Reimann et al: Perceptually Linear Display Devices 11

Reimann et al: Perceptually Linear Display Devices 11 [10] H.R. Blackwell. Contrast threshold of the human eye. Journal of the Optical Society of America, 36:624{643, November 1946. [11] A.A. Michelson. Studies in Optics. University of Chicago Press, Chicago, Illinois, 1927. [12] Eli Peli. Contrast in complex images. Journal of the Optical Society of America, 7(10):2032{2040, October 1990. [13] M. Ibrahim Sezan, Kwok-Leung Yip, and Scott Daly. Uniform perceptual quantization: Applications to digital radiography. IEEE Transactions on Systems, Man, and Cybernetics, SMC-17(4):662{634, 1987. [14] J.G. Rogers and W.L. Carel. Development of design criteria for sensor display devices. Technical Report HAC C6619, Hughes Aircraft Company, Culver City, California, December 1973. [15] Peter G.J. Barten. Physical model for the contrast sensitivity of the human eye. In Proceedings of the SPIE, volume 1666 of Human Vision, Visual Processing, and Digital Display III, pages 57{72, 1992. [16] Peter G.J. Barten. Spatio-temporal model for the contrast sensitivity of the human eye and its temporal aspects. In Proceedings of the SPIE, volume 1913, pages 2{14, 1993. [17] Scott Daly. Digital Images and Human Vision, chapter The Visible Dierences Predictor: An Algorithm for the Assesment of Image Fidelity, pages 179{206. The MIT Press, 1993. [18] Scott Daly. The visible dierences predictor: an algorithm for the assesment of image delity. In Proceedings of SPIE, volume 1666 of Human Vision, Visual Processing, and Didital Display III, 1992.

Reimann et al: Perceptually Linear Display Devices 10 quality is easily recognized. Typically, signicant improvement is rst noticed in regions of low luminance. However, careful inspection reveals changes in all regions with increased contrast at both high and low luminance and decreased contrast at intermediate luminance. Selection of the best human visual model and associated parameters for use in specic medical applications is a subject requiring further work. Notably, the features of pathologic lesions in medical images are quite dierent that the test patterns used to measure contrast detection thresholds. In our limited experience, we have preferred use of the models which predict slightly higher contrast thesholds at the high and low luminance values associated with typical CRT devices. Use of relatively small test pattern sizes (2 degrees) and high frequency (5 cycles/degree) provides more low luminance contrast improvement which appears to be benecial. For the tone manager implementation we have developed, the specication of model and parameters can be done by the user and grayscales easily recalculated. This facilitates an understanding of the eect of these parameters when compared to implementations which precompute the LUT. A large number of companies now sell workstations for use in medical imaging and many of these use X window environments for graphic display. These systems use a variety of color or monochrome CRT monitors with wide variation in their display luminance response. We believe that those centers using multiple systems which dier in ther model or their age are likely to have very inconsistent display grayscales on the dierent systems which could cause signicant variation in diagnosis. We urge manufacturers to explore the use of a common grayscale with human visual foundations and encourage diagnostic centers to seek perceptual linearity in their display systems. 8. Acknowledgements This work was supported in part by the National Science Foundation (BCS-9315580) and by Eastman Kodak Company. 9. References [1] Society of Motion Picture and Television Engineers. Smpte recommended practice specications for medical diagnostic imaging test pattern for television monitors and hard copy recording cameras. Journal of the SMPTE, RP 133:693{695, June 1986. [2] J.E. Gray, K.G. Lisk, D.H. Haddick, J.H. Harshbarger, A Oosterhof, and R. Schwenker. Test pattern for video displays and hardcopy cameras. Radiology, 154:519{527, 1985. [3] James Foley and Andries Van Dam. Fundamentals of interactive Computer Graphics, chapter Intensity and Color, pages 593{623. Addison-Wesley system Programming Series, 1984. [4] Hans Roehrig, Hartwig Blume, Ting Lan Ji, and M.K. Sundareshan. Noise of CRT display systems. In Image Capture, Formatting, and Display, Prcoceedings of the SPIE 1897, pages 232{245, 1993. [5] R.E. Johnston, J.B. Zimmerman, D.C. Rogers, and S.M. Pizer. Perceptual standarization. Proceedings of the SPIE 536, pages 44{49, 1985. [6] Hartwig Blume, H. Roehrig, M. Browne, and T.-L. Ji. Comparison of the physical performance of high resolution CRT displays and lm recorded by laser image printers and displayed on light-boxes and the need for a display standard. In Proceedings of the SPIE 1232, pages 97{144, 1990. [7] T. L. Ji, H. Roehrig, H. Blume, and J. Guillen. Optimizing the display for function of display devices. Proceedings of the SPIE 1653, pages 126{139, 1992. [8] Hartwig Blume, Scott Daly, and Edward Muka. Presentation of medical images on CRT displays: A renewed proposal for a display function standard. In Proceedings of the SPIE 1897, 1993. [9] B. M. Hemminger and R. E. Johnston. Perceptual linearization of video display monitors for medical image presentation. Paper 2164-23 at the SPIE Medical Imaging 1994 Conference.

