Threshold contrast visibility of micro calcifications in digital mammography

Transcription

1 Threshold contrast visibility of micro calcifications in digital mammography Ann-Katherine Carton*, Hilde Bosmans*, Dirk Vandenbroucke, Chantal Van Ongeval*, Geert Souverijns*, Frank Rogge*, Guy Marchal* * Dept. Of Radiology, University Hospitals of KUL, Herestraat 49, Leuven 3000, Belgium Agfa-Gevaert, Septestraat 27, 2640 Mortsel, Belgium ABSTRACT The purpose of this study is to describe a method that allows the calculation of a contrast-detail curve for a particular system configuration using simulated micro calcifications into clinical mammograms. We made use of simulated templates of micro calcifications and adjusted their x-ray transmission coefficients and resolution to the properties of the mammographic system under consideration (4). We expressed the thickness of the simulated micro calcifications in terms of Al equivalence. In a first step we validated that the thickness of very small Al particles with well known size and thickness can be calculated from their x-ray transmission characteristics at a particular X-ray beam energy. Then, micro calcifications with equivalent diameters in the plane of the detector ranging from 300 to 800 µm and thicknesses, expressed in Al equivalent, covering 77 to 800 µm were simulated into the raw data of real clinical images. The procedure was tested on 2 system configurations: the GE Senographe 2000 D and the Se based Agfa Embrace DM1000 system. We adapted the X-ray transmissions and spatial characteristics of the simulated micro calcifications such that the same physical micro calcification could be simulated into images with the specific exposure parameters (Senographe 2000D: 28 kvp-rh/rh, Embrace DM1000: 28 kvp-mo/rh), compressed breast thickness (42+/-5mm) and detector under consideration. After processing and printing, 3 observers scored the visibility of the micro calcifications. We derived contrast-detail curves. This psychophysical method allows to summarize the performance of a digital mammography detector including processing and visualization. Keywords: simulation, micro calcifications, phantom, digital mammography, observer study 1. INTRODUCTION Characterization of digital mammography systems is often performed by means of contrast detail curves using a phantom with inserts of different sizes and thicknesses (1) (2). As the success of digital mammography also depends on the optimal selection of image processing protocols for display and printing we aimed in this study for a threshold contrast-detail analysis with closer link to clinical practice. To do so, we simulated micro calcifications with different sizes and thicknessess into clinical images and we scored their visibility. The calculation of the size of the micro calcification that corresponds with a self-created template, required a validation phase first: we proved that the x-ray transmission factors of small well known Al particles can be used to estimate their true thickness. A similar procedure was used to deduce size parameters for the templates of the micro calcifications. Templates of micro calcifications were created for 2 different detector systems. We started from templates in a given detector. These were then corrected for differences in x-ray transmission coefficients (as other beam qualities were used) and spatial characteristics. These templates were then inserted in the raw data of clinical mammograms and these composed images were then processed and viewed as routine mammograms. Then the visibility of the micro calcifications was compared for the 2 digital mammography systems. SPIE USE, V (p.1 of 10) / Color: No / Format: A4/ AF: A4 / Date: :41:11

