Neonatal Chest Computed Radiography: Image Processing and Optimal Image Display

Transcription

1 Neonatal Computed Radiography Pediatric Imaging Clinical Observations Steven Don 1,2 Bruce R. Whiting 2 Jacquelyn S. Ellinwood 3 David H. Foos 3 Keith A. Kronemer 1 Richard A. Kraus 1 Don S, Whiting BR, Ellinwood JS, Foos DH, Kronemer KA, Kraus RA Keywords: chest, digital images, neonatal imaging, PACS, pediatric radiology DOI:1.2214/AJR Received April 29, 25; accepted after revision November 1, 25. Presented at the 24 annual meeting of the American Roentgen Ray Society, Miami Beach, FL. Supported in part by National Institutes of Health research grant 1 R41 HD The employment status of J. S. Ellinwood and D. H. Foos at Eastman Kodak Company did not influence the data in this study. 1 Mallinckrodt Institute for Radiology, St. Louis Children s Hospital, Washington University School of Medicine, 51 S Kingshighway, St. Louis, MO, Address correspondence to S. Don (dons@mir.wustl.edu). 2 Electronic Research Laboratory, Mallinckrodt Institute for Radiology, Washington University School of Medicine, St. Louis, MO. 3 Eastman Kodak Company, Rochester, NY. AJR 27; 188: X/7/ American Roentgen Ray Society Neonatal Chest Computed Radiography: Image Processing and Optimal Image Display OBJECTIVE. The purpose of this study was to determine soft-copy image display preferences of brightness, latitude, and detail contrast for neonatal chest computed radiography to establish a baseline for future work on low-dose imaging. CONCLUSION. Observers preferred brighter images with higher detail contrast and narrow to middle latitude for soft-copy display compared with the typical screen-film hard-copy appearance. Future research on low-dose neonatal chest imaging will be facilitated by an understanding of optimal soft-copy image display. raditional screen-film radiographs T have a characteristic density/log (exposure) response (Hurter and Driffield curve or tone scale) that is sigmoidal in shape and determines the density and contrast of screen-film images at a specific peak kilovoltage and effective tube current [1]. An underexposed image falls in the toe of the characteristic curve, and an overexposed image falls in the shoulder of the characteristic curve. The signal response in computed radiography (CR) is linear with exposure, resulting in a wider dynamic range: 1- to -fold greater than that of screen-film radiography. The result is better tolerance of underexposure and overexposure with CR, but the unprocessed raw image data do not have the proper image contrast and density necessary for image interpretation [2, 3]. Image processing of raw CR digital data frequently results in the screen-film radiographic appearance that radiologists are accustomed to interpreting. This appearance typically is achieved because of laser printing of the image on film. Hospitals, however, have begun to convert to soft-copy image display. Image processing, because of the unprocessed raw image data, must include recognition of the amount of exposure the X-ray plate receives in the relevant anatomic regions. Grayscale rendition is then applied to achieve appropriate overall brightness in the displayed image. Gray-scale rendition is performed to simulate the Hurter and Driffield curve and to produce an image acceptable to radiologists. Specialized image processing, such as edge enhancement and signal equalization, also may be used to emphasize features in the CR images [2, 3]. Because of concerns about patient radiation dose, it is desirable to perform CR examinations with a dose level as low as reasonably achievable while maintaining an acceptable level of image noise [4]. These requirements can be met by changing radiographic parameters, such as peak kilovoltage and effective tube current, or by using image processing to suppress noise while preserving image features. Many neonatal dose-reduction studies have been performed with laser-printed film and fixed image processing [5 7]. In those studies, investigators tested the effects of dose reduction on the diagnosis of hyaline membrane disease and pneumothorax [5 7]. Soft-copy image display parameters must be controlled to prevent interacting effects from confounding dose-reduction experiments. Soft-copy image display preferences for adult chest CR were determined in a previous study [8], but requirements for soft-copy image display preferences for neonatal chest CR are unknown. Because of concerns about radiation sensitivity in children, research is being conducted into methods of reducing radiation dose in CR. The purposes of this research are to determine user preferences for neonatal chest CR soft-copy image display as a baseline for future soft-copy work on low-dose imaging and to gain an understanding of which factors are important in the optimal presentation of neonatal images. The effects of low-dose exposure on observer performance in the detection of disease can then be studied AJR:188, April 27

