemented following the instructions presented in the referenced literature (Bradford et al 1999, Samei and Reimann 1997, Stierstorfer et al 1999). A slit camera with slit width 10 µm, slit length 10 mm, tantalum as slit material and slit thickness 1.5 mm (Model 07-624, GAMMEX-RMI, Middleton, WI) and a H¨uttner test pattern (Type 18, Faxil, Leeds, UK) with 26 line pair groups (1–20 cycles/mm) were used to perform the measurements. The edge used to measure the edge spread function was the border of the H¨uttner test pattern, made of a 30-µm-thick lead foil embedded in glass. 2.5. Spatial accuracy and laser beam function Accurate imaging requires preserved geometric relationships. A metal rule or plate may be used to obtain a test image that allows a comparison between actual and measured dimensions. Automatic edge detection was used to make measurements in different regions of the image. The analysis of the metal object edge profile is also useful to verify the beam laser function. Object edges should be straight and continuous over the full IP length. Under or overshoot of the scan lines in light to dark transitions along the object edge indicates a timing error, or laser beam modulation problem and results in occasional incorrect definition of the object edge. This effect can be indicated with the term ‘jitter,, generally used for any distortion of a signal or image caused by poor synchronization. In order to make quantitative evaluation of the jitter present, the edge profile was analysed in the following way: – two rectangular region of interest were automatically defined with a large portion (90%) of the edges inside (figure 3(a)); – for any horizontal row of the ROI indicated as 1 in the figure and for any vertical row of ROI 2, the position of the edge was determined by analysing the profile of pixel values, as the interpolated position of the midpoint between the left side and the right side of the profile; – a linear fit was calculated for the two series of edge position values, thus obtaining positions of an ideal edge with no jitter (figure 3(b)); – the differences between the actual edge positions and the fitted position values were plotted to check the presence of jitter above one pixel. The number of these jitters was the quality parameter that we used. 2.6. Resolution and spatial accuracy uniformity The previously described spatial accuracy and resolution tests only provide local information, whereas Samei et al 2001 suggest to characterize these properties across the entire image area, using a screen–film contact test tool (wire mesh test). In this way, it is possible to verify the resolution and spatial accuracy uniformity over the total area of the phosphor receptor, checking the sharpness and distortion of the obtained image over the whole field of view. The object used to perform this test was a mammographic film/screen contact test tool Gammex 157A (GAMMEX-RMI, Middleton, WI), a wire mesh encased in plastic with dimensions 25.5 × 31.5 cm2. The distance between the centre of the wires was 0.63 mm, with a frequency of 1.58 mm−1.

8

O Rampado et al

(a)

(b)

Figure 3. (a) Test image obtained exposing a metal plate and indication of ROIs for jitter and laser beam function analysis. (b) Plot of the edge position values and linear fit that represent the ideal edge with no jitter.

The degree of regularity and resolution governing a uniform pattern in digital image is most conveniently analysed not in the spatial domain, but in the frequency domain. In order to establish a quantitative criterion for this test we made a Fourier analysis of the test image. Within the acquired digital image array, a set of subarrays (size 128 × 128 pixels) arranged contiguously on the entire image area was identified. Prior to Fourier transformation, the subarray data were modified by multiplying the subarray values by a Hamming window function of the type commonly used in spectral estimation (Pratt 1991), in order to avoid aliasing from discontinuities at the edge of the subarray. The two-dimensional Fourier transform of each subarray was then computed. Power spectra in the fast-scan and in the slow-scan directions were estimated by averaging the central row or column and ±1 rows or columns within each transformed subarray. These power spectra resulted to have a main peak at the pattern frequency and a second peak at the double of the pattern frequency (second harmonic). Position and width at half amplitude of the main peak can be considered as indicators of absence of distortion. The position of the peak was determined with an accuracy of ±0.01 mm−1 using a Gaussian fit of the spectrum values near the maximum. The relative amplitude of the two peaks can be considered as an indicator of sharpness. This relative amplitude Ap was defined as follows: +1 (P S[fc + i] + P S[2fc + i]) Ap = i=−1 (7) 128 i=2 P S[i]


