Shortcomings of low-cost imaging systems for viewing computed radiographs

Computerized Medical Imaging and Graphics PERGAMON

Computerized Medical Imaging and Graphics 24 (2000) 25–32 www.elsevier.com/locate/compmedimag

Shortcomings of low-cost imaging systems for viewing computed radiographs J. Ricke*, E.L. Ha¨nninen, C. Zielinski, H. Amthauer, C. Stroszczynski, T. Liebig, M. Wolf, N. Hosten Department of Radiology Charite´, Campus Virchow-Klinikum, Humboldt-University Medical School, Augustenburgerplatz 1, 13353 Berlin, Germany Received 5 November 1998; accepted 5 October 1999

Abstract Objective: To assess potential advantages of a new PC-based viewing tool featuring image post-processing for viewing computed radiographs on low-cost hardware (PC) with a common display card and color monitor, and to evaluate the effect of using color versus monochrome monitors. Materials and Methods: Computed radiographs of a statistical phantom were viewed on a PC, with and without post-processing (spatial frequency and contrast processing), employing a monochrome or a color monitor. Findings were compared with the viewing on a radiological Workstation and evaluated with ROC analysis. Results: Image post-processing improved the perception of low-contrast details significantly irrespective of the monitor used. No significant difference in perception was observed between monochrome and color monitors. The review at the radiological Workstation was superior to the review done using the PC with image processing. Conclusion: Lower quality hardware (graphic card and monitor) used in low cost PCs negatively affects perception of low-contrast details in computed radiographs. In this situation, it is highly recommended to use spatial frequency and contrast processing. No significant quality gain has been observed for the high-end monochrome monitor compared to the color display. However, the color monitor was affected stronger by high ambient illumination. q 2000 Elsevier Science Ltd. All rights reserved. Keywords: Digital radiography; Image processing; Monitor review; ROC

1. Introduction Many digital hospital infrastructures use low-cost PCs to access electronic patient records. It would be desirable to use these ubiquitous machines for review of medical images as well [1]. However, display hardware and software may be of lower quality with respect to the radiological viewing task, compared to the expensive components used in dedicated radiological Workstations. No loss of diagnostic accuracy is suggested with the use of common display hardware and software for CT or MRI images if appropriate gray scaling is used [2,3]. Limitations exist for viewing computed radiographs at low-cost PCs if no contrast enhancing technique is implemented [4–6]. We seek to determine if the image post-processing, or the use of highend monochrome monitors, significantly improve diagnostic accuracy of low-cost PC review of computed radio* Corresponding author. Tel.: 149-30-450-57001; fax: 149-30-45057901. E-mail address: [email protected] (J. Ricke).

graphs. A viewing tool has been developed by our research group to enable PC-based contrast and spatial frequency processing of computed radiographs. Results of a ROC analysis are described, where viewing computed radiographs at a high-end radiological Workstation is compared to low-cost hardware and software with and without upgrades by image post-processing and a high-end display.

2. Materials and methods A ROC analysis employing digital radiographs was performed to compare low-cost to high-end viewing applications. A Siemens Magic View w featuring a Simomed 21inch monitor represented a high-end viewing tool. The monitor had a peak luminance of 600 cd/m 2. The luminance was 200 cd/m 2 for white, 0.2 cd/m 2 for black signals. The monitor resolution was 1000 × 1000 pixels. The Magic View employs Fuji’s CR image processing algorithm with gamma curve correction and unsharp mask filtering [7]. The

0895-6111/00/$ - see front matter q 2000 Elsevier Science Ltd. All rights reserved. PII: S0895-611 1(99)00036-1

26

J. Ricke et al. / Computerized Medical Imaging and Graphics 24 (2000) 25–32

Fig. 1. Settings for spatial frequency processing as proposed for review of computed chest radiographs on PC.