2 Neonatal Computed Radiography Materials and Methods Image Acquisition This study received the approval of our institutional review board. All image data were handled according to the Health Insurance Portability and Accountability Act (HIPAA). Images were obtained with a Kodak 4 CR system (Eastman Kodak Company). Five chest radiographs of healthy neonates and five chest radiographs of neonates with clearly discernible pneumothorax were selected. Raw image data were collected at a quality-control workstation with all image processing turned off. These data were transferred to another workstation for removal of all patient identification data and then were sent to an analysis workstation. A C Fig. 1 Normal neonatal chest radiographs illustrate response curve to brightness adjustments for neonatal chest computed radiography. Lungs appear progressively lighter from lowest brightness to highest brightness. A, Low-brightness image. B, Reference T-MAT G (Eastman Kodak Company) image. C, High-brightness image. D, Graph shows adjustment of characteristic curve from lowest to highest brightness. At fixed input code value, lowest brightness images have higher or more lung density output and highest brightness images have lower or more bone density output. Lung Output Code Value Bone 4, 4, 3, 2, 2, 1, 1,2 Imaging Processing The raw image data were processed with a prototype image-processing algorithm [8]. Three variables were altered: brightness, detail contrast, and latitude. A control image for each patient was acquired to simulate the characteristic response curve of T-MAT G film (Eastman Kodak Company). This screen-film combination is common in pediatric chest radiography. Brightness is increased on a CR image through a shift in the response curve to the right (Fig. 1). This control is analogous to the level in CT image display in which the lungs are dark at a mediastinal level and bright at a lung level. Detail contrast is increased on a CR image through an increase in the slope of the linear portion 1,4 1,6 1,8 2, 2,2 Input Code Value 2,4 Lowest brightness Middle brightness Highest brightness 2,6 2,8 of the response curve. Low-detail contrast images appear grayer, and high-detail contrast images are more black and white (Fig. 2). This control is analogous to the window in CT in which shallow lung window contrast minimizes the contrast difference between water and soft tissue and steep mediastinum window contrast accentuates the difference between water and soft tissue. adjustment on a CR image changes the overall contrast of an image without affecting local contrast or resolution. For a wide-latitude image, one compresses the large-area, low-frequency data of the histogram information and subtracts that from the histogram of the image. The overall appearance of the image is grayer yet retains local contrast. For ex- B D AJR:188, April

3 Lung Output Code Value Bone 4, 4, 3, 2, 2, 1, 1, 1,7 1,9 Highest contrast = 4. 2,1 2,3 2, Input Code Value T-MAT G contrast = 3.1 Lowest contrast = 2.5 2,7 2,9 A C A Fig. 2 Normal neonatal chest radiographs show detail contrast adjustment of response curve for neonatal chest computed radiograph. See reference image in Figure 1B for comparison. Difference in contrast between air in stomach and ribs is evident. A, Low-detail contrast image appears gray. B, High-detail contrast image is more black and white than A. C, Graph shows adjustment of response curve. As detail contrast increases, characteristic curve becomes steeper (T-MAT G, Eastman Kodak Company). Fig. 3 Normal neonatal chest radiograph with latitude adjustment of response curve for neonatal chest computed radiograph. See reference image in Figure 1B for comparison. Difference in free-in-air exposure and soft tissue around humerus is evident. A, Wide-latitude image appears gray. B, Graph shows adjustment of characteristic curve. As latitude increases, steepness of response curve decreases. This finding applies to low-spatial-frequency data only. Lung Output Code Value Bone 4, 4, 3, 2, 2, 1, = 1.46 =.56 1, 2, 2, Input Code Value 3, 4, B B ample, the contrast between a rib and adjacent lung is maintained, but the contrast between the free-inair exposure and soft tissues around the humerus is decreased (Fig. 3). This manipulation highlights features such as lung markings and bone detail and minimizes differences such as density difference between lungs. Examples of the effect of altering the image-processing parameters of brightness, detail contrast, and latitude for a neonate with pneumothorax are presented in Figure 4. Each case was processed with five brightness levels, four detail contrast levels, and seven latitude 114 AJR:188, April 27