Reimann et al: Perceptually Linear Display Devices 9 To evaluate the delity of a particular X display server, we have developed a test pattern which operates in conjunction with the tone manager. An independent application, called hvspat in our facility, accesses the default color map established by the tone manager in the display server and creates a image for which all tone manager values are presented from left to right with increasing luminance. For each tone manager value, the image contains three pixels of equal intensity. Thus for a typical display with.3 mm pixels and 200 tone manager values, the test pattern has an 18 cm horizontal dimension. In the vertical direction, the display intensity is modulated in a ten cycle bar pattern with six groups having contrast of 1, 2, 4, 8, 16, and 32 tone manager values (see Figure 11). The frequency of the pattern is chosen to be as close to 1 cycle/mm as possible depending on the display pixel size. Typically, about two pixels are at the higher luminance and two pixels at the lower luminance for one cycle of the pattern. The vertical dimension of the full test pattern is thus about 6 cm. Contrast L 1 Cycle/mm 10 Cycles L+1, L Luminance 1 Cycle/mm 10 Cycles L+1, L-1 1 Cycle/mm 10 Cycles L+2, L-2 1 Cycle/mm 10 Cycles L+4, L-4 Figure 11. The test pattern has proven to be a useful tool to evaluate and to demonstrate the eect of perceptually linear tone scales on low contrast detection over the full luminance scale. The tone manager can be switched between a calibrated and a non calibrated scale by striking scale selection keys and the dierence in the test pattern observed immediately. This is of value for understanding the eect of dierent human visual parameter on display quality; however, the test pattern is very dierent than most medical images. The most important value of the test pattern is in identifying potential limitations in small contrast depiction that come from the display hardware. Nonuniform, periodic contrast has been observed in the low contrast regions of the test pattern. These artifacts are believed to be due to the limited quantization of the intensity value by the DAC as the intensity is continuously changed through the full range of values. 7. Discussion We currently have 8 software applications running of 15 host computers which use the tone mgr application. Approximately 30 display servers are in use on our network. Our experience is using the default colormap in a display server as a common grayscale LUT for many applications has been very favorable. The recent addition of perceptually linear grayscales to this application have been a signicant enhancement. Our experience indicates that individual devices can be quickly calibrated and periodically recalibrated as a part of quality control activities. The attributes of the X window system with respect to unique screen identiers and color management provide excellent mechanisms for managing image display devices on a network. The use of grayscales computed such that the observer perceives equivalent contrast at all levels of brightness provides signicantly improved image display. Using the tools we have reported in this paper, any observer can easily change from a conventional grayscale to a scale computed on the basis of a human visual model. The improved display

