The British Journal of Radiology, 79 (2006), 981–990
Optimal beam quality selection in digital mammography 1
K C YOUNG,
PhD,
1
J M ODUKO,
PhD,
2
H BOSMANS,
PhD,
2
K NIJS,
BSc
and 3L MARTINEZ,
BSc
1
National Coordinating Centre for the Physics of Mammography, Royal Surrey County Hospital, Guildford GU2 7XX, UK, UZ Gasthuisberg, Department of Radiology, Herestraat 49, B-3000 Leuven, Belgium and 3Department of Physics, The Royal Marsden NHS Trust, London SW3 6JJ, UK 2
ABSTRACT. An experimental method of determining the optimal beam quality for digital mammography systems was applied to two systems (Fuji Profect and GE Senographe 2000D). The mean glandular dose (MGD) and contrast-to-noise ratio (CNR) were measured using Perspex breast phantoms simulating breasts from 20 mm to 90 mm thick. For each thickness, four combinations of tube voltage and target/filter were tested. Optimal beam quality was defined as giving a target CNR for the lowest MGD and was similar for the two systems. For breasts with a thickness of 21 mm or 32 mm, a tube voltage of either 25 kV or 28 kV and a Mo/Mo target/filter combination was optimal. For breast thicknesses of 45 mm and greater, the combination that had the highest X-ray energy (34 kV Rh/Rh) was optimal. Optimization using the higher energy beam quality required greater detector dose to compensate for the lower contrast. Thus for a 75 mm thick breast the 34 kV Rh/Rh combination required about a 90% greater detector dose than 28 kV Mo/Mo to achieve the same CNR because of the 25% reduction in contrast. Nonetheless, the MGD was reduced by 32% by choosing the higher energy spectra and achieving the same CNR. Current automatic exposure control (AEC) designs that aim for a fixed detector dose are not optimal and greater use of higher energy spectra should be accompanied by higher detector doses at all breast thicknesses which are average or above. This may result in slightly higher doses, but better image quality for these breasts.
Modern mammographic systems allow the operator to choose the beam quality by varying the tube voltage, and filter and target materials. The main issues to be considered in choosing a particular set of exposure parameters (and therefore beam quality) are the consequences for the absorbed dose to the glandular tissues and the image quality of the mammogram. A third consideration is the exposure time. In film–screen imaging, the appropriate dose to the detector is that required to achieve the correct optical density on the processed film. As a result, the main effect of using a higher X-ray energy with a film–screen system is to reduce patient dose at the expense of a loss in image contrast [1]. This has limited the use of higher energy X-rays to the small proportion of women with the greatest thickness on compression where a relatively small contrast loss may be acceptable for a large dose saving. In a review of radiation doses in the NHS Breast Screening Programme, it was found that only 1.4% of exposures used either a Rh/Rh or a W/Rh target/filter combination with film–screen systems [2]. Where digital mammography systems have been introduced there is sometimes a tendency to continue with similar practices. However, the optimization process is different with a digital system as the dose to the detector can be widely varied, in addition to the spectrum, without exposures going beyond the useful dynamic range of the system. There is also the potential for recovering perceptible Address correspondence to: Dr Kenneth C Young, Department of Medical Physics, Royal Surrey County Hospital, Egerton Road, Guildford GU2 7XX, UK. E-mail:
[email protected]
The British Journal of Radiology, December 2006
Received 19 January 2006 Revised 10 April 2006 Accepted 25 April 2006 DOI: 10.1259/bjr/55334425 ’ 2006 The British Institute of Radiology
contrast by means of post-processing algorithms [3], but inevitably any post-processing contrast amplification is limited ultimately by the inherent radiographic contrast, the quantum noise and the dynamic range of the system. Dance et al [4] conducted a Monte Carlo study on the influence of anode/filter and tube potential in digital mammography. They concluded that anode/filter combinations other than Mo/Mo, which result in higher energy spectra, are likely to be more suitable in digital mammography for all but the thinnest breasts. Until now there has been relatively little experimental data to confirm whether this conclusion holds in practice. Huda et al [5] investigated detection performance in digital mammography as a function of tube voltage and tube current–time product and lesion size, but only a single target/filter combination was used. Increasing tube voltage reduced the perceptibility and the results with changing mAs agreed well with the Rose model of image perception for small lesions, but it was found that for large lesions the perception was largely independent of exposure parameters. The images were computer generated and therefore represented an idealised system. With modern mammographic units it is necessary to go beyond this and incorporate the effects of spectral changes due to target and filter selection as well as tube potential. This paper describes an experimental procedure for optimizing the X-ray exposure factors on digital systems by assessing the un-processed radiographic contrast-to-noise ratio (CNR) and patient dose using polymethyl-methacrylate (PMMA), i.e. Perspex or Lucite, to simulate breasts of different thickness. The use of thin layers of aluminium to represent calcifications and 981
K C Young, J M Oduko, H Bosmans et al
estimate the contrast produced radiographically is well established in mammography and has been used by several other investigators [6–8]. Likewise, PMMA has been widely used to simulate breast tissue as it has similar radiographic properties; the thickness of PMMA equivalent in terms of absorption to selected thickness of breast tissue has been characterized using Monte Carlo techniques [9].