2 2. MATERIAL AND METHODS 2.1. Validation of Al-equivalent thickness Very small rectangular-like Al particles with a well known size (200 µm to 3600 µm) and nominal thickness (100 and 500 µm) were imaged with a Senographe 2000D detector (GE Medical Systems, Milwaukee). The Al particles were exposed on top of 10 mm PMMA at 27 kvp, Mo/Mo (half value layer (HVL) with compression plate=0.308 mm Al). Magnification factor 2 and small focus were used. Measurements were performed on the images with signal intensities proportional to the x-ray exposure (the raw data). The image segments including Al particles were extracted. The signal intensities in these segments were normalized to the average signal intensities of the background, i.e. the region without micro calcification. This made the pixel values adjacent to the Al particles close to one. The pixel values at the location of the Al particles represent the fraction of transmitted x-rays as detected by the detector. The measured x-ray transmissions I 1 were transformed to the x-ray transmissions I 2 of the Al particles as would be obtained in an ideally sharp detector (i.e. Modulation Transfer Function (MTF) = 1 at all frequencies) using: 1 2DFFT [] ([] 1 I ) 1 1 I = 2DFFT 2 MTF 2GE 2 X where [1]-I 2 and [1]-I 1 are the image segments with the signal profiles pixel by pixel. The function 2DFFT is the 2- dimensional discrete Fourier transform, DFFT -1 is the inverse 2-dimensional discrete Fourier transform. MTF2 GE2x is the 2-dimensional MTF of the GE detector in magnification mode. The MTF2 GE2x was computed from the horizontal and vertical MTFs derived with an edge phantom (3). Our edge was exposed under similar conditions as those used for the acquisition of the Al particles (on top of 10mm PMMA at 27kVp, Mo/Mo, small focus, magnification = 2x). The calculated 1 dimensional MTFs were then modelled by Lorentzian-type functions: a 1 a MTF ( f ) = mod 2 + el i 2 f f i i b c where a, b, c are parameters with a [0,1], b>0 and c>0. Linear interpolation of the curves for the horizontal and vertical direction resulted in the MTF2 GE2x. Then, we exposed large Al sheets with thickness from 0.1 mm to 0.7mm using the same Senographe 2000D, identical geometry and exposure parameters as those used for the small Al particles. We calculated their x-ray transmissions. We made a log fit through the nominal thickness of the Al-sheets along the measured x-ray transmissions Al-thickness = ln(x-ray transmission) We used the above relation to compute the thickness of the small Al particles from their x-ray transmissions as calculated for the ideally sharp detector. The calculated thickness of the small Al particles were found to be very similar to the real Al thickness. Figure1 illustrates this match. The minimum and maximum calculated values had deviations of less than 11% of the nominal value for the 100µm thick particles and less than 8.5% of the nominal value for the 500µm thick particles. Fig. 1: Calculated thickness of Al particles with nominal thickness of 100 µm (left) and 500 µm (right). The horizontal axes represent the average diagonal size of the small Al particles as measured under a stereomicroscope. The solid lines show the nominal Al thickness SPIE USE, V (p.2 of 10) / Color: No / Format: A4/ AF: A4 / Date: :41:11

3 2.2. Description of imaging systems The procedure was tested on 2 system configurations: the GE Senographe 2000D (Milwaukee, USA) and the Agfa Embrace DM1000 (Mortsel, Belgium). Figure 2 illustrates the resolution and noise of the two detectors. Both detectors have an almost radial symmetric behavior. Therefore, we only plotted their vertical Normalized NPS (NNPS) and MTF. Both NNPS and MTF of these systems were measured under conditions that are as close as possible to those used in clinical practice for the acquisition of a standard breast. For the GE Senographe 2000D the NNPS was calculated from a flat field image of 45mm PMMA obtained at 28kVp, Rh/Rh under the Automatic Exposure Parameter (AOP) mode. The STD setting was chosen. For the Embrace DM1000, 45mm PMMA was exposed using a manual look-up table for the exposure as a function of breast thickness and composition. Twenty-eight kvp, Mo/Rh and 55.2 mas were used. The MTF was computed from an exposure of an edge sandwiched between 40mm PMMA. Table 1 summarizes the properties of the two detectors. With these settings, we calculated Mean Glandular doses of 1.24 mgy for the Senographe 2000D and 1.40mGy for the Embrace DM1000. Notwithstanding this 11.4% higher MGD for the Embrace DM1000, its NNPS performance is only better up to 1.5 lp/mm. At 5 lp/mm its NNPS is 2.3 x higher than the NNPS of the Senographe 2000D. This may influence detectability at the higher frequency levels. Fig. 2: The normalized NPS (left) and MTF (right) of the GE Senographe 2000D (gray ) and the Agfa Embrace DM1000 (black). Table 1: Properties of the Senographe 2000D and the Embrace DM1000 Senographe 2000D Embrace DM1000 Vendor GE Agfa Detector Type Indirect Direct Manufactor GE Lorad Detector Material a-si Se Digital Matrix 1914x x3328 Pixel Size [µm] Image Depth [bits] Acquisition of mammograms Thirty-five mammograms of patients with a compressed breast thickness of 4.2 +/- 0.5 cm were acquired with the GE Senographe 2000D. Thirty-three mammograms were acquired with the Agfa Embrace DM1000 in the same thickness class. The mammograms had been exposed with the routine clinically used parameters (i.e. 28 kvp and Rh/Rh for the SPIE USE, V (p.3 of 10) / Color: No / Format: A4/ AF: A4 / Date: :41:11