gamma curve is a mathematical function that modifies the incoming digital signal values to the luminance values displayed on the monitor. In effect, displayed gray shades are adapted according to the optical range of the radiograph. The number of displayed gray levels can be adjusted. This algorithm can also process spatial frequency. An unsharp mask (B) of the image (A) is generated, splitting the image into high and low spatial frequency components. The blurred low spatial frequency component of the image can be individually modified, by changing the size of the kernel used. The unsharp mask is subtracted from the original image (A 2 B), and the result is weighted ( f ) with a nonlinear function ‰g…A†Š: Finally, the product is added to the original image to create the processed image A 0 : A 0 ˆ A 1 fg…A†…A 2 B† Parameters of image processing for review with the Magic View correspond to standard settings for thorax images as used in clinical routine (Spatial frequency processing: RN 5; RE 0,16; RT F; convolution mode 2; kernel mode 2. Gamma curve correction: GT E; GA 1.04; GC 564.1; GS 279,9. Window/level: 1024/512). The image processing application for the low-cost view-

Fig. 2. Computed radiographs of the statistical phantom. (a) Original image; (b) with both gamma curve correction and spatial frequency processing. Five out of 10 disks in each column carried a critical detail (arrows). For ROC analysis, viewers estimated the probability of the existence of a critical detail in each individual disk of a predetermined column. Viewing was repeated using five different images presenting details in varying positions.

ing system underwent gradual technical improvements during the study. The basic system featured a 133 MHz Pentium PC with 32 MB RAM. The graphic card employed was a 4 MB Matrox Millenium w. Visualization of 10 bit computed radiographs was performed with window/level at 1024/512. In the first step, a standard color monitor (20-inch, Multiscan sfII w, Sony Corporation, Tokyo, Japan) was used. The maximum luminance is 100 ^ 20 cd/m 2. The luminance for the test setting was 100 cd/m 2 for white, 0.2 cd/m 2 for black signals. The convergence error was ,0.3 mm, resolution chosen was 1152 × 864 pixels. In the next application, a high-end monochrome monitor (21-inch, SMM 2183 L, Siemens, Erlangen, Germany) connecting to the identical low-cost PC graphic card was evaluated. Its peak luminance could be increased to 650 cd/m 2, the luminance used in this study was 200 cd/m 2 for white, 0.2 cd/m 2 for black signals. The resolution chosen was 1000 × 1000 pixels. Evaluation of both color and monochrome monitors was repeated, applying software that enabled gamma curve correction as well as unsharp mask filtering of DICOM

J. Ricke et al. / Computerized Medical Imaging and Graphics 24 (2000) 25–32

27

Fig. 3. ROC curves of review at the PC with monochrome and color monitor with and without image post-processing. In the co-ordinate system, the ordinate shows the TPF (true positive fraction), the abscissa the FPF (false positive fraction). Note that the curves for color or monochrome monitor along with image processing indicate higher true positive (and true negative) results than the curves for both monitors without image processing. Az is the area under curve defined by a and b (a: signal to noise ratio, calculated by the mean values of the gauss curve with and without detail; b: sum of s divided by sum of m, the standard deviation of the gauss curve with and without detail).

Fig. 4. ROC curves of review at a radiological Workstation compared to PC with a monochrome monitor with and without image post-processing.

28

J. Ricke et al. / Computerized Medical Imaging and Graphics 24 (2000) 25–32

Fig. 5. Thoracic computed radiographs showing a metastasis of a hepato-cellular carcinoma (arrow). (a) Original image; (b) with gamma curve correction; (c) with spatial frequency processing; (d) with both gamma curve correction and spatial frequency processing.

images on a Personal Computer. This software was developed by our research group to improve low-cost PC visibility of computed radiographs. Its algorithm is nearly identical to the algorithm employed in Fuji-based CR systems. To ensure optimal settings of the PC-based contrastprocessing algorithm, two experienced radiologists evaluated a set of 10 thorax images. Varying pathologic findings comprised of high-frequency details (five patients with interstitial disease) or low-frequency details (five patients with pulmonary carcinoma or metastatic disease) were visually examined and compared to image display on the Magic View high-end Workstation. The kernel size selected for PC review was 5 mm, the weighting function as shown in Fig. 1. Window/level were set at 1024/512. Five images of a low-contrast phantom were acquired with a storage phosphor radiography system (Digiscan 1102, Siemens, Erlangen, Germany). Radiography parameters were 125 kV, 5.6 mA s and 2 m film-focus-distance. A Perspex layer of 5 cm was used for radiation scattering and attenuation. The phantom (Alvim R&D Ltd, Jerusalem, Israel) consists of 12 vertical columns of 10 disks each (Fig. 2a and b) [8]. Five out of 10 disks carry a drill hole, simulating a critical detail. The size of the drill holes increases from 0.9 to 2.0 mm, left to right. The order of