Methods Systems tested The optimization procedure was applied to two readily available digital mammography systems; a Fuji FCR Profect CS computed radiography (CR) system (Fuji Photo Film Co Ltd, Bedford, UK) used with a General Electric DMR+ mammography X-ray set (General Electric Medical Systems, Paris, France) and a General Electric Senographe 2000D. This latter system is of a type commonly called a digital radiology (DR) system, distinguished from the CR systems by having a detector integrated into the X-ray set. The detector for the automatic exposure control (AEC) on the DMR+ was always placed in the position closest to the chest wall edge. The same cassette and image plate were used for all measurements with the CR system. The CR system was used with a ‘‘fixed’’ exposure data recognizer (EDR) setting and a reading sensitivity or S value of 120. Each CR image comprised an array of 50 mm pixels. The DR system produced images with 100 mm pixel size in the detector plane. The DMR+ and the Senographe 2000D have three AEC modes (Standard, Dose and Contrast) which vary the balance between lowering the dose and increasing the image quality. The Dose mode is biased towards a lower dose and the Contrast mode biased towards higher contrast and therefore higher image quality. (Note that the two systems tested had the same three target/filter combinations available. The highest energy was obtained by selecting a Rh/Rh combination for these two systems. It was not possible to evaluate the use of alternative target/filter combinations, such as W/Rh, which are available on other manufacturer’s machines.)
Detector response and linearization of pixel values The detector response as a function of dose was established for each system using a phantom comprising 18 cm624 cm plates of PMMA with total thickness of 45 mm. This phantom was exposed using a tube voltage of 28 kV and a Mo/Mo target/filter combination and a range of tube current–time products from the minimum to maximum possible. The images were saved as unprocessed files and transferred to another computer for analysis. A 10 mm610 mm square region of interest (ROI) was positioned on the mid-line and 60 mm from the chest wall edge of each image. The mean pixel value and the standard deviation of pixel values within that region were measured for each image. The relationship between average pixel values and the mAs was determined. This is generally linear for digital systems 982
with integrated detectors such as the DR system tested here [10], but there may be an offset such that a linear fit to the data does not pass through the origin. For CR systems, the relationship is generally non-linear due to pre-processing [11, 12]. These data were used to convert all the original pixel values measured in CR images to a new pixel value which was linear with exposure, and therefore energy absorbed in the detector. (From this point on, the use of the term ‘‘pixel value’’ when referring to the CR system refers to the linearized values.) No adjustment to the pixel values was necessary with the images from the DR system as they were already linear with the energy absorbed in the detector.
Noise analysis The standard deviations for the pixel values in the ROI for each image were used to investigate the relationship between dose to the detector and image noise. The noise was assumed to be strongly dependent on the total energy absorbed by the detector and therefore on the pixel value linearized with dose (p). It was further assumed that this noise would comprise three components; electronic noise, structural noise and quantum noise with the relationship shown in Equation (1). These three noise sources are discussed in more detail by Evans et al [13]. sp ~
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ke 2 zkq 2 pzks 2 p2
ð1Þ
where sp is the standard deviation in pixel values within an ROI with a uniform exposure and a mean pixel value p, and ke, kq and ks are the coefficients determining the amount of electronic, quantum and structural noise in a pixel with a value p. For simplicity, the noise is generally presented here as relative noise defined as in Equation (2). Relative noise~
sp p
ð2Þ
The variation in relative noise with mean pixel value was evaluated and fitted using Equation (1), and nonlinear regression used to determine the best fit for the constants and their asymptotic confidence limits (using Graphpad Prism Version 4.03 for Windows; Graphpad software, San Diego, CA, www.graphpad.com). This established whether the experimental measurements of the noise fitted this equation, and the relative proportions of the different noise components. This also enabled the effect of varying the dose on the CNR to be modelled. In fact, the relationship between noise and pixel values was found empirically to be approximated by a simple power relationship as shown in Equation (3). sp ~kt p{n p
ð3Þ
where kt is a constant. If the noise was purely quantum noise, the value of n would be 0.5. However, the presence of electronic and structural noise means that n can be slightly higher or lower than 0.5. The British Journal of Radiology, December 2006
Optimal beam quality selection in digital mammography
CNR and dose measurement PMMA blocks with an area of 180 mm6240 mm and a total thickness ranging from 20 mm to 70 mm were used to simulate breasts of typical composition. An aluminium square (10 mm610 mm) with a 0.2 mm thickness was placed on top of the 20 mm thick block, with one edge on the midline and 60 mm from the chest wall edge. This position avoids the area above the AEC chamber of the X-ray set when testing the CR system. Additional layers of PMMA were added on top to vary the total thickness. The aluminium square was maintained at a fixed height and position so that the image of the square always occupied the same size and position on the detector array. This also made it easier to automate the subsequent ROI analysis. For each system, four different tube voltage and target/filter combinations were used covering a wide range of X-ray energies: 25 kV Mo/Mo, 28 kV Mo/Mo, 30 kV Mo/Rh and 34 kV Rh/Rh. To assess the dose, the X-ray factors (tube voltage, target material, filter material and mAs) were recorded for each exposure. The X-ray set output, half value layer (HVL) and the distance from the focus to table top were measured allowing the entrance surface air kerma at the top of the PMMA to be calculated. The method described by Dance et al [9] was used to calculate the mean glandular dose (MGD) to typical breasts (in the age range 50–65 years) with attenuation equivalent to the PMMA. For each beam quality and PMMA thickness three exposures were made. The first was designed to achieve a standard pixel value, similar to that selected when using the AEC. The next two images were recorded using approximately double and half that mAs to provide a wide range of doses. For each image, the average pixel values for ROIs in the centre of the aluminium square, mean(Al), and in the adjacent background area, mean(bgd), were measured. The standard deviation of the pixel values in the background ROI, sd(bgd), and the aluminium ROI, sd(Al), were also determined. These data were used to calculate the contrast and CNR for each image, as described in Equations (4) and (5). This procedure was repeated with the different tube voltage and target/filter combinations. Contrast~
mean(bgd){mean(Al) |100% mean(bgd)
mean(bgd){mean(Al) CNR~ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sd(bgd)2 zsd(Al)2 2
ð4Þ
ð5Þ
The expected relationship between CNR and dose is described by Equation (6). This can be derived from Equation (3) since the mean glandular dose (D) is proportional to the pixel value and contrast is independent of dose for a given beam quality and thickness of PMMA. CNR~kDn
ð6Þ
k is a constant that depends on the beam quality and thickness of PMMA, but is independent of dose. This The British Journal of Radiology, December 2006
equation was used to interpolate the effect on CNR of doses between the experimentally determined values for a given beam quality. The appropriate value of n was determined for each system from the analysis of the noise as a function of the mean pixel value. In practice, this was done by finding the value of n that provided the best fit for Equation (3) to the experimental data. The value of k was determined by regression analysis using Equation (6). In both cases non-linear regression was performed using Graphpad Prism. The PMMA test blocks were also exposed at each thickness tested using the X-ray set in automatic mode. For the DR system, all three automatic modes were used. For the CR system, only the clinically used Standard mode was evaluated. Since the choice of exposure factors depends on both attenuation and breast thickness, expanded polystyrene spacers were added to the PMMA blocks before compression to a standard force of 100 N. The spacers were positioned to avoid the central part of the image that is used to determine the total attenuation. The thickness of each spacer was selected to ensure that the total thickness equalled the equivalent breast thickness. Such spacers were not used with the 20 mm and 30 mm thicknesses of PMMA as the thickness adjustment was small (i.e. # 2 mm). The beam quality selected, MGD, background pixel value and CNR were determined for each image.