Reimann et al: Perceptually Linear Display Devices 8 Host Computer 1 tone_mgr Ethernet Network Host Computer 2 Table Server Host Computer 3 Imaging Apps Figure 10. The tone mgr application allows customization at startup of display specic parameters. The maximum number of color table colors to allocate is specied, however this number may be reduced if other applications are running. In this case, all remaining color entries are used. Individual applications may interact with the tone scale manager to specify an upper and lower level which can provide global contrast adjustments to all applications. When the command cursor is directed to the tone manager window, key commands can be used to select the desired scale (calibrated and uncalibrated gray scales, and various color scales) or restore the upper and lower level to a normal value. The selection of a perceptually linear tone scale is provided for X display screen monitors which have measured calibration data les. After calibration data les are graphically veried, they are place in a directory used by a centralized network table server. The table server is queried by the tone mgr application upon startup for the luminance table corresponding to the display device being used. The monitor response to an input command level is retreived from the table server, a polynomial is t to the measured data, the perceptually linear response for the measure luminance range is calculated. The LUT is then loaded with values producing a perceptually linear response for that particular CRT and specied visual model. The parameters for the visual model to be used are specied, along with model dependent variables, in les located in users home directories and read when the tone mgr application is executed. The tone mgr application thus allows runtime customization of the tonescale, and visual model. Keystroke provision is provided so that a user may change a human visual model parameter and generate a new perceptually linear grayscale. Computation of a new LUT for with new human visual model parameters can be done in about a second. 6. Display quality evalutation The determination of the desired X window (r,g,b) value using these methods and the 16 bit precision with which the (r,g,b) values are stored in the X window environment will produce a high delity display with respect to small contrast changes only if the display monitor has low noise characteristics. Fixed pattern noise in the CRT phosphor may degrade small contrast detection [4] and, in general, cannot be uniformity corrected. Additionally, noise may be present in the electronic circuits which control electron beam current. For X window display servers, the nal analog voltage sent to the CRT is created by a digital to analog converter (DAC). For most available X terminals and host computer display boards, the DAC has an 8 bit precision. Typically, 10 precision is needed to establish the required number of just noticeable gray values and the dierential luminance between any two values must be continuous, free of noise, and proportional to the desired contrast threshold [9]. Thus high delity X window display servers need at least a 12 bit DAC to achieve a perceptually linear tone scale.

Reimann et al: Perceptually Linear Display Devices 7 Figure 8. subset of colorcells from the default X colormap. This subset forms the LUT used to transform image gray level into displayed gray level. Our research laboratory has approximately 30 image display stations which use various X Window based image display and analysis software. The imaging applications we have developed all coordinate gray scale selection and window/level operations through the tone mgr application. Typically about 200 gray values are allocated for image display by the tone scale manager application. This allows 56 free colors for other applications. An image is mapped onto the range of 200 gray levels by assigning each pixel value to the appropriate index in the color table. The color table controls the input signal to the CRT and can be dynamically updated. Parameters which control the scale are stored on the X server as atoms on the root window manager, and are accesible by all client applications regardless of the host computer where they are executing. An example of the tone mgr application and 2 other imaging applications are shown in Figure 9. Figure 10 depicts the logical communications among the applications. Figure 9.