the disks in each column was changed for each image viewed in this study. For ROC analysis [9], five images of the statistical phantom were presented to five similarly experienced Radiologists (3rd and 4th year residents) at the high-end radiological Workstation and at the low-cost PC with color or monochrome monitor, with and without additional contrast processing. Images were randomly presented under identical ambient luminance viewing conditions with no dark adaptation. The probability of seeing a critical detail in a specific position was expressed using a rating scale of five categories: 1 ˆ definitely positive; 2 ˆ probably positive; 3 ˆ equally positive or not positive; 4 ˆ probably negative; 5 ˆ definitely negative. No time limitation was applied for viewing. Statistical significance was calculated with Student’s t-test.

3. Results The area under curve (AUC) observed for review at the high-end Workstation with image post-processing was 0.87 ^ 0.02 (Fig. 3). The AUC observed for the PC with color monitor without image post-processing was

J. Ricke et al. / Computerized Medical Imaging and Graphics 24 (2000) 25–32

29

Fig. 6. Cranial computed radiographs showing an orbital fracture on the left (arrow). (a) Original image; (b) with gamma curve correction; (c) with spatial frequency processing; (d) with both gamma curve correction and spatial frequency processing.

0.82 ^ 0.03 (Fig. 4). For PC review with color monitor and image post-processing, the AUC was 0.85 ^ 0.03 (Fig. 4). Review at the PC with monochrome monitor without image post-processing revealed an AUC of 0.80 ^ 0.03 (Fig. 4). The AUC observed for PC review with monochrome monitor and image post-processing was 0.85 ^ 0.02 (Fig. 4). The Workstation viewing was significantly superior to the viewing of post-processed images at the low-cost PC with monochrome or color monitor …p , 0:05†: The Workstation viewing was superior to the PC viewing without image postprocessing, irrespective of the type of monitor used …p , 0:001†: For low-cost PC viewing, improved perception of low-contrast details using image post-processing was significant with color or monochrome monitor, respectively …p , 0:05†: No significant difference was observed in perception of low-contrast details when comparing PC viewing of color and monochrome monitors. This result was confirmed for viewing with and without image post-processing.

4. Discussion This study attempts to answer two questions. Firstly, does

image post-processing improve the diagnostic accuracy of low-cost imaging systems, and, if it does, to what extent? Secondly, does a high-end monochrome monitor significantly improve diagnostic accuracy compared to a standard color monitor? To determine improvement of visibility of low-contrast details by image post-processing for low-cost PCs, an algorithm similar to the software method applied in Fuji CR systems was developed for evaluation. Our study shows that image post-processing significantly improves the visibility of low-contrast details in computed radiographs from low-cost PC systems. Adapting the optical density to the image by adjusting the gamma curve improves image contrast. Unsharp masking enhances the sharpness of object contours (Fig. 5a–d). Current literature supports this result for radiographs of bone pathology and interstitial lung diseases as well (Fig. 6a and b) [10–13]. Contrary to the anticipated outcome, no significant differences were observed for low-cost PC viewing with a standard color monitor compared to a high-end monochrome monitor. In theory, two main factors limit the performance of the color monitor more than the monochrome monitor: less luminance and convergence errors of the color tube. The color monitor we used had three electron beams targeting red, green and blue, through an aperture grill pitch.

30

J. Ricke et al. / Computerized Medical Imaging and Graphics 24 (2000) 25–32

Fig. 7. ROC of viewing at a radiological Workstation and PC with color monitor and no contrast processing. Viewing had been undertaken with ambient light just slightly dimmed.