Results In the results that follow, the errors shown indicate the 95% confidence limits unless specified otherwise.
Detector response and noise analysis The CR system was found to have the expected logarithmic relationship between dose and pixel values. This relationship was used to transform all measured pixel values to ones with a linear relationship with dose as illustrated in Table 1. The results of the noise analysis for the CR system are shown in Figure 1. The curves for the noise components and the total noise were derived by fitting Equation (1) to the data. The coefficients for the three noise components are shown in Table 2. There was a small amount of electronic noise and the structural noise was 0.54% of the pixel value. The largest noise component was the quantum noise. It can be seen from Figure 1 that the curve representing the total noise is very similar in shape to that representing the quantum noise. It was approximated within the pixel value range measured by the power relationship shown in Equation Table 1. Linearization of mean pixel values for CR images Set tube current–time product (mAs) using 28 kV Mo/Mo
Mean pixel value measured in background ROI
Linearized pixel value (p)
32 45 64 90 125
301 455 604 762 911
32.0 45.3 63.2 90.0 125.6
983
K C Young, J M Oduko, H Bosmans et al
Figure 1. Noise components as a function of linearized pixel
Figure 2. Noise components as a function of pixel values for
values for the CR system. The data points come from the background regions of interest used to measure the contrast-to-noise ratio at all the thicknesses of PMMA, each of the beam qualities tested and the different dose levels.
the DR system. The data points come from the background regions of interest used to measure the contrast-to-noise ratio at all the thicknesses of PMMA, each of the beam qualities tested and the different dose levels.
(3), where kt is 0.115¡0.002 and n50.471¡0.005, with a correlation coefficient R of 0.997. The individual measurements of the noise deviated from this approximation within about 2% (95% confidence limit). The noise components for the DR system are shown in Figure 2 and the coefficients derived shown in Table 2. The curves shown in Figure 2 for the noise components and the total noise were derived by fitting Equation (1) to the data. For this system Equation (3) was also an excellent fit with kt50.50¡0.06 and n50.536, with a correlation coefficient R of 0.987.
Equation (6) as shown in Figure 4 for a 75 mm thick breast. (A CNR of 8.5 was typical for the current operation of the system with a 45 mm thickness of PMMA. If a higher or lower value had been selected, the optimal choice for the beam quality would be relatively unaffected.) The doses, pixel values and contrast corresponding to a CNR of 8.5 are shown in Table 3 for each spectrum and thickness tested using the CR system. The MGD for each spectrum is plotted against equivalent breast thickness in Figure 5. Exposure times have also been calculated for the different beam quality selections and were always shorter the higher the energy of the spectrum. The exposure times for a breast thickness equivalent to 90 mm were 4.6 s for 28 kV Mo/Mo, 3.3 s for 30 kV Mo/Rh and 2.9 s for 34 kV Rh/Rh. The data were also interpolated to find the CNR and MGD if the mAs was adjusted to achieve a background pixel value of 50 in each image regardless of the beam quality used, and the results are shown in Table 4. A pixel value of 50 was chosen as this was approximately the mean linearized pixel value resulting from using the X-ray set in its AEC mode regardless of what beam quality was selected. The amount of relative noise was calculated for each of the interpolated points. The thickness had no effect on the noise, but there was a small difference in noise between the beam qualities. This is analysed in Table 5 where the average noise is shown to be slightly greater for the two higher energy
CNR and dose: CR system The contrast measurements for the CR system are shown in Figure 3. Each point is the average of the measurements at the three dose levels with a 95% confidence limit of 0.1%. For each PMMA thickness, the calculated CNR was plotted against the mean glandular dose for the different beam qualities tested, as illustrated in Figure 4. A curve of the form shown in Equation (6) was fitted to the points corresponding to the three dose levels. It was assumed that satisfactory image quality would be achieved if a specific CNR were achieved. For the purpose of comparison, a CNR of 8.5 was selected and the dose necessary to achieve this determined by interpolation of the CNR data using
Table 2. Coefficients determined from the noise analysis (errors indicate 95% confidence limits) System
Electronic noise coefficient (ke)a
Structural noise coefficient (ks)
Quantum noise coefficient (kq)a
n
Fuji Profect GE Senographe 2000D
0.11¡0.05 3.42¡0.06
0.0054¡0.0007 0.0036¡0.0009
0.122¡0.003 0.363¡0.009
0.471¡0.005 0.536¡0.018
a
The values of the electronic and quantum noise coefficients are dependent on the relationship between pixel value and detector dose, therefore the values for the two systems cannot be compared.
