Improved visualization of simulated nodules by adaptive enhancement of digital chest radiography

Improved visualization of simulated nodules by adaptive enhancement of digital chest radiography

Improved Visualization of Simulated Nodules by Adaptive Enhancement of Digital Chest Radiography Jong Hyo Kim, PhD, 1,2,3 Jung-Gi Im, MD, 1 Man Chung ...

3MB Sizes 0 Downloads 19 Views

Improved Visualization of Simulated Nodules by Adaptive Enhancement of Digital Chest Radiography Jong Hyo Kim, PhD, 1,2,3 Jung-Gi Im, MD, 1 Man Chung Han, MD, 1 Byoung Goo Min, PhD, 2 and Choong Woong Lee, PhD 3

Rationale and Objectives. Although digital radiography of the chest has a wide dynamic range, difficulties still remain in visualizing the acquired images faithfully on gray-scale displays (cathode-ray tube [CRT] displays), which have a much lower level of luminance than film view boxes. We propose an adaptive-enhancement algorithm for digital chest radiography that provides faithful visualization of the chest on the CRT. Methods. We investigated the contrast sensitivity of a CRT monitor and developed an image processing algorithm that compresses the dynamic range and enhances image contrast selectively in the mediastinal area and that transforms the gray scale to visualize the image by use of the full effective dynamic range of the CRT. We performed a receiver-operating characteristic (ROC) study by using simulated nodules to evaluate the clinical value of the proposed algorithm. Results. The processed images provided improved visualization of both mediastinal and lung regions. The area under the ROC curve for retrocardiac or subdiaphragmatic nodule detection increased significantly (from 0.69 to 0.79; P < 0.05). The area under the ROC curve for lung nodule detection also increased (from 0.64 to 0.75; P < 0.1), although not to the level of statistical significance.

Conclusion. The proposed algorithm allows improved visualization of nodules on digital chest radiographs with the CRT display. From the 1Department of Radiology and the 2Instituteof BiomedicalEngineering,Seoul National University Hospital,28 Yongon-dong,Chomgmo-gu, Seou1110-744,and the 3Departmentof Electronics Engineering, Seoul National University, 56 Shilimdong, Kwanak-ku,Seou1151-742,South Korea. Address reprint requests to Jong Hyo Kim, PhD, Departmentof Radiology,Seoul NationalUniversity Hospital, 28 Yongon-dong, Chomgmo-gu, Seoul 110-744,SouthKorea. Received October14, 1993, and acceptedfor publicationafter revision May 18, 1994. Acad Radiol 1994;1:93-99

9 1994, Association of University Radiologists

Key Words. Image processing; chest radiography; digital radiography; receiver-operating characteristic curve.

he introduction of digital technology to chest radiography provides the potential to overcome the inherent limitations of conventional chest radiography, such as the limited dynamic range of radiographic film and scattered radiation [1]. Several image-processing algorithms for digital chest radiography have been suggested in an attempt to provide faithful visualization. These include unsharp masking, adaptive unsharp masking, and regionally adaptive histogram equalization [2-4]. In these algorithms, how-

T

93

KIM ET AL.

ever, a film view box display was assumed and little consideration was given to soft-copy display on a cathode-ray tube (CRT). Soft-copy display on a CRT is being used in clinical practice because of the increasing interest in picture archiving and communicating systems (PACSs). However, the dynamic range of CRT displays is significantly inferior to that of conventional film view b o x displays, and CRT displays show granular noise because of the phosphor, which results in reduced detection of lowcontrast objects [5, 6]. Faint lung nodules m a y be overlooked m o r e frequently with CRT displays than with film view b o x displays. This problem can be serious with PACSs and necessitates the application of image processing to image displays to achieve faithful visualization of radiographic images on the CRT. We propose an adaptive-enhancement algorithm for digital chest radiographs that provides faithful visualization of the chest on a CRT. To derive such an algorithm, we investigated the contrast sensitivity of the CRT monitor with a just-noticeable-difference (JND) measure and incorporated the results into an image-processing algorithm. This algorithm compresses the dynamic range and enhances contrast selectively in the mediastinal region and transforms the gray scale to project the image data onto the full dynamic range of the CRT. In addition, we evaluated the clinical value of the prop o s e d method in an observer performance study.

