Preliminary Assessment of High Dynamic Range Displays for Pathology Detection Tasks CIS/Kodak New Collaborative Proposal CO-PI: Karl G. Baum, Center for Imaging Science, Post Doctoral Researcher CO-PI: Maria Helguera, Center for Imaging Science, Assistant Professor Also Involved: James A. Ferwerda, Center for Imaging Science, Associate Professor Kent M. Ogden, Department of Radiology, SUNY Upstate Medical University, Associate Professor Abstract A two-alternative forced choice experiment was conducted to determine if diagnostic benefits exist when using a high dynamic range display for a lesion detection task. The potential benefits resulting from both an increased dynamic range and an increased number of displayable just noticeable grayscale differences will be evaluated. A custombuilt high dynamic range display system, constructed at the Munsell Color Science Laboratory using an LCD panel and DLP projector, was used. The display, which provides a much larger contrast ratio than available with typical medical LCDs, was calibrated using the DICOM standard and its performance compared with that of an LCD.
Introduction Liquid crystal displays (LCD) have replaced film as the display of choice in radiologic imaging. Use of LCD displays and associated digital images provide conveniences that film can not, such as improved workflow efficiency and image enhancement options. However, film viewed over a light box in many cases still has a diagnostic advantage due to a larger dynamic range. Traditional light boxes have a peak brightness of around 4000 cd/m 2, providing a 3000:1 contrast ratio (maximum brightness divided by black offset) 1. This is well in excess of the 600-900 cd/m 2 and 600:1 contrast ratio of diagnostic LCD displays. For this reason it is believed that a display with an increased dynamic range and an increase in the number of perceivably different grey levels (just noticeable differences JND) may provide diagnostic benefits. To evaluate this hypothesis, a custom high dynamic range (HDR) display was calibrated to the DICOM standard and two-alternative forced choice (2-AFC) studies were conducted to determine if the HDR display provides an advantage for pathology detection. Methods HDR System The Munsell Color Science Laboratory (MCSL) at the Center for Imaging Science, Rochester Institute of Technology has developed an HDR display with a dynamic range of five orders of magnitude commensurate with the fully adapted human visual system and the highest gamut volume possible. The MCSL HDR system (see Figure 1) includes a Plus U5-232 DLP projector (2000 Lumens, 2000:1 contrast ratio, XGA (1024 768) resolution, F = 2.6 2.9, F = 18.4 22mm) and an Apple 15 LCD panel with backlight removed (768 x 1024). The luminance channel of the projector is further modulated by the LCD panel, resulting in a bright display with a very low black level. The system includes a 150mm achromatic focusing lens, a Fresnel lens (custom Reflexite 24 inch), and a diffuser (Reflexite BP331), to focus and collimate the projector beam on the plane of the LCD panel. Fig. 1 MCSL HDR Display
System Calibration Following the DICOM standard calibration techniques 2, the system characteristic curve was determined. Since the system luminance is multiplicative in nature, two characteristic curves, one for the projector and one for the LCD panel, were independently measured using an LMT Colormeter C1210. First the projector was set to white and the driving of the LCD varied from 0 to 255, then the LCD was set to white while the driving level of the projector was varied from 0 to 255. Multiplying the two curves provides a two-dimensional characteristic surface. Given a desired luminance, this surface can be used to identify the appropriate digital driving levels for both the LCD panel and the projector in order to match the grayscale standard display function defined in the DICOM standard. DICOM calibration of this display provided a unique challenge since it is driven by two separate 8-bit graphic units. Different methods of selecting the best driving levels for the LCD and projector in order for the system to have a characteristic curve matching the DICOM standard were investigated. It was determined that the best option was one that guaranteed that the projector was monotonically increasing while allowing the LCD to wander a certain calibration width in either direction of the projector s driving level. This allowed the LCD to provide fine-tuning of the displayed luminance in order to provide a high quality calibration. The display was found to have a low black level and a peak luminance of around 1,800 cd/m 2, providing a contrast ratio spanning around 900 JNDs. The HDR display was calibrated with multiple configurations: 1) max luminance of 1,200 cd/m 2 and 768 DDLs, 2) max luminance of 700 cd/m 2 and 768 DDLs, 3) max luminance of 1,200 cd/m 2 and 512 DDLs, 4) max luminance of 700 cd/m 2 and 512 DDLs, 5) max luminance of 450 cd/m 2 and 512 DDLs, 6) max luminance of 1,200 cd/m 2 and 256 DDLs, 7) max luminance of 700 cd/m 2 and 256 DDLs, 8) max luminance of 450 cd/m 2 and 256 DDLs. Figure 2 shows the characteristic surface of the HDR display and the path taken through it when using 768 quantization level with different maximum luminance. As specificed the projector is guaranteed to be monotonically increasing with the LCD providing fine adjustments of the luminance.
Fig. 2 Calibration Surface and Sample Calibration. To verify the calibration output, input patches spaced equally apart JNDs were measured using a Konica Minolta Chroma Meter CS-100A. Each calibration was evaluated using multiple metrics including the theoretically achievable JNDs, realizable JNDs, delta JND and the standard deviation of delta JND. As can be seen from the calibration results, shown in Table 1, a high quality calibration was achieved. Plotting the systems calibrated luminiance along with the DICOM standard display function provides a way to visually show the quality of the calibration (Figure 3). Grey Levels Max Luminance (cd/m2) Theoretically Achievable JNDs Realizable JNDs ΔJND std(δjnd) 256 450 685 256 2.69 0.62 256 700 751 256 2.94 0.61 256 1200 827 256 3.24 0.62 512 450 685 488 1.34 0.60 512 700 751 498 1.47 0.64 512 1200 827 512 1.62 0.57 768 700 751 660 0.98 0.66 768 1200 827 689 1.08 0.64 Table 1 Metrics showing the quality of the calibration of the HDR system with different display settings.