984
The British Journal of Radiology, December 2006
Optimal beam quality selection in digital mammography
Figure 3. Contrast for 0.2 mm aluminium for different breast thickness and beam quality using the CR system. (Error bars indicate 95% confidence limits).
beam qualities (30 kV Mo/Rh and 34 kV Rh/Rh) than the two lower ones (25 kV Mo/Mo and 28 kV Mo/Mo). In each case this difference was significant (p , 0.01 Student t-test).
Contrast-to-noise ratio and dose: DR system The contrast, dose and CNR measurements on the DR system were performed and analysed as for the CR system. However, in view of the higher CNR achieved a
Figure 4. Contrast-to-noise ratio (CNR) vs mean glandular dose (MGD) for a simulated 75 mm thick breast for the CR system. There are no data for the 25 kV Mo/Mo combination as the backup timer prevented exposures of the 60 mm thickness of PMMA. (Error bars indicate 95% confidence limits).
The British Journal of Radiology, December 2006
CNR of 12 was selected for comparison and the dose and pixel values necessary to achieve this for each beam quality are shown in Table 6. Figure 6 illustrates the interpolation process for a 75 mm thick breast. The MGD for each spectrum with a CNR of 12 is plotted against equivalent breast thickness in Figure 7. Exposure times have also been calculated for the different beam quality selections and were again always shorter the higher the energy of the spectrum. The exposure times for a breast thickness equivalent to 90 mm were 7.3 s for 28 kV Mo/ Mo, 4.9 s for 30 kV Mo/Rh and 3.0 s for 34 kV Rh/Rh. The data were also interpolated to find the CNR and MGD if the mAs were adjusted to achieve a pixel value of 900 in each image regardless of the beam quality used, and the results are shown in Table 7. A pixel value of 900 was chosen as this was approximately the average pixel value resulting from using the X-ray set in its automatic modes regardless of what beam quality was selected. The effect of this strategy (i.e. a fixed pixel value of 900) on the CNR at different breast thicknesses is shown for the four beam qualities in Figure 8. The beam qualities, pixel values, doses and CNR when the system was used in its three automatic modes are shown in Table 8.
Discussion The optimization procedure described here assumes that the choice of beam quality primarily affects the contrast and noise in the image, and the dose to the patient, and that it is the CNR that should be optimized. There is an implicit assumption that the MTF of the imaging system is not significantly affected by changing beam quality. This is an assumption that may need to be verified for some designs. In practice one also needs to ensure that exposure times are not too long. For the two systems studied here a consideration of the exposure times would not alter the optimal choice, as the highest energy spectrum (34 kV Rh/Rh) gave the shortest exposure times and the target CNR were achieved within reasonable exposure times (# 3 s) even for the largest simulated thickness. It is also implicitly assumed that the signal difference or contrast generated by the 0.2 mm thickness of aluminium is representative of the clinically important tissues in the breast and that CNR is the relevant measure affecting their visibility. The use of aluminium to simulate the attenuation of calcifications has been employed by a number of authors and was recently subject to experimental verification by Carton et al [14]. These authors compared the attenuation properties of biopsied microcalcifications with aluminium at 25 kV and 31 kV and a Mo/Mo target/filter combination. They concluded that aluminium and calcifications had a similar energy dependence. A comparison of the linear attenuation coefficients of calcifications and aluminium shows that they are remarkably similar across the full range of energies encountered in mammography. The attenuation characteristics of other breast tissues were measured by Johns and Yaffe [15]. They considered the normal breast to be composed of fat and fibrous tissues, commonly referred to as glandular tissues by other authors. They also found that the attenuation of infiltrating duct carcinoma was similar to fibrous tissues. 985
K C Young, J M Oduko, H Bosmans et al Table 3. Dose, pixel value and contrast for the CR system with a CNR of 8.5 (errors indicate 95% confidence limits) PMMA thickness (mm)
20
30
40
45
50
60
70
Equivalent breast thickness (mm)
21
32
45
53
60
75
90
Mean glandular dose (mGy)
0.24¡0.01 0.24¡0.01 0.26¡0.01 0.31¡0.01 24¡1 26¡1 35¡2 59¡2 21.2¡0.1 20.1¡0.1 18.0¡0.1 14.7¡0.1
0.47¡0.01 0.46¡0.01 0.49¡0.02 0.52¡0.02 29¡2 33¡1 45¡3 69¡1 19.6¡0.1 18.4¡0.1 16.4¡0.1 13.6¡0.1
0.92¡0.04 0.91¡0.03 0.96¡0.06 0.89¡0.04 34¡1 41¡1 55¡3 85¡3 18.4¡0.1 17.0¡0.1 14.9¡0.1 12.4¡0.1
1.29¡0.03 1.33¡0.03 1.27¡0.03 1.17¡0.06 35¡1 46¡1 59¡1 91¡4 17.7¡0.1 16.1¡0.1 14.2¡0.1 11.8¡0.1
1.78¡0.03 1.88¡0.03 1.66¡0.04 1.51¡0.07 36¡1 50¡1 60¡2 98¡4 17.4¡0.1 15.4¡0.1 13.9¡0.1 11.5¡0.1
3.85¡0.15 3.35¡0.14 2.64¡0.04 60¡2 76¡3 116¡2 14.0¡0.1 12.5¡0.1 10.5¡0.1
7.25¡0.26 6.16¡0.56 4.41¡0.11 70¡4 90¡12 132¡5 13.0¡0.1 11.6¡0.1 9.8¡0.1
Pixel values
Contrast (%)
25 28 30 34 25 28 30 34 25 28 30 34
kV kV kV kV kV kV kV kV kV kV kV kV
Mo/Mo Mo/Mo Mo/Rh Rh/Rh Mo/Mo Mo/Mo Mo/Rh Rh/Rh Mo/Mo Mo/Mo Mo/Rh Rh/Rh
CR, computed radiography; CNR, contrast-to-noise ratio.