MATERIALS AND METHODS Contrast Sensitivity of the CRT To evaluate the performance of the CRT in visualizing diagnostic information, we investigated the contrast sensitivity of the CRT with the JND measure. Two different gray-scale monitors were ~fsed. Both display 1024 lines at a 60-Hz frame rate. The type I monitor (Image Systems, H-opkins, MN) has a m a x i m u m luminance of 60 foot-Lamberts (fL), and the type II monitor (Sampo, Norcross, GA) has a m a x i m u m luminance of 45 fL. They were calibrated with the Society of Motion Picture and Television Engineers test pattern [7]. We performed a 20-alternative-forced-choice experiment in which an observer was presented with an image that contained a signal and was forced to choose the position (of the 20) that was likely to contain the signal [8]. A 1-cm 2 object was used as the visual target. Unlike previous studies in which large visual targets

94

Vol. 1, No. 2, October 1994

were used, in this study the small visual target simulated the nodules that are frequently encountered in routine radiology practice. The observers were two third-year residents of radiology. The ambient lights were dimmed to prevent glare on the CRT. We prepared three sets of 20 images to measure the JND at a particular background gray level, with each set having a different signal difference. The three signal differences were selected to be close to the barely visible signal difference at that background level in a preliminary test. Each observer read the images at a viewing distance of 50 cm and made 20 responses to a given set. The n u m b e r of correct responses divided by 20 yielded the percentage of correct detection for a given signal difference for an observer. We averaged the percentages of correct detection for the two observers. We defined the JND as the signal difference that produces 50% correct detection. This value corresponds to a point on the receiver-operating characteristic (ROC) curve that represents 5% false-positive and 50% true-positive. From a m o n g the three measurements of percentage of correct detection, w e selected the two that were closest to 50% and determined the JND from the interpolation of the two signal differences. By repeating this procedure, we measured the JND at 10 different background gray levels.

Image Acquisition Chest images of 50 patients without pulmonary nodules were obtained with a custom digital radiography system (DR-1000; Institute of Biomedical Engineering, Seoul National University Hospital, Seoul, South Korea) [9]. This system uses a phosphor screen-photodiode array as a radiographic detector and provides a 1024 • 1024 matrix image of 12 bits in a 1-sec scan time. Images were transferred to a personal computer (PC/486DX 33 MHz) for processing. During the processing, image data were manipulated in a 12-bit gray scale and then transformed to an 8-bit scale for display.

Adaptive Image Enhancement The given image is divided into the low-frequency and high-frequency components, which are processed separately. The low-frequency image (F L) is obtained by smoothing the original image (F) with a square filter mask that measures 15 m m (50 pixels) per side (equation 1).

IMPROVED

Vol. 1, No. 2, October 1994

FL(I,j ) = ~

1

VISUALIZATION

OF SIMULATED

NODULES

N/2 -1

(1)

~, ~ F(i- k,j-I) k,I = -N/2 where N = 50

The high-frequency image (F H) is calculated as the difference between the original image and FL (equation 2).

F H = F- F L

(2)

Ft is modified by the nonlinear gradation function (H) (equation 3) to reduce the dynamic range. (3)

F H' = H [ F L] F

0_
where H = IFL/2+T/2 ~FL

between the lung and the mediastinum. This gray-level adaptive modification of high-frequency and low-frequency components achieves selective processing for the lung and the mediastinum; no processing is done for the lung while dynamic range reduction and local contrast enhancement are performed for the mediastinum. The gray-level threshold T was selected as the gray value that corresponds to the 50% level in the cumulative pixel histogram9 This threshold selection resulted from the observation that the lung and the mediastinum occupy about the same amount of space in a typical chest radiograph. Figure 2 shows this threshold selection scheme. The modified high-frequency and low-frequency components are combined to make the modified image F' (equation 5). F' = F L" + F H'

FH is amplified by the function c~ of FI. (equation 4) to enhance the local contrast.

F H' = (X(FL) • ll

H

(4)

Finally, the minimum and maximum pixel values are found by use of the histogram of the modified image F', and the gray scale between them is linearly transformed to the full effective dynamic range of the CRT (equation 6). F' = G [ F ' ]

0 <-FL< T - 512 (mediastinal range)

where ~= _sin{l_.0~(FL_T2)} +2,0 T - 5 1 2 ~ F L < T + 5 1 2 (interface) T+512
L

(5)

240

lung range

where G = ,

(6) 0-
225

(F,_T2) § 15

Tm~
| r a i n - ~rnax

"i5

Figure 1 shows an example H and the gain ~t. Both H and r change at the point T of the gray-level threshold

Tmax _
rain [ F'(i,j)]

Trnax = max [ F'(i,j)]

4

4096

100

Jl

_

3

N

~

3072 ~

80

g

60

,-I

2

~

2046

'~i ~

e~

r~

~

-/

,

stogram

~9 ,~

~Cumu~tiv~

1

/

1024

._>0 4 0

T o

I

0

Y

I

1024 2046 Pixel Value

o

I

3072 of FL

4096

FIGURE 1. Nonlinear gradation function (H) and gain function ((x), both of which are functions of the low-frequency image (FL). H has slopes of unity for the lungs and 0.5 for the mediastinum, thereby compressing the dynamic range of the mediastinum selectively. The ~ has maxima of 3.0 for the mediastinum and unity for the lungs. Multiplication of the high-frequency image by a enhances the local contrast in the mediastinum selectively, A smooth transition around the threshold (T) was necessary to minimize the ringing artifact near the lungmediastinum interface.