4 GE Senographe 2000D 28 kvp and Mo/Rh for the Agfa Embrace DM1000). The mammograms did not show any pathology Simulation of micro calcifications All micro calcifications were simulated starting from the x-ray transmissions of templates of real micro calcifications exposed at 27kVp, Mo/Mo on top of 40mm PMMA with a prototype of a CR plate of Agfa and using magnification (4). An overview of the different steps is shown in Figure 3. First, we computed the x-ray transmissions of the templates at 27kVp, Mo/Mo on top of 40mm PMMA as would be obtained for an ideally sharp digital detector. These templates were then expressed in their Al-equivalent thickness for the ideally sharp detector. The x-ray transmission properties of these (ideal) templates were then computed for the x-ray beam energy and compressed breast thickness under consideration in this study (4). We calculated the x-ray transmissions of the templates as if they had been exposed with an ideally sharp detector for 28 kvp, Rh/Rh (GE Senographe 2000D) and 28kVp Mo/Rh (Agfa Embrace DM1000) on top of 40 mm PMMA. To do so, radiographs of large Al sheets with different thicknessess on top of 40 mm at 28 kvp, Rh/Rh and 28kVp, Mo/Rh were acquired. The average x-ray transmissions of the different Al thicknesses were calculated from the raw images. Exponential fits were applied through the data points (Figure 4).. It shows the x-ray transmission as function of Al-thickness for an ideally sharp detector, as relatively large sheets of Al had been exposed. These relationships were used to adjust the templates of the ideally sharp digital detector to the exposure settings. Corrections for the differences in the spatial resolution were made using the method described in (4). Our templates of the ideally sharp detector had a pixel size of 50 µm (4). Therefore, the templates of the Embrace DM1000 (pixel size = 70µm) and the GE Senographe 2000D (pixel size = 100µm) had to be adapted for the difference in pixel size. The templates for both FFDM were finally corrected based on their pre-sampled MTFs. Figure 5 shows the median, 25- percentile and 75-percentile of the minimum x-ray transmissions of the final templates (i.e. after correction for beam energy and resolution characteristics of the 2 systems under consideration). x-ray Transmission of Template 27 kvp Mo/Mo 40 mm PMMA Prototype CR plate of Agfa x-ray Transmission of Template 27 kvp Mo/Mo 40 mm PMMA Ideally Sharp Digital Detector Al-equivalent of Template 27 kvp Mo/Mo 40 mm PMMA Ideally Sharp Digital Detector x-ray Transmission of Template X-ray Beam Energy Ideally Sharp Digital Detector x-ray Transmission of Template x-ray Beam Energy Digital Detector 2 Fig. 3: Overview of the different steps necessary to simulate a particular micro calcification. The procedure starts from templates acquired at 27 kvp, Mo/Mo on top of 40 mm PMMA with a prototype CR plate of Agfa (4). From these templates, new templates are calculated for a particular x-ray beam energy (i.e. 28kVp, Rh/Rh and 28kVp, MoRh on top of 40mm PMMA in this study) with a particular digital detector 2 (i.e. Senographe 2000D & Embrace DM1000 in this study). SPIE USE, V (p.4 of 10) / Color: No / Format: A4/ AF: A4 / Date: :41:11