experimentally here. Johns and Yaffe pointed out that the visibility of carcinomas without calcifications is not likely to be limited by quantum noise, but rather by the structural noise (or clutter) of the background normal fibrous tissues. This was supported more recently by Huda et al [5] who concluded that radiographic technique factors have little effect on detection performance for lesions larger than about 0.8 mm, but that the visibility of smaller lesions is affected by quantum mottle. This suggests that optimizing the technique factors using CNR measurements as described here is important for ensuring adequate detection of small micro-calcifications, but may have little effect on the detection of the larger features of masses greater than about 1 mm in scale. Smaller features such as spicules may be affected and are likely to be more easily seen where the CNR is relatively high. A very simple measure of noise has been used in calculating CNR. A more complete analysis would involve the evaluation of the noise power spectra [16] and breast structure noise [17]. However, CNR is expected to be a valid tool for optimization where the dominant noise source is quantum noise, as found here.
Figure 5. Mean glandular dose for the CR system interpolated for a contrast-to-noise ratio (CNR) of 8.5. (Error bars indicate 95% confidence limits).
Thus the detection of a carcinoma identifiable as a mass depends primarily on the shape, appearance and location of the mass rather than on the mass having different attenuation characteristics from the surrounding normal glandular tissue. Dance et al [4] calculated the optimal spectra for 5 mm of glandular tissue as well as a 200 mm calcification detail using Monte Carlo modelling and found that the optimal factors were broadly similar, and in line with the results found
Detector response and noise analysis The CR system had a logarithmic response to detector dose as described by the manufacturer. The DR system had a linear response to detector dose. The noise analysis showed that quantum noise was the dominant noise
Table 4. Dose and CNR for the CR system with a pixel value of 50 PMMA thickness (mm)
20
30
40
45
50
60
70
Equivalent breast thickness (mm)
21
32
45
53
60
75
90
Mean glandular dose (mGy)
0.50¡0.02 0.46¡0.02 0.38¡0.01 0.26¡0.01 12.1¡0.2 11.5¡0.3 10.1¡0.3 7.8¡0.1
0.80¡0.03 0.69¡0.02 0.55¡0.02 0.38¡0.01 11.0¡0.3 10.3¡0.1 8.9¡0.2 7.3¡0.1
1.36¡0.05 1.12¡0.04 0.87¡0.03 0.53¡0.02 10.2¡0.2 9.4¡0.1 8.1¡0.2 6.6¡0.1
1.83¡0.06 1.46¡0.05 1.08¡0.04 0.64¡0.02 10.0¡0.1 8.9¡0.1 7.9¡0.1 6.4¡0.1
2.45¡0.09 1.90¡0.07 1.38¡0.05 0.77¡0.03 9.9¡0.1 8.5¡0.1 7.8¡0.1 6.2¡0.1
3.19¡0.11 2.22¡0.08 1.14¡0.04 7.8¡0.2 7.0¡0.1 5.7¡0.1
5.21¡0.18 3.43¡0.12 1.68¡0.06 7.3¡0.1 6.5¡0.3 5.4¡0.1
CNR
25 28 30 34 25 28 30 34
kV kV kV kV kV kV kV kV
Mo/Mo Mo/Mo Mo/Rh Rh/Rh Mo/Mo Mo/Mo Mo/Rh Rh/Rh
CR, computed radiography; CNR, contrast-to-noise ratio.
986
The British Journal of Radiology, December 2006
Optimal beam quality selection in digital mammography
using analogue mammography systems. This is understandable because the dose selection was effectively fixed to that necessary to achieve the correct optical density on the film. However, in digital mammography the detector dose can be freely varied. In this case, the optimal spectral choice depends on achieving the best CNR for the lowest dose. This can be determined by inspection of Table 3. For the smallest simulated breast thicknesses (21 mm and 32 mm) the optimal spectrum was selected by using either 25 kV or 28 kV and a Mo/ Mo target/filter combination. For the 45 mm thickness and above, the optimal choice was the spectrum that produced the highest energies (34 kV Rh/Rh). However, it should be noted that these results were achieved by greatly increasing the detector dose to compensate for the lower contrast produced at higher energies. The reduction in MGD achieved by using the highest energy spectra was quite small, 4% and 12% for the 45 mm and 53 mm breast thicknesses, respectively. However, the dose savings were large at greater thicknesses. Thus for a 75 mm thick breast the 34 kV Rh/Rh combination required a 93% greater detector dose than 28 kV Mo/ Mo to achieve the same CNR. Nonetheless, the MGD of 2.64 mGy was 32% lower where the higher energy spectra was used in preference to 28 kV Mo/Mo, to achieve this CNR. One can compare this dose with what may be expected with the current AEC designs by considering the doses in Table 4. In this case, the detector dose is fixed and the MGD depends strongly on the beam quality selected. A combination of 28 kV Mo/Mo leads to an MGD of 3.19 mGy for the 75 mm thick breast. In practice, modern sets are unlikely to select such a combination at this thickness and a combination such as 30 kV Mo/Rh would be more usual, leading to an MGD of 2.22 mGy. It is unlikely that a Rh/Rh target/filter combination would be selected, but if it were the MGD would be much lower at 1.14 mGy. Table 4 also shows that, with current AEC designs, the lower MGD achieved by employing higher energy combinations also leads to reduced CNR and therefore image quality. The proposal in this paper that AECs should be designed to aim for a specific target CNR may lead to relatively high doses for the larger thicknesses of breast with CR systems. Acceptable limits for different thicknesses of breast are provided in European Guidelines for digital mammography systems and are designed to ensure that doses
Table 5. Mean relative noise for different beam qualities at the absorbed detector dose used clinically Mean relative noise (%)¡2 standard error in the mean Beam quality
Fuji Profect (Pixel value550)
GE 2000D (Pixel value 5 900)
25 28 30 34
1.777¡0.006 1.792¡0.006 1.811¡0.008 1.816¡0.007
1.22¡0.01 1.27¡0.01 1.33¡0.01 1.36¡0.01
kV kV kV kV
Mo/Mo Mo/Mo Mo/Rh Rh/Rh
source for both systems, with smaller amounts of structural and electronic noise. There was slightly more structural noise with the CR system, 0.54% of the pixel value as opposed to 0.36% for the DR system. The total noise was about 1.8% of the pixel value for the CR system at the clinically used dose. The total noise was 1.3% of the pixel value for the DR system. The noise for each system was almost completely determined by the pixel value and therefore dose to the detector. However, the energy spectrum used also had a small effect with slightly higher noise at higher energies. Thus using the Mo/Rh or Rh/Rh target/filter combinations resulted in about 2% more noise than using Mo/Mo with the CR system. However, this is a much smaller effect than the corresponding contrast reduction, which is about 25% to 30%. For the DR system, the beam quality had a larger effect on the noise. Thus using the Rh/Rh combination resulted in 7% more noise than using the 28 kV Mo/Mo tube voltage and target/filter combination for the same pixel value. The increased noise at higher energies may be expected because there are fewer quanta for the same detector dose.