4 Neonatal Computed Radiography levels. One hundred ten combinations of 14 possible combinations of each patient s image were made. The processing parameter matrix was not symmetric, and the narrower-latitude and lowerdetail contrast images were not acquired because observation [8] had shown the range of lowest latitude and contrast performed the worst. With this criterion, 3 image combinations per patient were eliminated without alteration of the results, allowing reduction of computational time and enhancement of observer participation. Combinations were achieved by adjustment of the tone scale with a fixed shape that approximately matched the tone scale curve of T-MAT G film. The A B C D E F Fig. 4 Neonatal chest radiographs with pneumothorax. Images show effects of brightness, detail contrast, and latitude adjustment on neonatal chest radiographs. A, Reference image. B, Low-brightness image. C, High-brightness image. D, Low-contrast image. E, High-contrast image. F, Wide-latitude image. middle brightness level was chosen as the level that would yield an optical density of approximately 1.65 on a T-MAT G film image of the lung region. Brightness adjustments were achieved by shifting the tone scale along the input axis in increments of 75 and 15 code values above and below the middle brightness code value for each image to yield lowest, lower, middle, high, and highest brightness settings. At the reference detail contrast setting, that adjustment would yield optical densities in the lung region of, from low to high, 2.25, 1.95, 1.65, 1.35, and Detail contrast adjustments were achieved by pivoting the tone scale at set increments around the input value corresponding to a density of 1.. adjustments were achieved by reducing the low-frequency contrast at set increments while maintaining the detail contrast, achieved through a signal-equalization filtering step. The control image-processing parameters were chosen as the image at middle brightness (one that would yield an optical density of approximately 1.65 on T-MAT G film in the lung region), low-detail contrast (T-MAT G film contrast of 3.1), and narrow latitude (one that would maintain the tonal characteristics of low-frequency data for a detail contrast level of 3.1). These parameters would simulate on soft-copy display the hard-copy screenfilm appearance of neonatal chest radiographs. AJR:188, April

5 Image Viewing One hundred ten images were acquired for each of 1 sets of neonatal CR images for a total of 1,1 images. These images were randomized and presented in 1 sessions of 11 images. The images were viewed with a stand-alone PC workstation on a high-resolution (2, 2,) cathode ray tube monitor (model DR 11, Data Ray). The monitor was calibrated, and lighting was controlled to eliminate glare on the screen. The workstation was in a quiet room for elimination of extraneous noise. Each test image was presented along with the control processed image. The observer toggled between the test image and the control image and controlled the toggle rate and amount of observation time before rating. Three pediatric radiologists with certificates of added qualification in pediatric radiology participated in the image review. These radiologists were familiar with soft-copy interpretation. Each radiologist rated each image on a nine-point viewing scale relative to the control image as in previously published adult work [8]: 4, image quality markedly better, diagnosis likely altered; 3, image quality clearly better, diagnosis might be altered; 2, image quality somewhat better, diagnosis should be the same; 1, image quality slightly better, diagnosis will be the same;, no difference; 1, image quality slightly worse, diagnosis will be the same; 2, image quality somewhat worse, diagnosis should be the same; 3, image quality clearly worse, diagnosis might be altered; 4, image quality markedly worse, diagnosis likely altered. Statistical Analysis A regression model with a second-order polynomial was used to analyze the data. The factors of brightness, latitude, and detail contrast were treated as continuous. A