0

1024

2048 Pixel V a l u e

3072

4096

of F

FIGURE 2. Threshold selection scheme. The gray-level threshold (T) between the lung and the mediastinum was selected as the gray value that correspondsto the 50% level in the cumulative histogram. This value corresponds well to the histogram valIey between the Iungs (I_)and the mediastinum (M). F = original image.

95

KIM ET AL. 100

~

vol. 1, No. 2, October 1994

-

256

80

192

'g

_J .......... Histogram --Transformation

o

u_

-

128

",~

i

~ 40 m

== G

g:

_

~ 20

i /

o

Tmi~/ 0

""\,

~

64

,

o

I

1024

2048 Pixel Value

3072

4096

of F'

FIGURE 3. Gray-scale transformation G. Tmin and Tmax are the minimum and maximum pixel values of the modified image F', respectively. G linearly transforms the gray scale between Tmin and Tmax to the full dynamic range of the CRT.

Figure 3 shows an example of gray-scale transformation. On the basis of the results of the JND experiment, we defined the effective dynamic range of a CRT as the 226 gray levels of the 256, excluding the lowermost and u p p e r m o s t 15 gray levels, at which the contrast sensitivity is significantly low. After the gray-scale transformation, the processed image (Fp) has an 8-bit scale and is sent to the display frame buffer.

Observer Performance Test Simulated nodules were created and added to the images by use of a computer program as described elsewhere [10]. A nodule was selected from the library of prepared nodules and moved to a desired location, and the contrast and margin sharpness were adjusted to selected degrees. The size and contrast of simulated nodules rar{ged from 5 m m to 15 m m and from 5 m m to 12 mm, respectively. From a set of 50 chest images, we selected 13 for placement of simulated nodules in ehe lung fields and another 12 for placement of simulated nodules in the mediastinal and subdiaphragmatic areas. We placed a single simulated nodule in each selected image. The 50 images were processed with the proposed algorithm. Therefore, a total of 100 images were prepared as a reading set for the observer performance study. Images were displayed on the type I monitor in a r a n d o m sequence by use of the computer program. Ambient lights were dimmed, and the CRT monitor was calibrated before observation. To familiarize the observers with the CRT display and the processed images, we prepared a preliminary reading session with images that were not included in the test reading set. Two board-certified radiologists and 96

three senior radiology residents interpreted these images. Observation time was not limited, but it took about 40 min to finish reading the total set of images. Observers were told that each image contained at most one nodule and were told not to manipulate the brightness and contrast controls. They were asked to indicate their level of confidence with regard to the existence of a nodule by using a five-point rating scale, with 1 indicating definitely absent and 5 indicating definitely present. ROC analysis was performed with these data. The ROC curves and the areas under the ROC curve (Az) for each observer were obtained with the ROCFIT program (C. E. Metz, University of Chicago). ROC curves for the observers were generated from the pooled data for the lung and mediastinal nodules in the processed and unprocessed images. The statistical significance of differences between ROC curves was calculated by applying a paired Student's t test to the Az for each observer [11].

RESULTS The JND curves obtained from the two types of CRT monitors are shown in Figure 4. There were large variations in JND across the background levels. The JND at the dark end reached u p to 15 times the minimum value. The middle-to-low range of the gray scale between 50 and 120 s h o w e d the highest contrast sensitivity; a very low gray-level difference (1 or 2) could often be distinguished. The JND increased again slowly as the gray level increased to the brighter end.

20

15

10

0

~

~Type

'

-I

50

I

'

I

'

I

rnoni!or

I

100 150 200 Background Gray Level

'

250

FIGURE 4. Just-noticeable-difference (JND) curves obtained from a psychophysical experiment. The JND was highest at the dark end of the gray scale and lowest in the middle to low range between 50 and 120, The JND increased slowly as the gray scale increased to the brighter end. Error bars represent standard deviation.

Vol. 1, No. 2, October 1994

Figure 5A is a typical chest radiograph, and Figure 5B is its processed version. In the processed image, the structures of the bronchial tree, details of the vertebral column, and retrocardiac vascular structures are visualized quite well. Figures 6A and 6B show another pair of unprocessed and processed images that include multiple

IMPROVED

VISUALIZATION

OF SIMULATED

NODULES

nodules in the retrocardiac and subdiaphragmatic areas. The processed image shows significantly improved visualization of nodules in these areas. The results of the ROC analysis are summarized in Figures 7A and 7B. There was a significant improvement in the accuracy of mediastinal nodule detection (Fig. 7A). The ROC areas for these nodules were 0.69 for the unprocessed images and 0.79 for the processed

A A

B FIGURE 5. Routine chest radiograph before (A) and after (B) processing. In the unprocessed image, mediastinal structures are poorly depicted. In the processed image, visualization of the tracheobronchial tree, vertebral columns, and pulmonary vessels behind the heart and the diaphragm is enhanced.