9

where PS[i] are the power spectrum values and fc is the position of the characteristic frequency of the pattern. These indicators should have the same value for every subarray of the test image. In order to establish a range of performance levels of this test we used a set of reference test images. The resolution and spatial accuracy uniformity of a CR system was first tested by repeating the MTF analysis and geometrical measurements several times with slit camera and small metallic rules positioned in different parts of the imaging plate. This CR system was used to obtain a set of images of the wire mesh test. These images were modified with digital image processing tools provided by the ImageJ software to simulate problems that might occur during the acquisition process. Specifically, we applied a scaling factor to the test images in both scan directions, in order to simulate incorrect velocity in these two directions. In other cases, we applied Gaussian blurring filters with different kernels to simulate a loss of sharpness that might be experienced with a laser spot of increased cross-sectional diameter (Rowland 2002). 2.7. Erasure thoroughness The plate, if improperly or insufficiently erased, can potentially give rise to image artefacts. The test of the erasure capability is performed by exposing an erased IP (unused for 1 h before the test) at high exposure levels (50 mR) with a centrally placed high-contrast test object (a thick lead block), reading the plate and re-exposing the plate to a uniform incident exposure of about 1 mR. The re-exposed image should be free from ghost artefacts. The ghost signal is quantified in our software by the percentage difference between the average pixel value in the region previously occupied by the high-contrast object and in the surrounding area. 3. Results and discussion The image processing tools were used to analyse images obtained from quality controls performed on a total of 125 IPs (82 Kodak GP25, 4 Kodak HR and 39 Fuji ST-VN). Kodak IPs were tested with all the three Kodak CR readers and Fuji IPs with the unique Philps CR reader. The uniformity test and the dark noise test were applied to all the IPs, the other tests were performed on a single IP for each type and size of the screen and for each CR reader. The number of test images required for each single reader and IP type and size was 34, and the number of images processed in this study was 430, with an exposure and acquisition time of 8 h and a total processing time of 2 h. 3.1. Dark noise, exposure calibration and linearity The exposure indicator values, obtained on flat field images exposed at a given exposure level, were correct according to the tolerance specified in quality control protocols for the 90% of the IPs which we analysed. 13 IPs (8 Kodak and 5 Fuji) had exposure indicator values out of the tolerance. The 8 Kodak IPs had all the same size (24 × 30) and also the 5 Fuji IPs (18 × 24). The CR reader had a guided procedure to re-calibrate the exposure indicator, which involved the readout of an IP exposed to a known value (e.g. 20 mR for Kodak) for every IP size and type. We found that the calibration date of the IPs with the exposure indication out of tolerance was older (more than 3 years) than the calibration date for the other IPs (about 1.5 year). After a re-calibration of the system all IPs had correct exposure indicator values. Also, the linearity was good for all the IPs and CR systems, with a minimum correlation coefficient value of 0.9994 between the exposure calculated from exposure indicator and the

10

O Rampado et al

Figure 4. Box and Whisker plots of different uniformity indices for the IP with apparent defects (group Y) and without (group N).

actual one. We found dark noise values above the specified tolerance for the Kodak CR 900 reader only, with EI values ranging between —110 and 140, the specified tolerance being of 80. We discussed these data with the manufacturer technician and hence supposed that the problem might have been related to the calibration of the laser-photomultiplier chain, or to the infiltration of ambient light inside the CR reader. These hypothesis were then discarded because a re-calibration of the laser-photomultiplier chain and the variation of the room light (also trying the total darkness) did not change obtained EI values. No other differences in image quality indicators were found between this CR reader and the others. 3.2. Uniformity The uniformity analysis has two purposes: the evaluation of a global uniformity index and the detection and classification of artefacts. In order to study the information related to the different uniformity indices previously defined, the IPs were divided into two groups: IPs with artefacts (group Y) and IPs without artefacts (group N). 22 IPs (16 Kodak and 6 Fuji) were assigned to group Y, and 103 IPs (70 Kodak and 33 Fuji) to group N. This classification was made by using the algorithm previously described to highlight the presence of defects. According to Cesar et al 2001, the most frequent artefacts are white irregular lines, caused by cracks or scratches on the IP; white straight lines, caused by dirt on the light guide in the IP reader; wide bright areas, caused by damage to the IP protective layer. In our sample we found two IPs with small white irregular lines, 18 IPs with wide bright areas and two IPs with both these artefacts. The wide bright areas were often located near the corners of the IPs. The distribution of results from the four uniformity indexes for the two groups is shown in figure 4. A Student t-test was applied to assess the significance of the differences between the two groups, and we found p-values lower than 0.005 for all indices. This means that all indices were influenced by the presence of artefacts, but from figure 4 it is easy to see that sometimes the same index value was obtained for IPs with and without artefacts. As a consequence it was difficult to define a threshold value for these indices which could be used to automatically distinguish IPs with artefacts. Only for the Udif index the two distributions were almost separated, and a reference value could be 5% because all IPs with Udif index >5% belonged to the Y group. Nevertheless, a visual inspection of the uniformity test image should always be performed; in fact, we found, for example, that the two IPs with small irregular lines