logarithmic transform of the latitude was used to improve conformance to the polynomial fit. Brightness, detail contrast, and log-latitude levels were centered to reduce the correlation between predictor values, thus clarifying significance testing. In addition, health condition (presence or absence of pneumothorax) was included as a discrete and fixed effect. Observer and patient effects were treated as discrete and random, and the patient effect was considered nested within the health condition. The peak soft-copy viewing preference range of detail contrast and latitude cell at each brightness level was determined. The best-rated cell at each brightness level was identified. The 95% CI was determined for the difference between the predicted mean of the cell with the best rating and the predicted mean of the cell being checked. All cells identified as not significantly different from the best rating cell were included in the peak preference range for each brightness level. Results The regression results showed that brightness (p <.1) was the single most important factor in determining user imaging and diagnostic preferences (Table 1). Also significant in order of contribution to the variance were observer, detail contrast, latitude, and patient (Table 1). The presence or absence of pneumothorax was not statistically significant (p =.14). There was a peak cell-rating difference of three points between the user-preference predicted best cell rating at the lowest brightness ( 1.7) and the user-preference predicted best cell rating at the highest brightness (1.3). At the lowest brightness, all scores were significantly less than, indicating that the images at this brightness level, regardless of detail contrast and latitude, were less desirable than the reference image in image and diagnostic quality (Table 2). At the middle brightness level, which included the reference image, the predicted ratings were near (Table 3). The ratings at the highest brightness and highest detail contrast and narrow latitude had the highest predicted score (Table 4). At none of the image brightness settings was there a statistically significant difference between the ratings for healthy patients and those for patients with pneumothorax. Discussion Despite the presence of CR in pediatric imaging for nearly 2 years [9], to our knowledge no scientific methodology has been developed for addressing specific soft-copy viewing preferences. Years ago, a hard-copy CR film included two images, one imitating a screen-film radiograph and a second that enhanced edge structures, such as catheters [8, 9]. Currently, one hard-copy CR film melds the screen-film appearance and edge enhancement. Our findings clearly indicate that soft-copy viewing preferences in neonatal chest CR are different from the established hard-copy screen-film radiographic preferences [8]. Our findings establish baseline soft-copy settings for future use in image display and dose-reduction research. More radiology departments are shifting to soft-copy display and eliminating film entirely. Much work has been done to meet the challenge of PACS and soft-copy display [1]. Research has focused on comparing soft-copy display with hard-copy film [11], resolution of the monitor [11], CR with other digital radiographic systems [12, 13], and monitor luminance [14 16]. In these studies, the images were initially displayed from fixed, preset image-processing algorithms. In our study, the image-processing parameters were the tested variables. In three of the previous studies [12 14], it was explicitly stated that radiologists were allowed to adjust window and level but not spatial filtering. In one study [14], the radiologists manipulated window and level settings 9% of the time. In only one study [8], with adult subjects, were preferred soft-copy display settings evaluated. Our study of neonatal chest CR images showed that pediatric radiologists find improved image and diagnostic quality of softcopy display with the brightest image setting and with high-detail contrast and narrow to middle latitude processing settings. Adjusting the CR image display resulted in a higher rating by the radiologists compared with the reference screen-film radiographic appearance. The CR viewing preference was different from the screen-film radiographic appearance of the characteristic response curve of T-MAT G film, which is middle brightness, narrow latitude, and less detail contrast. Wide-latitude images were not preferred for routine viewing. The parameters that were varied in this experiment (brightness, detail contrast, and