CNR and dose: CR system Figure 3 shows that the lower energy spectra produced the greatest contrast at all breast thicknesses. It also shows how contrast declines with increasing breast thickness. This result is as expected for any mammographic imaging system and was also found here for the DR system. It is the reason why there has been a historical preference for lower energy spectra when
Table 6. Dose, pixel value and contrast for the DR system with a CNR of 12 (errors indicate 95% confidence limits) Equivalent breast thickness (mm)
21
32
45
53
60
75
90
Mean glandular dose (mGy)
25 28 30 34
kV kV kV kV
Mo/Mo Mo/Mo Mo/Rh Rh/Rh
0.36¡0.03 0.35¡0.02 0.39¡0.02 0.46¡0.03
0.68¡0.05 0.69¡0.05 0.72¡0.05 0.76¡0.02
1.34¡0.06 1.37¡0.05 1.35¡0.10 1.32¡0.08
1.85¡0.06 1.95¡0.20 1.87¡0.33 1.71¡0.04
2.72¡0.08 2.57¡0.18 2.21¡0.11
5.46¡0.35 10.33¡0.23 4.68¡0.25 8.94¡1.80 3.69¡0.20 6.24¡0.38
Pixel values
25 28 30 34
kV kV kV kV
Mo/Mo Mo/Mo Mo/Rh Rh/Rh
426¡32 481¡30 678¡40 1046¡66
497¡39 593¡46 837¡53 1239¡25
579¡24 721¡28 1023¡79 1501¡87
605¡20 801¡81 1138¡199 1603¡42
856¡26 1242¡88 1708¡88
1038¡67 1447¡78 1993¡106
25 28 30 34
kV kV kV kV
Mo/Mo Mo/Mo Mo/Rh Rh/Rh
20.8¡0.1 19.8¡0.1 17.4¡0.1 14.4¡0.1
19.6¡0.1 18.3¡0.1 15.9¡0.1 13.3¡0.1
18.3¡0.1 16.8¡0.1 14.8¡0.1 12.3¡0.1
17.8¡0.1 16.3¡0.1 14.4¡0.1 12.0¡0.1
15.5¡0.1 13.7¡0.1 11.5¡0.1
14.3¡0.1 13.1¡0.1 12.5¡0.1 11.7¡0.1 10.7¡0.1 10.0¡0.1
Contrast (%)
1215¡27 1802¡364 2331¡141
DR, digital radiography; CNR, contrast-to-noise ratio.
The British Journal of Radiology, December 2006
987
K C Young, J M Oduko, H Bosmans et al
Figure 6. Contrast-to-noise ratio (CNR) vs mean glandular dose (MGD) for a simulated 75 mm thick breast for the DR system. There are no data for the 25 kV Mo/Mo combination as the backup timer prevented exposures of the 60 mm thickness of PMMA. (Error bars indicate 95% confidence limits).
are not significantly greater than currently accepted with film–screen systems [18]. In the example of a 75 mm thick breast simulated with a 60 mm thickness of PMMA, the acceptable limit is set at 4.5 mGy, which is well above the MGD necessary when aiming for a CNR of 8.5 with a 34 kV Rh/Rh combination. In fact the acceptable limits are well above the doses shown in Table 3 for 34 kV Rh/Rh at all the thicknesses. The acceptable limits in the European Guidelines have recently been incorporated into guidance to the NHS Breast Screening programme on commissioning and
Figure 7. Mean glandular dose for the DR system interpolated for a CNR of 12. (Error bars indicate 95% confidence limits).
988
routine testing of full field digital mammography systems [19]. Since only four well-separated beam qualities were used in this study, it is not possible to determine exactly which tube voltage and target/filter combination would be optimal. This is an issue that will be addressed in further work. These findings are consistent with the conclusion reached by Dance et al [4] that for digital systems, target/filter combinations delivering relatively high energy spectra (e.g. Rh/Rh and W/Rh) are preferable for all but the smallest breasts. In clinical use, the CR system was used with an AEC which aimed for a fixed detector dose regardless of which spectrum was selected. The effect of this can be seen in Table 4. In this case, the highest CNR and therefore best image quality was always achieved by choosing the lowest energy spectra, at the expense of an unnecessarily high dose. It can be concluded that the current practice of using AECs that aim for constant dose to the detector are not optimally adjusted for such CR systems. If they choose higher energy spectra for large breasts, the image quality will be reduced in just the type of breast where image quality is already reduced due to increased scatter and beam hardening.