5 Fig. 4: x-ray transmission as function of the Al-equivalent thickness of the large Al-sheets exposed on top of 40 mm PMMA at 27 kvp and Mo/Mo using the prototype CR plate of Agfa, at 28 kvp and Mo/Rh using the Agfa Embrace DM1000, at 28 kvp and Rh/Rh using the Senographe 2000D flat panel detector. The size of the simulated micro calcifications in the detector plane, expressed in equivalent diameter d eq ( d 2 # pixels eq = pixelsize ), as they would be obtained with an ideally sharp detector ranged from 200 µm to 800 µm. π The peak contrast, expressed in Al equivalent thickness, as they would be obtained with an ideally sharp detector, varied from 77µm to 800µm. In total 600 micro calcifications were simulated for each FFDM. Software phantoms were then created as follows: 0 to 8 simulated micro calcifications were randomly distributed in 2cmx2cm frames. They were then embedded in the linear data of the real mammograms in the parts with a constant breast thickness (5). In each mammogram, different types of background were chosen: fatty tissue, tissue that is a nearly homogeneous mixture of fat and glandular tissue and purely glandular tissue. Slow and medium varying anatomical variations were used. Different phantom configurations were made to avoid memorization. To the best of our knowledge, the phantoms were positioned in regions without any real micro calcification. Fig. 5: Minimum x-ray transmissions of the final templates, i.e. at 28 kvp, Rh/Rh corrected for spatial resolution of the Senographe 2000D (left) and at 28 kvp, Mo/Rh corrected for the spatial resolution of the Embrace DM1000 (right) for the different equivalent diameter groups as function of their peak Al-equivalent thickness that would be obtained in an ideally sharp digital detector. Error bars are 25-percentile and 75-percentile minimum x-ray transmissions. SPIE USE, V (p.5 of 10) / Color: No / Format: A4/ AF: A4 / Date: :41:11

6 The composed raw images were compressed, processed and printed as routinely obtained images. The Embrace DM1000 images were processed with MUSICA (6). The dedicated GE processing was used for the Senographe 2000D images. The Dryview 8610 (Eastman, Kodak) high resolution laser printer was used to print the GE, Senographe 2000D images. The Agfa Embrace DM1000 mammograms were printed with the high resolution Drystar 4500M (Agfa, Mortsel, Belgium) Observer experiment Three radiologists with experience in digital mammography viewed the full set of images. It was explained to the observers that each phantom contained between 0 to 8 micro calcifications. The observers were asked to indicate the locations of the micro calcifications and rate their confidence about their presence using a discrete 4-point scale (with 4=very likely to be a micro calcification, 1=probably not a micro calcification). There were no constraints on reading time and reading distance. The observers were encouraged to use a magnifying glass. Prior to the study, the observers first viewed a training set of 10 phantoms and got feed-back after rating these phantoms. A True Positive (TP) result was assumed when a detected micro calcification was within the contours of a simulated micro calcification. The False Negatives (FN) were given the score Data analysis The results of the observer experiments were analyzed using the AFROC methodology (7). The analyses were performed using ROCKIT (8). We calculated the areas under the AFROC curves A l which give a scalar summary of detectability. The confidence intervals on the area parameters were also computed. The analyses were performed for both systems, for all micro calcification sizes and for each individual observer. For each observer, the fractions of detected micro calcifications for each Al-equivalent thickness group were plotted along each equivalent diameter group. To do so, we used a binary score to calculate the fraction of detected micro calcifications for each of the 20 size groups. The scores less than or equal to 1 were set equally to 0. The scores 2, 3 and 4 were set equal to 1. The fraction of the binary score 1 was calculated for each of the 20 size groups. As AFROC analysis only takes account of the highest rated FP event in each phantom, and all other FP responses on that phantoms are neglected, we plotted the number of FP along with their scores (i.e. 1, 2, 3, 4) for each observer. 3. RESULTS The results of the AFROC can be found in table 2. No significant results are seen between the A l of the Senographe 2000D and the Embrace DM1000. Figure 6 summarizes for both FFDM the detected fraction of simulated micro calcifications for each Al-equivalent thickness along their equivalent diameter. For the micro calcifications with equivalent diameter >500µm and Alequivalent thickness of >400µm both detectors perform very similarly. For the micro calcifications with an equivalent diameter >500µm, and Al-equivalent thickness group > µm, the average detected fraction is overall larger with the Embrace DM1000: the detected fraction of micro calcifications is on average more than 50%. The Senographe 2000D detector wins for the > µm equivalent diameter group with Al-equivalent thickness >400µm. The detected fraction of the micro calcifications in the Senographe 2000D system is poor compared to the Embrace DM1000 for micro calcifications with an equivalent diameter > µm, independently of their Alequivalent thickness. Figure 7 shows the scores assigned to the FP. Each observer assigned less FP to the Senographe 2000D detector than to the Embrace DM1000: r1 assigned 3.03x less FP, r2 assigned 1.96x less FP and r3 assigned 1.14x less FP. SPIE USE, V (p.6 of 10) / Color: No / Format: A4/ AF: A4 / Date: :41:11