11

Figure 5. CNR values for different details of the TOR [CDR] phantom and for three exposure levels.

had an Udif value lower than 5%. In our opinion, the segmentation of uniformity images with a repeatable criteria like the one used in this study allows us to perform the visual analysis in a more objective way. The other indices had a more consistent overlapping between the two data distributions, and therefore it was more difficult to define a threshold value related to the presence of artefacts. For Uquad and Upvsd, this overlapping was caused by the fact that often the artefacts were located totally or partially outside the ROI used for evaluation, or in some cases the spatial extension of the artefacts was small in respect to the ROI area. The evaluation of Uint or Uquad allows us to investigate the overall uniformity, which is influenced, for example, by the presence of a gradual variation between two opposite edges of the plate that is not perceived as an artefact (and probably does not affect the Udif index). 22 IPs were analysed before and after the periodic cleaning procedure (performed every month). We found that Upvsd is the best index to use when checking the cleanliness status of the IP, with an average difference in value for the same IP of 31% (range 25–41%). The other uniformity indices here considered were not strongly influenced by the difference in PV due to dirt on the IP because they were all based on ROI average values. All the IPs without artefacts had a Upvsd value lower than 0.5 mR; therefore a greater value could indicate that the cleaning procedure is not being performed with the correct periodicity or accuracy. 3.3. Noise and low-contrast resolution CNR values in the range of different details and at different exposure values are shown in figure 5. In order to assess the uncertainty in the CNR evaluation, new test images were produced five times at each of the three exposure levels, using the same CR reader, but with different IPs of the same type and size. The error bars in the figure indicate the standard deviations for CNR values; with CNR values above 0.4 relative standard deviations were

12

O Rampado et al

Figure 6. Presampled MTF of a Kodak 18 × 24 IP determined using the slit, the edge and the bar pattern methods.

below 20% for exposures of 0.1 mR, and below 10% for other exposure levels. We used this value of 0.4 as the threshold needed when checking the constancy of the number of details with a higher CNR, resulting in an average number of 6, 12 and 13 for the three exposure levels. These values should be evaluated during the acceptance test and should be considered as reference for constancy tests, with a tolerance of two visible details. With regard to the assessment of contrast detectability, an important issue is the post processing of the produced test image. According to the proposed quality control protocols, the low-contrast phantom image should be analysed once the window width and level have been optimized. This introduces another subjective factor in the test procedure. The histogram analysis on the CR reader modifies the pixel value according to the selected window width and window level. An advantage of using the CNR to test the low-contrast resolution is that adjustment of window and level does not affect CNR, because the average pixel values and standard deviations are modified in a linear manner, given that the vendor-specific, nonlinear image processing is disabled. Infact, a n-bit pixel value PV is modified by the application of an adjustment of window width (WW) and level (WL) according to the relationship (Passariello 2000): 2n 2n 2n · (P V − W L) + 2n−1 = · PV − · W L + 2n−1 = αP V + β. (8) P VW = WW WW WW As a consequence, the mean value and the standard variation of a ROI are modified in the following manner: P VW = αP V + β

σW = ασ.