Thus, convergence errors increasing display noise are inevitable. Monochrome monitors do not have those convergence errors, since they employ just one electron beam to stimulate a phosphor layer, which emits luminance at a defined intensity. In both monitors, background noise comprises of temporal and spatial noise. Temporal noise refers to random fluctuations at the same screen location; spatial noise corresponds to the local and stationary graininess of the luminance output for constant input [14]. It should also be noted that the signal transfer curve, the convergence error or the uniformity of luminance might vary among, otherwise identical, displays due to the statistical deviation of analog display technology [15]. To study the effect of increased ambient illumination on monitor imaging, ROC curves observed during a preliminary study were determined. Except for just slight dimming of ambient lights, the study was identical to that used for review of phantom images at the high-end Workstation and the low-cost PC with color monitor and no image postprocessing. AUC for the Workstation was 0.86 ^ 0.02 (0.87 ^ 0.02 with moderate ambient illumination: no dark adaptation necessary). The AUC for the low-cost PC with color display and no image processing was 0.71 ^ 0.02 (0.82 ^ 0.03 with moderate ambient illumination) (Fig. 7). We conclude that increased ambient illumination affects viewing of the high-end monochrome monitor less than viewing of the standard color monitor. Viewing images from the high-end Workstation proved to be statistically superior to that of the low-cost PC irrespective of improvements with image post-processing or high-

end monochrome monitor. Three sources may potentially account for the consistent decrease in diagnostic accuracy of low-cost PC viewing: graphic card, monitor and image processing. Displaying properties of the monochrome monitors show advantages of the one used with the PC over the one used with the Workstation (i.e. maximal luminance). Since monitors supplied with radiological Workstations frequently do not work with PC graphic cards, an exchange of monitors between Workstation and PC could not be done. Hence, further evidence that the monochrome monitor has not been a factor decreasing diagnostic accuracy could not be provided and should be researched further. Secondly, it should be recognized that the image processing algorithms in the Workstation and in the PC were not identical. Whereas the original Fuji algorithm delivers a beta curve that is sigma shaped in part, it is discontinuously linear in the PC application (Fig. 1). The gradient of the gamma curve used in the PC is minimally higher than the one used in the Workstation. However, previous studies have shown no significance of discrete deviations in spatial frequency or contrast processing, respectively [16,17]. By exclusion of monitor or image post-processing we suggest that the graphic card was the main cause for the decreased diagnostic accuracy of low-cost PC viewing. A graphic card transforms digital pixel values to appropriate voltage values for signal intensity. The performance of a graphic card is indicated first by the maximum value of the signal output in proportion to the incoming signal; second by the gradient of the signal onset. Studies testing the performance of PC graphic cards have shown that variations of the expected

J. Ricke et al. / Computerized Medical Imaging and Graphics 24 (2000) 25–32

signal output may range from 35 to 100% between low-cost and high-end graphic cards [18]. Furthermore, the CE-standard (Commission Europe´enne) requires extensive damping of high frequencies in a monitor signal. According to the same studies, less expensive graphic cards commonly used in low-cost PCs may contribute to sub-optimal filter functions [18]. In conclusion, contrast and spatial frequency processing increase diagnostic accuracy of viewing computed radiographs using low-cost PCs. A high-end monochrome monitor did not improve the perception of low-contrast details under standardized test conditions with moderate ambient lighting. However, high-end monochrome monitors show improvements under low ambient illumination for image viewing outside radiological departments, such as on wards or in outpatient clinics. Our results suggest that the standard graphic card employed is a major factor causing decreased image quality of the low-cost PC when compared to the more expensive Workstation.

5. Summary The study described herein attempts to answer two questions. Firstly, does image post-processing improve the diagnostic accuracy of low-cost imaging systems, and, if it does, to what extent? Secondly, does a high-end monochrome monitor significantly improve diagnostic accuracy compared to a standard color monitor? To determine improvement of visibility of low-contrast details by image post-processing for low-cost PCs, an algorithm similar to the software method applied in Fuji CR systems was developed for evaluation. Computed radiographs of a statistical phantom were viewed on a PC with and without postprocessing (spatial frequency and contrast processing) employing a monochrome or a color monitor. Findings were compared to viewing on a high-end radiological Workstation and evaluated with ROC analysis. Our study shows that image post-processing significantly improves the visibility of low-contrast details in computed radiographs from low-cost PC systems. Adapting the optical density to the image by adjusting the gamma curve improves image contrast. Unsharp masking enhances the sharpness of object contours. No significant differences were observed for low-cost PC viewing with a standard color monitor compared to a high-end monochrome monitor. Increased ambient illumination affected viewing of the high-end monochrome monitor less than viewing of the standard color monitor. Viewing images from the high-end Workstation proved to be statistically superior to that of the low-cost PC, irrespective of improvements with image post-processing or highend monochrome monitor. In conclusion, lower quality hardware (graphic card and monitor) used in low cost PCs negatively affects perception of low-contrast details in computed radiographs. In this

31

situation, it is highly recommended to use spatial frequency and contrast processing. With optimal ambient illumination, the high-end monochrome monitor offers no advantages over the color display.