7 Fig. 6: Fractions of detected micro calcifications as a function of the equivalent diameter for the different Al-equivalent thicknessess. r1=reader 1, r2=reader 2, r3=reader3. Left curves represent the GE Senographe 2000D, Right curves represent the Agfa Embrace DM1000 SPIE USE, V (p.7 of 10) / Color: No / Format: A4/ AF: A4 / Date: :41:11

8 Fig. 7: Number of FP along their scores for each observer, r1= reader1, r2=reader2, r3=reader3. Left curves represent the GE Senographe 2000D, Right curves represent the Agfa Embrace DM1000 Table 2: A l and corresponding 95% confidence intervals (between brackets) under the AFROC curves for the Senographe 2000D and the Embrace DM1000 Senographe 2000D Embrace DM1000 Reader [0.465,0.558] [0.470,0.552] Reader [0.508,0.681] [0.525,0.654] Reader [0.485,0.569] [0.501,0.585] 4. DISCUSSION AND CONCLUSIONS Present paper compares the detectability of micro calcifications using simulated templates in 2 systems as they are operated in clinical routine today. The templates had been carefully constructed taking into account the detected differences in x-ray transmissions and differences in spatial characteristics of the FFDM. The observer performance test that we performed is very close to the clinical task for detection of micro calcifications. The observers had to scrutinize the images and locate the specific image regions that were suspicious for containing micro calcifications. Different from clinical practice, where it is searched for clusters, was the fact that the micro calcifications were distributed in uncorrelated locations. This restriction was made because AFROC analysis requires independent observations. The two different flat panel detectors under consideration have markedly different image physical characteristics at clinical used exposures. The Embrace DM1000 performs better for resolution (Figure 2) whereas the Senographe 2000D has superior noise characteristics for > 1.5 lp/mm (Figure 2). That are not the only parameters that determine which detector is better for clinical tasks. The clinical used processing parameters and viewing may also have their influence. The effect of processing and printing are included in this study. But, it was impossible to derive their impact on the visibility of micro calcifications. Comparison of different processing or viewing conditions would be straightforward with current procedure and is under study. In this preliminary study a conflict is seen between the overall AFROC results on one hand and the detected fraction for the different micro calcification sizes and the number of FP on the other hand. Whereas the AFROC results look to SPIE USE, V (p.8 of 10) / Color: No / Format: A4/ AF: A4 / Date: :41:11