Reimann et al: Perceptually Linear Display Devices 6 The total number of just noticeably dierent display values, J = total JNDs, is often used as a display gure of merit which indicates the number visible graylevels that a particular display device can support for it's particular values of L min and L max. Figure 6 illustrates perceptually linear display luminance response curves constructed in this fashion for a display monitor having L min = 0:6 cd/mm 2 and L max = 134 cd/mm 2. The curves are similar for the three human visual models and result in about 900 total JNDs. This suggests that 10 bit grayscales having 1024 levels at specic luminance values are needed for this luminance range. 4. Display Luminance Measures The perceptually linear model requires perceptually equivalent luminance steps for unit changes of any display value. The desired luminance change is determined from the models discussed in the previous section. The luminance of the display monitor as a function of input signal level can be measured for a small number of input levels and interpolated over the entire range. It is useful to consider the inverse function, namely X command level as a function of luminance. The composition (Figure 7) of the desired perceptually linear response functon and the measured display luminance function can give the desired display value to X command level relationship required to display images with a uniform contrast response. PERCEPTUALLY LINEAR MONITOR LUMINANCE RESPONSE MEASURED MONITOR LUMINANCE RESPONSE Log L Log L Figure 7. I JND X Value We have written a program, xlum, to present 21 dierent input signal levels, from zero to full signal, in order to characterize the response of all display servers on our network. The contrast and brightness controls are initially adjusted with the SMPTE test pattern and, when possible, access to these controls restricted. Gray levels corresponding to 21 specic X command levels are then generated in a uniform eld about 8 2 8 cm in size. A luminance meter (International Light 1400A) with a cylindrical sensor (SEL 033) specically designed for CRT devices is used to measure luminance (see Figure 8). The measured luminance values for each of the 21 levels are entered to the xlum window and stored along with the X command value in a data le. Files are named using the unique X designation for a display device to insure proper association of the calibration data with devices. The curve is plotted and a parameter corresponding to the rst input signal value for which a change in luminance is observed is entered into the data le. The measurement procedure is relatively simple, requiring about 5 minutes. 5. Network Tone Manager Application We have previously developed an X application, tone mgr (tone manager), which allows multiple image display programs to share a colormap with dynamic LUT modication. Once the tone mgr application is running on an X display, multiple imaging applications from dierent computer systems can communicate their color requirements to the tone mgr application and simultaneously display images. Several color scales are provided and are user selectable at startup, or by key selection once the application is running. The tone mgr application allocates a

Reimann et al: Perceptually Linear Display Devices 5 model at high luminance. Daly's normalization for peak contrast sensitivity, P, has been set to 300. This is between the value of 250 used by Daly [18, 17] and within the range of 255-420 considered by Blume [8]. 300 has been used to obtain consistent results with all models. The contrast threshold is strongly dependent on frequency with minima in the range of 2 to 10 cycles/degree and signicant increase at higher frequency. Figure 5 plots the value of C t for three dierent spatial frequencies of 1, 4, and 10 and a test pattern size of 2 degrees. The signicant increase in C t for high frequencies and low luminance suggest the merit of image lters which selectively enhance high frequencies in regions of an image which have low luminance. C t is also notably inuenced by the size of the test pattern and to a lesser extent by pupil size, viewing distance, light adaptation or background luminance, and orientation. The test conditions contributing to C t involve test patterns that are quite dierent than the patterns observed in medical radiographs. Little work has been done to establish which human visual parameters best represent a medical radiograph and therefore what relationship for C t versus luminance should be used to construct a perceptually linear tone scale. We have obtained good preliminary results by using a test pattern size of 2 degrees and a spatial frequency of 5 cycles/degree. A perceptually linear tone scale is constructed by establish a luminance response such that a unit change in the image display intensity value produces a change in luminance which makes 1L=L just noticeable. A \just noticeable dierence" or JND is assumed to exist if 1L=L is equal to C t. Beginning with the minimum luminance, a luminance response can then be constructed numerically as: L i = ( Lmin ; if i = 0; L i01 + L i01 2 C t (L i01); if 0 < i < J; L max ; if i = J. MODELS FOR PERCEPTUALLY LINEAR RESPONSE 100 Log(L) 10 "Rogers and Carel" "Barten" "Daly" 1 0 100 200 300 400 500 600 700 800 900 1000 Ijnd Figure 6.