(9)

These relationships may be used to verify that the CNR defined by (6) is not affected by the adjustment of the window width and level: P VW in − P VW out αP Vin + β − αP Vout − β α(P Vin − P Vout ) CNRW = = = = CNR. 2 2 σW2 in + σW2 out α 2 σin2 + α 2 σout α σin2 + σout (10) 3.4. Spatial resolution Figure 6 shows the measured presampled MTF in the fast-scan and in the slow-scan directions for a GP 18 × 24 plate using the slit, the edge and the bar pattern devices. The results obtained from the three methods were similar. Within the frequency range of 1–4 cycles/mm, the MTF


13

Table 3. Frequencies at 0.1 MTF for CR readers and IP types and sizes used in this study. Frequency (lp mm−1) at 0.1 MTF CR reader

IP type

Image size (cm2)

Fast scan

Slow scan

Kodak CR-800

GP GP HR GP HR GP GP HR GP HR GP GP HR GP HR ST-VN ST-VN ST-VN

35 × 43 24 × 30 24 × 30 18 × 24 18 × 24 35 × 43 24 × 30 24 × 30 18 × 24 18 × 24 35 × 43 24 × 30 24 × 30 18 × 24 18 × 24 35 × 43 24 × 30 18 × 24

3.11 3.71 4.55 3.79 4.95 3.09 3.71 4.57 3.78 4.92 3.04 3.68 4.49 3.76 4.90 3.21 3.18 3.43

4.1 4.31 5.50 4.29 5.39 4.25 4.19 5.43 4.26 5.33 4.19 4.13 5.42 4.20 5.31 3.25 3.22 4.02

Kodak CR-900

Kodak CR-950

Philips AC-PCR 5000

values obtained with the three methods were compared at the cycles/mm values of the bar pattern device (about every 0.25 cycles/mm), observing MTF differences less than 0.028, with a mean difference of 0.023. Values obtained from MTF measurements in terms of frequency at 0.1 MTF are shown in table 3. Generally, there were little differences in the MTF values measured on the three Kodak readers for each IP type and size, with maximum differences of 2% in the fast-scan direction and of 4% in the slow-scan direction. Similar to the Kodak plates, the Fuji MTF values corresponding to the slow-scan direction were higher than the fast-scan direction. However, there was a much larger difference between the slow-scan and fast-scan MTF values for the Kodak plates. The main goal of these MTF measurements was neither the evaluation nor the comparison of the performance of digital radiographic systems, but rather the quality control of the reproducibility of the spatial resolution measurement. As a consequence, test images were obtained without particular attention to the test device position and alignment. The angulation of the slit or the edge was in the range 5◦ –8◦ and was then measured by the software in the presampling process. Repetition of measurements on the same CR system gave results which were very similar with a maximum difference of 4%. We think that a tolerance of 10% (which is more than two times the maximum difference between repeated measurements) could thus be used in order to test the constancy of the spatial resolution, according to Seibert et al. 3.5. Spatial accuracy and laser beam function Measurements of spatial accuracy and laser beam function were performed using the four CR readers afore mentioned and one plate for each size and type. Differences between measured and actual distances in the orthogonal directions were less then 2% for all the imaging plates analysed. Figure 7 shows jitter values in pixels, evaluated as indicated in the material and methods section (difference between the actual edge positions and the fitted position values).

14

O Rampado et al

Figure 7. Differences between the actual edge positions and the fitted position values.

Figure 8. Peak position values of the power spectrums obtained from a wire mesh test image processed with several scaling factors. The two dotted lines define the range of the values obtained from the original image (mean value ±3 σ ).

Maximum differences were less than 0.5 pixels. All plates analysed had no jitter above 0.5 pixels. 3.6. Uniformity of resolution and uniformity of spatial accuracy In order to determine the sensitivity of the measurements for uniformity of resolution and for uniformity of spatial accuracy, we modified images of the wire mesh object using rescaling and blurring digital image processing tools and applied the analyses to the modified images. We investigated the effect of a scaling factor on the peak position of the power spectra. Theoretically, the peak position after the rescaling is simply obtained by the ratio of the old peak position and the scaling factor. For example, with a scaling factor of 0.5, the new peak position will be twice the value of the old peak position. Figure 8 shows peak position values obtained from the Fourier analysis of adjacent subarrays, after the application of different scaling factors. The error bars were defined as three times the standard deviation σ , because all the obtained values are comprised in the ±3 σ range. The two dotted lines define the range of the values obtained from the original image (1.58 ± 0.02). A local warping of ±4%, which is not visible with a simple visual inspection, results in values exceeding the indicated range, and is thus detectable with this analysis. We repeated the analysis with partial scaling of the images, and ROIs located in the re-scaled region show the peak value out of the established