Acknowledgements The authors would like to acknowledge the support of the monitor study by R. Bra¨uer and E. Litzler from Siemens AG, Karlsruhe, Germany.

References [1] Gillespy T, Rowberg AH. Radiological images on personal computers: introduction and fundamental principles of digital images. J Digit Imaging 1993;6:7–81. [2] Barnes JE. Characteristics and control of contrast in CT. Radiographics 1992;12:37–825. [3] Ricke J, Wolf M, Hosten N, et al. How accurate is telediagnosis for tomographic images? Fortschr Ro¨ntgenstr 1997;166:66–70. [4] Ishida M, Doi K, Loo LN, et al. Digital image processing: effect on the detectability of simulated low contrast radiographic patterns. Radiology 1994;150:569–75. [5] Balter S. Fundamental properties of digital images. Radiographics 1993;13:41–129. [6] Ricke J, Wolf M, Zielinski C, et al. Digital radiology: ROC-Analysis of diagnostic accuracy in reviewing thorax images of digital luminance radiography on film, medical workstation or standard PC. In: Lemke H, Inamura K, Vannier M, et al., editors. Proceedings of CAR, Amsterdam: Elsevier, 1997. p. 82–6. [7] Prokop M, Schaefer-Prokop CM. Digital image processing. Eur Radiol 1997;7(3):S73–82. [8] Gurvich V, Wolf M. Introducing a phantom to determine radiation dose in diagnostic radiology. In: Schmidt T, Stieve FE, editors. Digitale Bildgebung in der diagnostischen Radiologie, Bildqualita¨t-Strahlen exposition, Berlin: H. Hoffmann, 1996. p. 287–92 (in German). [9] Swets J, Pickett R. Evaluation of diagnostic systems. New York: Academic Press, 1982. [10] Busch HP. Digital radiography for clinical applications. Eur Radiol 1997;7(3):S66–72. [11] Freedman M, Steller D. Digital radiography of the musculoskeletal system: the optimal image. J Digit Imaging 1995;8(1):37–42. [12] Muller RD, Hirche H, Voss M, Buddenbrock B, John V, Gocke P. ROC analysis in post-processing of image data in digital thoracic radiography. Fortschr Rontgenstr 1995;162:163–9 in German. [13] Prokop M, Schaefer C, Oestmann JW, Galanski M. Improved parameters for unsharp mask filtering of digital chest radiographs. Radiology 1993;187:521–6. [14] Blume H, Roehrig H, Ji TL. High-resolution, high-brightness crt display systems: up-date on state of the art. SID 94 Digest 1994:219–22. [15] Roehrig R, Dallas W, Ji TL, et al. Physical evaluation of CRTs for use in digital radiology. Proc SPIE 1989;1091:262–78. [16] Prokop M, Galanski M, Oestmann JW, et al. Storage phosphor versus screen-film radiography: effect of varying exposure parameters and unsharp mask filtering on the detectability of bone defects. Radiology 1990;177:109–13. [17] Loo LN, Doi K, Metz CE. Investigation of basic imaging properties in digital radiography. 4. Effect of unsharp masking on the detectability of simple patterns. Med Phys 1985;12:209–14. [18] Bertuch M. The graphic card: cause of unsharp monitors. C’t Magazin fu¨r Computer und Technik 1996;9:272–8.

32

J. Ricke et al. / Computerized Medical Imaging and Graphics 24 (2000) 25–32

Jens Ricke MD is a fellow in radiology. He is head of the Digital Radiology Working Group at the Charite´ Virchow-Klinikum.

Christian Stroszczynski MD is a radiological resident.

Thomas Liebig MD is a radiological resident. Enrique Lopez Ha¨nninen MD is a radiological resident.

Christoph Zielinski is a computer scientist.

Martin Wolf PhD is a physicist, with special interest in both medical image processing and medical image display.

Holger Amthauer MD is a radiological resident.

Norbert Hosten MD is Associated Professor of Radiology.