Reimann et al: Perceptually Linear Display Devices 4 0.016 0.014 0.012 MODELS FOR VISUAL CONTRAST DETECTION THRESHOLD Display : xds6:0.0 HVS Model Rogers and Carel Barten Daly Contrast Threshold 0.01 0.008 0.006 0.004 0.002 X0 = Y0 = 2 deg., U = 4 c/deg. 1 10 100 Log(L) Figure 4. Contrast Threshold 0.04 0.035 0.03 0.025 0.02 0.015 HUMAN VISUAL CONTRAST DETECTION THRESHOLD Display : xds6:0.0 Daly HVS Model Frequency 1 c/deg. 4 c/deg. 10 c/deg. 0.01 0.005 0 X0 = Y0 = 2 deg. 1 10 100 Log(L) Figure 5.

Reimann et al: Perceptually Linear Display Devices 3 The use of the default colormap aords fewer colors for the application, but no sudden color change or \colormap ash" occurs when going between windows. Furthermore, applications can share colorcells in the default colormap without having to contend for colors. Thus many applications can share a larger number of gray levels. We have developed an application to manage the shared colormap among applications running on a particular X display. This application, tone mgr, is described in section 5. 3. Perceptually Linear Tone Scale A perceptually linear tone scale is a relationship between display luminance and image display value such that the perceived contrast associated with a small change in display value is constant for all luminance values between the minimum (i.e. dark elds) and the maximum (i.e. bright elds). The generation of a perceptually linear tone scale is usually done by considering the human visual response to small changes in luminance as measured with specic test patterns. In this section, we briey review models for human visual contrast detection and the use of these models to construct a tone scale. The ability of human observers to detect contrast in sine wave test pattern was reported in 1946 [10] and subsequently by numerous investigators. Within the test pattern, luminance is varied with a spatial sine wave pattern of a specic frequency and orientation (see Figure 3). The test pattern size subtends a specic angular eld of view at a specied viewing distance and is surrounded by a background luminance which may be dierent that the average luminance of the test pattern. The contrast threshold, C t, has been reported either as (L max 0 L min )=(L max + L min ) which is the Michelson contrast [11, 12], or as (L max 0 L min )=(L mean ) = 1L=L which is sometimes called the Weber fraction [13]. For this work, we have taken C t to be 1L=L. Figure 3. Rogers and Carel developed a polynomial t to experimentally measured data which accounts for the test pattern frequency, luminance, size and the background luminance [14]. We have used in this work the Rogers and Carel formula reported by Blume [8]. Barten has reported a physical model for both the spatial and the temporal response of the human eye which agrees with extensive experimental data [15, 16]. We use in this work the Barten spatial response model for C t and it's dependence on test pattern frequency, luminance and size, along with it's dependence on pupil size. Daly has reported a dierent human visual system model [17, 18] which we have implemented with it's dependence on test pattern frequency, luminance and size, along with it's dependence on orientation, light adaptation level, and viewing distance. Illustrated in Figure 4 is the value of C t versus luminance predicted by these three models for a test pattern subtending 2 degrees with a frequency of 4 cycles/degree. The three models produce very similar results with only a slight reduction in C t for the Daly model at low luminance and a slight reduction in C t for the Rogers and Carel

Reimann et al: Perceptually Linear Display Devices 2 xds1:0.0 xds2:0.0 Host Computer 1 X Clients Ethernet Network xds3:0.0 xds3:0.1 Host Computer 2 X Clients Figure 1. or more display screens. A key concept of X terminals is a uniquely identiable display screen name. For example, the name id.rad.hfh.edu:1.2 corresponds to screen 2 of display 1 of the display server id.rad.hfh.edu. The color of displayed information is easily selected with X. Each screen has a default colormap and application specic private colormaps. Colorcells in the colormap control the color of that particular colorcell. In early versions of X, the color was controlled by specifying a triplet (r,g,b) of red, green, and blue values. Newer versions (R5 and newer) allow specication using the CIE XYZ color model, however the (r,g,b) model is still supported. These (r,g,b) values are each typically specied by 16 bits, resulting in a possible 2 48 distinct colors. The number of simultaneously displayable values, and hence the colormap, is often much less, typically 256. The size of the colormap and the number of possible colors depends on the X display server hardware. For example, in a grayscale system with an 8 bit display, only one of the (r,g,b) triple values is used, so only 256 of the possible 65536 graylevels are simultaneously available. Several layers of abstraction hide X display server hardware from the imaging application. To display an image, one needs to perform several steps schematically shown in Figure 2. Digital image values are mapped to the display tonescale values with typically fewer distinct levels than in the input image. This mapped image is converted from the tonescale value to an X (r,g,b) color with a look up table (LUT). The (r,g,b) is converted to an analog signal driving the monitor to create a particular luminance. Once an image is loaded, the LUT can be changed rapidly. This facilitates dynamic modication, including fast window and leveling operations. Input Image Tone Manager Colormap Display A B C Gray Level Index (R,G,B) DAC Luminance R G B A B C min-max 0-N (0-64K,0-64K,0-64K) Color Figure 2.