CNR and dose: DR system The contrasts measured with the DR system for the 4 spectral choices (Table 6) were, as expected, very similar to those found with the CR system (Table 4) However, since the noise was lower, the CNR values were generally higher. The optimal spectrum choices at each thickness were very similar to those found for the CR system. Thus either 25 kV or 28 kV used with a Mo/Mo target/filter combination should be preferred for the two smallest thicknesses tested. For all thicknesses of 45 mm or above, the 34 kV Rh/Rh target filter combination was found to be best. Once again an equivalent CNR is achieved, when using the highest energy spectrum, by increasing the detector dose to compensate for the reduced contrast. Thus for example, for a 75 mm thick breast, the 34 kV Rh/Rh combination required a 91% greater detector dose than 28 kV Mo/Mo to achieve the same CNR. Nonetheless the MGD was reduced by 33% by using the higher energy spectra. This finding is virtually identical to that found with the CR system. It cannot be assumed that such a finding will be found with all detector systems. In clinical use the DR system is employed in one of its three automatic modes. In these automatic modes the AEC aims for a constant dose to the detector. This resulted in a pixel value of about 900, as shown in Table 8. As breast thickness increased, the preferred exposure settings gradually changed to ones producing higher energies. In fact the spectra selected in DOSE mode are quite similar to the optimal ones deduced from Table 6. However, the policy of using a fixed dose to the detector resulted in the CNR declining rapidly with increasing thickness (Table 7). (A similar result can be expected with CR systems that use the X-ray sets AEC system to achieve a fixed detector dose at all thicknesses.) The latest generation of digital mammography systems from this manufacturer have a different AEC The British Journal of Radiology, December 2006
Optimal beam quality selection in digital mammography Table 7. Dose and CNR for the DR system with a pixel value of 900 (errors indicate 95% confidence limits) Equivalent breast thickness (mm)
21
Mean glandular dose (mGy)
25 28 30 34
kV kV kV kV
Mo/Mo Mo/Mo Mo/Rh Rh/Rh
CNR
25 28 30 34
kV kV kV kV
Mo/Mo Mo/Mo Mo/Rh Rh/Rh
32
0.74 0.65 0.51 0.39
45
1.21 1.04 0.77 0.56
17.7¡0.7 16.6¡0.6 13.9¡0.4 11.1¡0.4
53
2.07 1.70 1.19 0.80
16.4¡0.7 14.9¡0.6 12.5¡0.4 10.1¡0.1
2.74 2.19 1.48 0.97
15.2¡0.3 13.5¡0.3 11.2¡0.5 9.2¡0.3
60
75
90
-
-
-
2.86 1.87 1.17
14.8¡0.3 12.8¡0.7 10.6¡1.0 8.9¡0.1
4.74 2.92 1.68
12.3¡0.2 10.1¡0.4 8.5¡0.2
11.1¡0.4 9.3¡0.3 7.9¡0.2
7.66 4.47 2.42 10.2¡0.1 8.3¡0.9 7.2¡0.2
DR, digital radiography; CNR, contrast-to-noise ratio.
Figure 8. Contrast-to-noise ratio for the DR system using a fixed pixel value of 900. (Error bars indicate 95% confidence limits).
design which does not aim for constant detector dose and it is planned that their performance will be the subject of a further study.
General Theoretical calculations have suggested that the use of higher energy X-rays may be preferable for digital
systems as compared with film–screen systems [4]. This is because the physical contrast loss caused by using a higher energy X-ray can be compensated for by increasing the detector dose and thereby reducing the quantum noise in the image. The net result being that the required CNR is achieved for a lower dose to the patient. However, it is not certain that this will happen in practice. Most of the current digital systems, such as the two systems tested here, have AECs that are designed to achieve a constant dose to the detector. This is suitable where a film–screen imaging system is used. If such an AEC is used with a digital system, the use of higher energy X-rays causes a substantial reduction in CNR, and best image quality will be achieved by using the lowest energy spectra possible within the limits of exposure time. This can be seen from Tables 4 and 7. Thus 25 kV Mo/Mo resulted in the greatest CNR for breast thicknesses up to about 60 mm, and 28 kV Mo/ Mo for greater thickness. However, such a strategy is not the most dose efficient way of achieving these levels of CNR. It would be preferable to set up AECs to maintain a target level of CNR across as wide a range of breast thicknesses as possible. The test procedure described here provides a simple means of checking whether AECs on digital systems are well adjusted. The method of measuring CNR and dose described here has been adopted in European and UK Guidance on testing the AECs of digital mammography systems [18, 19]. A reason why practical digital systems might not always benefit from the use of higher energy spectra is the presence
Table 8. Beam quality, pixel value, dose and CNR for the DR system used in automatic modes Equivalent breast thickness (mm)
21
32
45
53
60
75
90
Beam quality selected
DOSE STD CNT
28 kV Mo/Mo 28 kV Mo/Mo 31 kV Mo/Rh 31 kV Rh/Rh 31 kV Rh/Rh 32 kV Rh/Rh 32 kV Rh/Rh 27 kV Mo/Mo 27 kV Mo/Mo 28 kV Mo/Rh 28 kV Rh/Rh 28 kV Rh/Rh 32 kV Rh/Rh 32 kV Rh/Rh 25 kV Mo/Mo 25 kV Mo/Mo 26 kV Mo/Mo 26 kV Mo/Rh 26 kV Mo/Rh 31 kV Rh/Rh 32 kV Rh/Rh
Pixel value
DOSE STD CNT
863 877 854
Mean glandular DOSE dose (mGy) STD CNT CNR
DOSE STD CNT
0.63 0.64 0.70 16.4 17.1 17.8
890 884 886 1.02 1.08 1.19 15.0 15.3 16.6
927 893 863 1.21 1.27 1.88 11.4 12.1 14.6
858 865 913 0.99 1.21 1.81 9.8 9.9 12.5
863 872 923 1.24 1.50 2.34 9.3 10.4 12.1
873 890 894
879 886 893
1.78 1.81 1.90
2.63 2.63 2.66
8.5 8.5 9.0
7.9 7.9 8.0
DR, digital radiography; CNR, contrast-to-noise ratio.