latitude) are the building blocks of image quality. As such, the results can be generalized for use with most CR systems, provided the interface with image processing has the required flexibility. Although the particular image processing (rendering) parameters are different among CR systems, the functionality of image processing can be related to these three fundamentals of image quality. The results of this study can be applied to soft-copy and hard-copy interpretation, provided the image-processing parameters are adapted to compensate image appearance for the dynamic range, gray-scale resolution, and spatial resolution differences among printers and soft displays. The concept of the building blocks of image quality has been detailed [17]. In a previous study [8] of posteroanterior chest radiographs of adults, the radiologists preferred a wide-latitude (highly equalized) image with less detail contrast, different from the user preferences we found. One explanation may be that adult chest images have a broader histogram range than chest images of neonates. This difference is attributable to the size difference between neonates and adults. The anatomic structures of neonates are less attenuating than are those of adults. The typical peak kilovoltage used for neonatal chest radiographs is 6 7 kvp, whereas that for 1142 AJR:188, April 27

6 Neonatal Computed Radiography TABLE 1: Regression Model Results Source of Variation Degrees of Freedom F p Brightness 1 2, <.1 a Observer <.1 a Detail contrast <.1 a <.1 a Patient <.1 a Normal finding versus pneumothorax a Statistically significant difference. TABLE 2: Predicted Ratings for Lowest Brightness Contrast Narrowest (.44) Narrower (.5) Narrow (.56) Middle (.7) Wide (.92) Wider (1.9) Widest (1.46) Highest (4) High (3.5) Low (3.1) Lowest (2.5) Note Values in parentheses are actual latitude and contrast settings used. Peak preference range is displayed in boldface, and the values are not significantly different from the highest rated value. = Δlog (exposure), detail contrast = Δdensity / Δlog (exposure). TABLE 3: Ratings for Middle Brightness Contrast Narrowest (.44) Narrower (.5) Narrow (.56) Middle (.7) Wide (.92) Wider (1.9) Widest (1.46) Highest (4) High (3.5) Low (3.1). a Lowest (2.5) Note Values in parentheses are actual latitude and contrast settings used. Peak preference range is displayed in boldface, and the values are not significantly different from the highest rated value. = Δlog (exposure), detail contrast = Δdensity / Δlog (exposure). a Reference T-MAT G (Eastman Kodak) image processing. TABLE 4: Ratings for Highest Brightness Contrast Narrowest (.44) Narrower (.5) Narrow (.56) Middle (.7) Wide (.92) Wider (1.9) Widest (1.46) Highest (4) High (3.5) Low (3.1) Lowest (2.5) Note Values in parentheses are actual latitude and contrast settings used. Peak preference range is displayed in boldface, and the values are not significantly different from the highest rated value. = Δlog (exposure), detail contrast = Δdensity / Δlog (exposure). adult chest radiographs is much higher, in the range of kvp. The higher peak kilovoltage used for adult chest radiographs results in lower contrast and a wider range of optical density than on neonatal chest radiographs. A wider-latitude image compresses the range of optical density for optimal display of soft-copy chest images of adults. The presence or absence of pneumothorax did not significantly affect viewer preference. This finding implies that there is no need for routine alteration of image display for routine evaluation for pneumothorax. This study, however, did not test ability to diagnose subtle cases of pneumothorax. A previous study of neonatal chest images showed that edge enhancement, which was not investigated in this study, improved detection of subtle pneumothorax [18]. There were limitations to this study. Only three radiologists from a single site were used to select user preferences. The statistically significant difference between radiologist preferences (p <.1) (Table 1) indicates that image-display preference is a highly personal choice. In addition, only one vendor s CR equipment was used. The results cannot be extrapolated to other types of monitors, uncontrolled viewing situations, and other vendors CR and digital radiography systems. Direct extrapolation of the findings of this study to current clinical practice as a department transitions to soft-copy interpretation is problematic given the limitations of this study. Setting up and performing experiments such as ours require access to raw image data, imaging