A System to Maintain Perceptually Linear Networked Display Devices David A. Reimann, Michael J. Flynn, and James J. Ciarelli X-Ray Imaging Research Laboratory, Department of Radiology, Henry Ford Health System, Detroit, MI 48202 ABSTRACT The CRT is commonly used to display digital medical image data. The brightness as a function of input signal is dierent for each CRT, is nonlinear and poorly matched to the human eye's perception of brightness change, decreases over time, and is easily changed with hardware contrast and brightness adjustments. Previous studies have suggested display brightness functions which are based on human visual experiments and which produce the same contrast perception for small display value changes at all brightness levels. The best scale depends on CRT luminance and scene dependent variables. We have developed an X Window based client server approach to maintain perceptually linear display scales using unique transformation tables for many display devices. Color use on X window systems is described. Perceptually linear human visual models are described. Finally we present a method to implement and maintain these models on a networked collection of X displays. 1. INTRODUCTION Cathode Ray Tubes (CRT) used as display screens on workstations are increasingly being used to display medical images, particularly in computed tomography and magnetic resonance imaging where primary diagnosis may be performed using electronically displayed images. The grayscale of these devices varies signicantly for dierent systems, is easily changed by external brightness and contrast controls, and changes with time. The SMPTE medical imaging test pattern [1] was specically developed as an aid in setting up CRT displays [2]. CRT contrast and brightness settings are now frequently established by use of this pattern. The display luminance response of a particular system is often poorly matched to the manner in which human observers perceive contrast in visual patterns. When viewing medical images or the SMPTE test pattern, contrast in low luminance regions is typically poor. For most CRT devices the luminance (brightness) of the display surface is related to the display value and video voltage by the relationship L = cv [3, 4], where gamma can vary from about 1.5 to 2.3. Signicant improvements is display quality can be obtained by changing the display luminance response to a perceptually linear luminance response establishing by using data on human visual response [5, 6, 7, 8, 9]. In this paper, we report methods for measuring the response of network display devices and maintaining response data in a centralized server for use in establishing perceptually linear response. A tone scale manager is used to provide the desired display scale on a network device for many applications which may be running in multiple network host computers. The methods previously reported for perceptual linearization are supported for three models of human visual performance and for various human visual dependencies. 2. X Window System The X Window System, or simply X, provides an extensible vendor neutral platform for graphical interaction. The most current version is X11R6, denoting version 6 of release 11. One extension available in R6 is XIE, the X Image Extension. The X windowing system is available for most workstations, personal computers, and on dedicated hardware X terminals. This direct support for image display has made X a natural choice for high performance imaging applications, including medical imaging. In this section, we review aspects of X relevant to image grayscales. A typical X display server consists of a keyboard, mouse, CRT, and a control unit connected via a communications channel (typically ethernet for an X Terminal, or socket for a Unix workstation) to a host computer. An X window system consists of client applications running on remote hosts, which communicate with an X display server which services the graphical requests of the client application, as shown in Figure 1. Each X display server consists of one