15

Figure 9. Peaks’ relative amplitude Ap and MTF values obtained from images test (wire mesh and slit camera) after the application of Gaussian blurring filters with different kernels. The two dotted lines define the range of the values obtained from the original image (mean value ±3 σ ). The effect of the blurring filters on a detail of the slit camera image and of the wire mesh image is also shown.

range. Depending on the pixel dimension, the 128 dimension of the subarray corresponds to a square ROI of 0.6–2 cm. A gradual loss in resolution power was generated convolving the wire mesh test images with Gaussian blurring filters with several kernels of amplitude in the range 1–5 pixels. This should simulate the effect of an increase in the cross-sectional diameter of the laser beam, the intensity of which has a Gaussian profile (Rowland 2002). The same processing tool was applied to the image of a slit camera, in order to evaluate the effect on MTF values. Figure 9 shows the peaks’ relative amplitude Ap previously defined by equation (7), and MTF values at 1 and 2 lp mm−1 versus Gaussian blurring kernel. The error bars of Ap were defined as three times the standard deviation and also in this case all the values obtained are comprised in the ±3 σ range (corresponding to about 10% of Ap). A value of the peaks’ relative amplitude out of this range means that there is a loss of resolution power similar to that caused by the application of a blurring filter with kernel greater than 1. The loss in MTF values at 1 lp lp mm−1 and 2 lp mm−1 for a kernel of 1 was respectively −10% and −30%. The analysis was repeated with partial blurring of the image and examples of results are shown in table 4. We found that a good criterion to apply in order to automatically find resolution non-uniformity could be to verify that the relative amplitude of the peak values are comprised in the range Ap mean value ±0.07 (proposed tolerance range in the table). 3.7. Erasure thoroughness The percentage difference between the average pixel value in the region previously occupied by the high-contrasthigh-contrast object and the surrounding area was less then 0.9%, that is less then the typical standard deviation of 2% of the IP at this exposure level. The uniformity of

16

O Rampado et al Table 4. Examples of Ap values obtained from the Fourier analysis after the application of a Gaussian blurring filter with kernel = 2 on different portions of a wire mesh test image. Ap values

Percentage of image area with blurring (%)

Mean

Min

Max

Tolerance range

0 5 10 20 50 75

0.68 0.66 0.65 0.64 0.62 0.58

0.62 0.48 0.47 0.49 0.52 0.47

0.73 0.72 0.72 0.72 0.72 0.72

0.61–0.75 0.59–0.73 0.58–0.72 0.57–0.71 0.55–0.69 0.51–0.65

response of the analysed IP is not affected by a previous high exposure with high-contrasthighcontrast objects. 4. Conclusions We have developed a package of software tools for the implementation of a complete set of quality tests on CR systems. We applied it to CR systems of two manufacturers. The uniformity analysis with different approaches allowed us to understand more deeply the correlation between the presence of artefacts and the obtained parameter values. Upvsd and Uquad are good indices in quantifying an overall uniformity of the IP, but their values are not correlated with the presence of the most frequent artefacts. The Udif index may be used to automatically detect the presence of artefacts such as wide bright areas, but it is not sensitive to small irregular lines. The constancy of low-contrast resolution was verified using the contrastto-noise ratio values obtained from an automatic ROI analysis, avoiding any subjectivity in the evaluation procedure. The parameters defined to assess the spatial accuracy and resolution uniformity proved to be suitable quantities to perform such tests. The analysis of the peak positions of the power spectra allowed us to detect a local distortion of 4%, which is not visible with a simple visual inspection. The relative amplitude of the first and second harmonic peaks Ap was found to be sensitive to a loss of resolution power similar to that caused by the application of a blurring filter with kernel greater than 1. The time required to complete one set of measurement depends on the number and types of IPs. As an example for a single CR reader with 39 IPs of three different sizes, the time required for exposure and acquisition was about 150 min and the processing time was less than 30 min. In conclusion, the implemented analysis algorithms allow us to perform a quality control for CR with a short processing time and with absence of subjectivity in the evaluation of the parameters. Acknowledgments The study was supported, in part, by grants from the project ‘Riduzione del rischio associato all’esposizione a radiazioni ionizzanti per fini medici—Compagnia San Paolo di Torino’. References AAPM 2005 Assessment of Display Performance for Medical Imaging Systems Task Group 18 AAPM On-Line Report No. 03 http://www.aapm.org/pubs/reports/public/OR 03.pdf