The British Journal of Radiology, December 2006
989
K C Young, J M Oduko, H Bosmans et al
of other noise sources than quantum noise. Structural noise is not reduced by increasing detector dose, so it may not be possible to reduce the overall noise sufficiently to compensate for contrast losses. In this case, relatively low energies may be preferable. For some detector systems, higher energies may cause a loss in detection efficiency, and again lower energies may be preferable. It is therefore important to experimentally verify the optimal beam quality for each design of digital system. The method described here can help in choosing the most appropriate beam quality for a given digital system. However, it is not yet clear what CNR is necessary for an optimally adjusted system. One approach to this would be to relate CNR data to measurements of threshold contrast as described in European Guidance [18]. A limitation in the method of optimization described here is that only four beam qualities were tested. This is something that requires further work before manufacturers can fully optimize AEC systems for digital mammography.
References 1. Young KC, Ramsdale ML, Rust A, Cooke J. Effect of automatic kV selection on dose and contrast in mammography. Br J Radiol 1997;70:1036–42. 2. Young KC, Burch A, Oduko JM. Radiation doses in the UK Breast Screening Programme in 2001 and 2002. Br J Radiol 2005;78:207–18. 3. Baydush AH, Floyd CE. Improved image quality in digital mammography with image processing. Med Phys 2000;27:1503–8. 4. Dance DR, Thilander Klang A, Sandborg M, Skinner CL, Castellano Smith IA, Alm Carlsson G. Influence of anode/ filter material and tube potential on contrast, signal-to-noise ratio and average absorbed dose in mammography: a Monte Carlo study. Br J Radiol 2000;73:1056–67. 5. Huda W, Ogden KM, Scalzetti EM, Dudley EF, Dance DR. How do radiographic techniques affect mass lesion detection performance in digital mammography? In: Proceedings of SPIE Medical Imaging 2004. SPIE 2004;5372:372–82. 6. Pachoud M, Lepori D, Valley JF, Verdun FR. A new test phantom with different breast tissue compositions for image quality assessment in conventional and digital mammography. Phys Med Biol 2004;49:5267–81. 7. Carton AK, Bosmans H, Vandenbroucke D, Souverijns G, Van Ongeval C, Dragusin O, et al. Quantification of
990
8.
9.
10.
11.
12.
13.
14.
15. 16.
17.
18.
19.
Al-equivalent thickness of just visible microcalcifications in full field digital mammograms. Med Phys 2004;31: 2165–76. Cowen AR, Launders JH, Jadav M, Brettle DS. Visibility of micro-calcifications in computed and film screen mammography. Phys Med Biol 1997;42:1533–48. Dance DR, Skinner CL, Young KC, Beckett JR, Kotre CJ. Additional factors for the estimation of mean glandular dose using the UK mammography dosimetry protocol. Phys Med Biol 2000;45:3225–40. Evaluation of the IGE Medical Systems Senographe 2000D full field digital mammography unit (Report 01041) London, UK: Medical Devices Agency (MDA), 2001. Computed radiography systems for mammography: Fuji FCR 5000MA and FCR Profect CS. (Report 04094) London, UK: Medicines and Healthcare products Regulatory Agency (MHRA), 2004. Kato H, Miyahara J, Takano M. Computed radiography with scanning laser stimulated luminescence. In: AAPM summer school 1984; Recent developments in medical imaging. AAPM 1985:237–55. Evans DS, Workman A, Payne M. A comparison of the imaging properties of CCD-based devices used for small field digital mammography. Phys Med Biol 2002;47:117–35. Carton AK, Bosmans H, Van Ongeval C, Souverijns F, Rogge F, Marchal G. Contrast visibility of simulated microcalcifications in full field mammography systems. In Proceedings of SPIE Medical Imaging 2003. SPIE 2003;5034:412–23. Johns PC, Yaffe MJ. X-ray characterization of normal and neoplastic breast tissues. Phys Med Biol 1987;32:675–95. Dobbins J. Image quality metrics for digital systems. In: Van Metter RL, Beutel J, Kundel HL, editors. Handbook of medical imaging, physics and psychophysics, volume 1. Bellingham, WA: SPIE, 2000. Burgess AE, Jacobsen FL, Judy PF. Human observer detection experiments with mammograms and power law noise. Med Phys 2001;28:419–37. Van Engen R, Young KC, Bosmans H, Thijssen M. The European protocol for the quality control of the physical and technical aspects of mammography screening. Part B: Digital mammography. In: European Guidelines for Breast Cancer Screening, 4th edn. Luxembourg: European Commission, 2005 (In press and available online at www.euref.org [Accessed 26 June 2006]). Commissioning and routine testing of full field digital mammography systems. (NHSBSP Report 0604) Sheffield: NHS Cancer Screening Programmes, 2006 (available online at www.cancerscreening.nhs.uk [Accessed 26 June 2006]).
The British Journal of Radiology, December 2006