Fig. 3 DICOM standard display function shown with the HDR display function when calibrated with 768 quantization levels and a maximum luminance of 1200 cd/m 2. Experimental Setup Images of a male RANDO Phantom (The Phantom Lab, Salem, NY) acquired using a GE Lightspeed RT 4-channel scanner were provided by SUNY Upstate. Images consist of sixty-four axial slices in the chest obtained at 180 mas and 120kVp with automatic ma adjustment disabled. Reconstruction was done at 1.25mm using the standard reconstruction filter. Lesions were simulated by mathematically projecting a sphere with a diameter of 100 pixels of arbitrary size and a maximum signal value of 100 (Hounsfield units). This primary lesion was then blurred by a Gaussian with a standard deviation of 1 pixel to remove the sharp edge. The lesion can then be scaled to produce lesions of appropriate size for the pixel dimensions of the image the lesion is inserted into. Lesion sizes of 2.5, 3.75, 5.5, 8.5, and 12.5 mm were created in this manner. Fig. 4 Chest (left) and abdomen (right) images used for the 2AFC study. Different size lesions were inserted surrounded by a cue circle. The study followed a previously established methodology 3. During each 2-AFC experiment, a single image was presented for each trial, in which a lesion was randomly placed within one of two non-overlapping pre-determined regions. A low contrast cue circle was placed around the lesion, and a second dummy cue circle placed randomly in the other region (see Figure 4). The observer s task was to select the cue circle
containing the lesion. A staircase psychophysical procedure was used 4. Initially a high contrast level was set, then as the observer answers correctly the contrast level is decreased until a mistake is made or a predetermined contrast level reached. If a mistake is made, the contrast level is increased, and then three consecutive correct responses will be required before the scale factor is further reduced. Each experimental session consisted of 64 trials, and was repeated by each observer four times for each lesion size, with both the chest and abdomen images. Contrast detail curves representing the 92% correct detection rate were generated for the three observers. One example curve is shown in Figure 5. As expected the curves follow a power-law where increasing contrasted is needed for smaller lesions 5. Fig. 5 Example contrast detail curve. The area under the contrast detail curve was used to compare the different display settings. The plots in Figure 6 compare the calculated areas.
Fig. 6 Plots showing the area under the curve for different display settings. A smaller area indicated improved observer performance.
Conclusions and continuation of work A calibration procedure for a dual driven display was designed and successfully applied to calibrate the HDR display with eight different display settings. The observer performance for each of the display settings was then evaluated using an AFC study. The results are very informative. It was found that increasing the display luminance without increasing the luminance can be detrimental to observer performance. On the other hand, when dealing with a larger number of quantization levels the increased dynamic range proved beneficial. For a display with a fixed luminance there does not appear to be any benefit to increasing the number of quantization levels beyond the 256 supported by most displays. These results are preliminary and not conclusive as the study is ongoing and additional data is still to be collected and evaluated. The results presented in figure 6 are a subset of that which was collected from the three study participants and while they show interesting trends the differences in display settings are not statistically different at this phase of the project. Additional trials are being performed to confirm these results and determine the significance. Student involvement: 1. Natalie Tacconi worked on this project in the spring quarter 20083. She was responsible for installing the software provided by Dr. Kent Ogden and did the initial selection of images. 2. Robert Harrigan worked on this project during the summer quarter (20084) and the beginning of fall quarter (20091). He was responsible for the calibration of the system, preparation of images for the experiments, submission of IRB consent form, and initial analysis of results. 3. Allen Greer started working on this project midway through the fall quarter (20091). He is responsible for the analysis of data. Conference presentations: 1. Baum, K.G., Harrigan, R.L., Tacconi, N., Phillips, J., Heckaman, R., Ferwerda, J.A., Ogden, K.M., Helguera, M., Calibration and Evaluation of a Dual Layer High Dynamic Range Display for Pathology Detection Tasks, Proceedings of the Imaging Technologies in Biomedical Sciences Symposium, 2009. 2. Harrigan, R., Tacconi, N., Baum, K.G., Helguera, M. DICOM Calibration of a Dual Layer High Dynamic Range Display, RIT Undergraduate Research Symposium, 2009. Journal paper in preparation: Baum, K.G., Harrigan, R.L., Tacconi, N., Ogden, K.M., Helguera, M., Calibration and Evaluation of a Dual Layer High Dynamic Range Display for Pathology Detection Tasks, Journal of Display Technology, 2009. Error! Bookmark not defined.
1 Guarnieri, G., Image Splitting Techniques for a Dual Layer High Dynamic Range LCD Display, IEEE Medical Imaging, 2008. 2 Digital Imaging and Communications in Medicine (DICOM), Part 14: Grayscale Standard Display Function. National Electrical Manufacturers Association, 2004. 3 Ogden, K.M., Huda, W., Shah, K., Scalzetti, E.m., Lavallee, R.L., Roskopf, M.L, A Method for Analyzing Contrast-Detail Curves, Medical Imaging 2007, Proc. of SPIE, 6515, 6515H-1-H-9, 2007 4 Gescheider, G.A., Psychophysics: The Fundamentals, 3rd Edition. Erlbaum, 1997. 5 A. E. Burgess, F. L. Jacobson and P. F. Judy, "Human observer detection experiments with mammograms and power-law noise," Med.Phys. 28 (4), 419-437 (2001)