physicists, and time. To aid clinical departments in the transition, radiologists should encourage the vendors of digital imaging equipment and PACS workstations to develop a library of cases that include various ages, body parts, and disease processes. When new equipment is installed, radiologists can review the cases to find their personal preferences. The results of this study establish baseline soft-copy display parameters for neonatal chest CR. These parameters are clearly different from those for the screen-film hard-copy laser-printed images to which radiologists have become accustomed for interpreting CR images. Scientific research on image processing and dose reduction requires establishing baseline image processing. Future improvement in image processing and studies on the effects of dose reduction on detection of disease processes can be conducted with controlled imaging parameters for reducing and eliminating confounding factors in statistical analysis. In conclusion, this study showed that pediatric radiologists prefer high brightness, high detail contrast, and a narrow to middle latitude image processing for viewing neonatal chest CR images compared with a soft-copy control image on which tonal characteristics and brightness are mapped to yield an optical density of approximately 1.65 in the lung region. The presence or absence of pneumothorax did not affect viewing preferences. AJR:188, April

7 References 1. Curry TS, Dowdey JE, Murry RC. Photographic characteristics of x-ray film. In: Christensen s physics of diagnostic radiology. Malvern, PA: Lea & Febiger, 199: Hillman BJ, Fajardo LL. Clinical assessment of phosphor-plate computed radiography: equipment, strategy, and methods. J Digit Imaging 1989; 2: Cowen AR, Workman A, Price JS. Physical aspects of photostimulable phosphor computed radiography. Br J Radiol 1993; 66: Don S. Radiosensitivity of children: potential for overexposure in CR and DR and magnitude of doses in ordinary radiographic examinations. Pediatr Radiol 24; 34[suppl 3]:S167 S Don S, Cohen MD, Kruger RA, et al. Volume detection threshold: quantitative comparison of computed radiography and screen-film radiography in detection of pneumothoraces in an animal model that simulates the neonate. Radiology 1995; 194: Don S, Hildebolt C, Sharp T, et al. Comparison of computed radiography with film-screen radiography in detecting pulmonary edema in a rabbit model simulating the neonate. Radiology 1999; 213: Roehrig H, Krupinski EA, Hulett R. Reduction of patient exposure in pediatric radiology. Acad Radiol 1997; 4: Flynn M, Couwenhoven M, Eyler W, et al. Optimal display processing for digital radiography. In: SPIE medical imaging 21: visualization, display and image guide procedures. Piscataway, NJ: IEEE, 21: Dietrich RB, Boechat MI, Huang HK. Experience with phosphor imaging plates: clinical experience in pediatric radiology. Proc SPIE 1989; 191: Arenson RL, Chakroborty DV, Seshadri SB, Kundel HL. The digital imaging workstation. Radiology 199; 176: Otto D, Bernhardt TM, Rapp-Bernhardt U, Ludwig K, Liehr UB, Dohring W. Subtle pulmonary abnormalities: detection on monitors with varying spatial resolutions and maximum luminance levels compared with detection on storage phosphor radiographic hard copies. Radiology 1998; 27: Goo JM, Im J, Kim JH, et al. Digital chest radiography with a selenium-based flat-panel detector versus a storage phosphor system. AJR 2; 175: Goo JM, Im J, Lee HJ, et al. Detection of simulated chest lesions by using soft-copy reading: comparison of an amorphous silicon flat-panel detector system and a storage-phosphor system. Radiology 22; 224: Goo JM, Choi J, Im J, et al. Effect of monitor luminance and ambient light on observer performance in soft-copy reading of digital chest radiographs. Radiology 24; 232: Ikeda M, Ishigaki T, Shimamoto K, et al. Influence of monitor luminance change on observer performance for detection of abnormalities depicted on chest radiographs. Invest Radiol 23; 38: Ishihara S, Shimamoto K, Ikeda M, et al. CRT diagnosis of pulmonary disease: influence of monitor brightness and room illuminance on observer performance. Comput Med Imaging Graph 22; 26: Couwenhoven M, Senn R, Foos D. Enhancement method that provides direct and independent control of fundamental attribute of image quality for radiographic imagery. In: SPIE Medical Imaging 23: visualization, display and image guide procedures. Piscataway, NJ: IEEE, 24: Goo HW, Kim H, Song K, et al. Using edge enhancement to identify subtle findings on softcopy neonatal chest radiographs. AJR 21; 177: AJR:188, April 27