17

AGFA 1999 ADC compact auto QC software user manual Agfa-Gavaert N.V. 2217B GB 19991125 Bradford C D, Peppler W W and Dobbins J T 1999 Performance characteristics of a Kodak computed radiography system Med. Phys. 26 27–37 Cesar L J, Schueler B A, Zink F E, Daly T R, Taubel J P and Jorgenson L L 2001 Artefacts found in computed radiography Br. J. Radiol. 74 195–202 Christodoulou E G, Goodsit M M, Chan H and Hepburn T W 2000 Phototimer setup for CR imaging Med. Phys. 27 2652–58 Fuji 1993 Automatic setting functions for image density and range in the FCR system Technical Review No. 3 Fuji Photo Film Co. Ltd., Tokyo, Japan Fujita H, Tsai D Y, Itoh T, Doi K, Morishita J, Ueda K, Tsai D Y and Ohtsuka A 1992 A simple method for determining the modulation transfer function in digital radiography IEEE Trans. Med. Imaging 11 34–9 Goldman L W 2004 Speed values, AEC performance evaluation, and quality control with digital receptors Medical Physics Monograph No. 30 Medical Physics Publishing) pp 271–97 IEC 2003 Medical electrical equipment—characteristics of digital imaging devices— part 1: determination of the detective quantum efficiency International Electrotechnical Commission, International Standard IEC 62220-1 Kengyelics S M, Davies A G and Cowen A R 1998 A comparison of the physical imaging properties of Fuji ST-V, ST-VA, and ST-VN computed radiography image plates Med. Phys. 25 2163–9 Kodak 2001 Guidelines for acceptance testing and quality control Kodak DirectView CR 800 and CR 900 System Technical and Scientific Bulletin Kwan A L C, Filipow L J and Le L H 2003 Automatic quantitative low contrast analysis of digital chest phantom radiographs Med. Phys. 30 312–20 Masden M T 1997 A method for quantifying SPECT uniformity Med. Phys. 24 1696–700 Passariello R 2000 Radiologia: elementi di tecnologia (Napoli: Idelson-Gnocchi) Pratt W K 1991 Digital Image Processing (New York: Wiley) Rowland J A 2002 The physics of computed radiography Phys. Med. Biol. 47 123–66 Samei E and Reimann D A 1997 A method for measuring the presampled MTF of digital radiographic systems using an edge test device Med. Phys. 25 102–13 Samei E, Seibert J A, Willis C E, Flynn M J, Mah E and Junck K L 2001 Performance evaluation of computed radiography systems Med. Phys. 28 361–71 Seibert J A et al Acceptance testing and quality control of photo-stimulable phosphor imaging systems Report of the American Association of Physicists in Medicine (AAPM) Task Group No. 10 (unpublished, in the final review process) Stierstorfer K and Spahn M 1999 Self-normalizing method to measure the detective quantum efficiency of a wide range of x-ray detectors Med. Phys. 26 1312–9 Tucker D M and Rezentes P S 1997 The relationship between pixel value and beam quality in photostimulable phosphor imaging Med. Phys. 24 887–93 Willis C E and Slovis T L 2005 The ALARA concept in pediatric CR and DR: dose reduction in pediatric radiographic exams—a white paper conference Am. J. Roentgenol. 184 373–4

See endnote 1

See endnote 2

Endnotes (1) Author: Please update reference ‘Seibert et al’. (2) Author: Please check whether the year in reference ‘Stierstorfer and Spahn 1999’ is OK, as set. Reference linking to the original articles References with a volume and page number in blue have a clickable link to the original article created from data deposited by its publisher at CrossRef. Any anomalously unlinked references should be checked for accuracy. Pale purple is used for links to e-prints at arXiv.