9 promise no significant difference between both mammography systems, the performance of both systems is different when splitting up for the different micro calcifications sizes and when taking into account all indicated FP. We believe that the detected fraction for each micro calcification size and the number of the number of FP reflects much more the clinical reality than the AFROC results. We tried to perform AFROC analyses for the different micro calcification sizes but a too small statistical power for the analyses has been observed. The results of our observer study suggest that, despite the higher noise level and lower x-ray transmission of the micro calcifications, the Embrace DM1000 allows overall for greater detectability of micro calcifications than the Senographe 2000D. The higher resolution (pixel size + MTF) of the Embrace DM1000 very probably explains this performance difference for the greater part. The overall larger number of FP for the Embrace DM1000 may be due to the higher noise content of the Embrace DM1000. This indicates that readers were not sure about the true nature of the indicated artifacts, and would hesitate to interpret them as micro calcification in real practice. This preliminary study shows that there is a more or less constant detectability level for the micro calcifications of the two smallest equivalent diameter groups (> µm and > µm) in the Embrace DM1000 images. In the images of the Senographe 2000D, an increased detectability is observed for the micro calcifications of the > µm equivalent diameter group compared to these of the > µm equivalent diameter group (except for the smallest Alequivalent thickness group). The larger pixel size of the Senographe 2000D seems to dissolve the smallest micro calcifications in the background. This result could have been predicted by a simple viewing of the templates. Again, the clinically used processing parameters, printer and film for the two FFDM could also partly have determined the results. Further research of the separate parts of the imaging system, rather than a global analysis is needed here. Present study quantifies the sizes of detected micro calcifications for 2 FFDM system using physical methods, clinical mammograms and human reading. The results may be important and straightforward information for radiologists in charge of assessing the role of digital mammography in their (screening) practice or of comparing FFDM systems. It remains to be seen in larger studies whether clinical practice will confirm the present observations regarding the detectability of lesions. It goes without saying that also more different systems, such as computed radiography plates or slot scan systems, could be evaluated in the same way. Finally, noise, contrast and spatial resolution are the very fundamental properties with an influence on image quality. The digital nature of digital images adds post processing and viewing to this list. In this study, 2 detectors were involved with different properties and studied as they are used in the clinical routine. The detector has certainly a big influence on the final result. This preliminary study suggests that the influence of resolution on the detectability of small lesions like micro calcifications can compensate for a higher noise level. It must be recognized however that the chosen exposure settings determine image contrast and the dose level influences the noise in the images. Post processing can further enlarge or decrease the sharpness or the global appearance of the lesions. Further studies will unravel the complete link between the detectability of particular lesions and the noise, contrast and resolution properties of an imaging system and, in this way, guide the user towards optimal exposure settings. Present procedure may add complementary information to more established approaches. 5. ACKNOWLEDGMENT We would like to thank A. Chèvrement of GE for all practical help with the printing of the Senographe 2000D images. Octavian Dragusin is greatly acknowledged for his help with the acquisition of the results. This work was partly sponsored by DIMONDIII, an EC funded research project. REFERENCES 1. R.E. Van Engen, W.J.H. Veldkamp, L.J. Oostveen, M.A.O. Thijssen, N. Karssemeijer, Image Quality of a Dual Side Reading Fuji Computed Mammography System Compared to the GE FFDM System and a Film/Screen Mammography System, IWDM th International Workshop on Digital Mammography, p X.J. Rong, C.C. Shaw, D.A. Johnston, M.R. Lemacks, X. Liu, G.J. Whitman, M.J. Dryden, T.W. Stephens, S.K. Thompson, K.T. Krugh and C-J. Lai, Microcalcification detectability for four mammographic detectors: Flatpanel, CCD, CR and screen/film, Med.Phys., 29 (9), September 2002, SPIE USE, V (p.9 of 10) / Color: No / Format: A4/ AF: A4 / Date: :41:11

10 3. E. Samei, M. J. Flynn, and D. A. Reimann, A method for measuring the presampled MTF of digital radiographic systems using an edge test device, Med. Phys. 25, 102 (1998) 4. A.-K. Carton, H. Bosmans, C. Vanongeval, G. Souverijns, F. Rogge, A. Van Steen, G. Marchal, Development and validation of a simulation procedure to study the visibility of micro calcifications in digital mammograms, Med. Phys., 2003, August 2003, A.E. Burgess, Mammographic structure: Data preparation and spatial statistics analysis, Medical Imaging 1999: Image Processing, K. Hanson (ed.), Vol.3661, , SPIE, Bellingham WA P. Vuylsteke, P. Dewaele, E. Schoeters, "Optimizing Computed Radiography Imaging Performance", in "The Expanding Role of Medical Physics in Diagnostic Imaging", G.D. Frey, P. Sprawls. Eds., Proceedings of the 1997 AAPM Summer School, pp , D. P. Chakraborty, The FROC, AFROC and DROC variants of the ROC Analysis, Chapter 16 in Handbook of medical imaging. Vol. 1: Physics and Psychophysics. SPIE Editors: J. Beutel, H.L. Kundel, R.L. Van Metter 8. SPIE USE, V (p.10 of 10) / Color: No / Format: A4/ AF: A4 / Date: :41:11