B FIGURE 6. Chest radiographs with multiple lung nodules. A. Unprocessed image showing hardly visible nodules (arrows) behind the heart and the diaphragm. B. Processed image showing clear visualization of the nodules (arrows).

97

KIM ET AL.

Vol. 1, No. 2, October 1994

DISCUSSION

1.0

O "

0.8 ./.. i..............

..~.........~... .-''''''-''''''-''~''''''''''''''*'''''''''/

0.6

"

O

r

O

0.4

..........Unpr0r

13..

p
ID b,.

I-"

0.2

0.0

I

'

I

'

I

'

I

'

0.2 0.4 0.6 0.8 False Positive Fraction

0.0

1.0

A 1.0

g

,m

0.8 / ./ . . ./ ... . ... . .. .. j. .... ... . ... ..j ...... . .//// ./"""

r

u_

0.6

, m

r

o 0.4

a_

/'~npr0cessed(Az-0.64) p<0.1

/" n,..

0.2

0.0

'

0.0

'

I

'

I

'

I

0.2 0.4 0.6 0.8 False Positive Fraction

'

1.0

B FIGURE 7. Comparison of averaged receiver operating characteristic (ROC) curves for simulated lung nodule detection with processed and unprocessed images. Nodules were placed in the lung field (A) and behind the heart and the diaphragm (B). Az = area under the ROC curve.

images ( P < 0.05). For the lung nodules, a better level of detection was achieved (Fig. 7B). The Az values for the lung nodules were 0.64 for the unprocessed images and 0.75 for the processed images. However, the statistical significance was low ( P < 0.1).

98

We have investigated the JND characteristics of a CRT monitor to evaluate the performance of the CRT display in visualizing diagnostic information. A typical CRT monitor is able to display 256 gray levels. However, the JND characteristics indicate that the perceivable dynamic range of the CRT should be narrower than the available 256 gray levels. The sensitivity at both ends of the gray scale was significantly low. The low sensitivity at the dark end seems to occur because of light reflection at the CRT surface and because of the reduced sensitivity of the human eye at a low luminance level [6]. The low sensitivity at both ends of the gray scale may lead to the reduction of observer performance in the detection of low-contrast lesions, as has been noted in a number of studies [12, 13]. Therefore, we restricted the effective dynamic range of the CRT to the moderate 226 gray levels, excluding both ends, to avoid the possible loss of important diagnostic information. We believe that a carefully designed image-processing technique could carry the diagnostic information conmined in an image to an observer faithfully with the CRT monitor. We attempted to achieve two objectives with the proposed method. The first was to compress the dynamic range of image data to fit in the effective dynamic range of the CRT, while preserving clinically important information as much as possible. The second was to compensate for the variations in the contrast sensitivity of the CRT shown by the JND measure to provide equal visibility of objects having various background densities. To achieve these objectives, we first suppressed the dynamic range of the low-frequency component of the mediastinum to one-half by applying the nonlinear gradation function H, which has a slope of 0.5 in the mediastinal range and unity slope elsewhere. Slopes of 0.3 and 0.2 were tested, but they resulted in unacceptable images with excessive suppression of the mediastinum. We did not attempt to suppress the low-frequency component of the imaged lung area because it might have obscured the minute density differences between large objects in the lung field that are crucial to the diagnosis of certain lung diseases. Along with the selective suppression of the dynamic range, we enhanced the local contrast of the mediastinum to compensate for the relatively low contrast sensitivity of the CRT in the bright range of the gray scale. We

Vol. 1, No. 2, October 1994

did this by amplifying the high frequency with the function c~, having a maximum gain of 3.0 for the mediastinum and unity gain for the lungs. The contrast sensitivity of the CRT in the bright range is about half of that in the moderate range (Fig. 3). With a maximum gain of 2.0, mild enhancement was observed. However, it was not sufficient to provide improved observer performance in the preliminary evaluation test. Changing the gain to higher than 5.0 resulted in excessive enhancement. Image noise and artifacts in the subdiaphragm became noticeable, and the excessively enhanced mediastinal structures appeared distracting. With the smooth transition of gain between the lungs and the mediastinum, we could not find the ringing artifact that has been noted in some previous studies [2, 14]. For the choice of filter kernel size, we used the results of McAdams et al. [2]. They suggested that the 50 x 50 kernel (17 x 17 ram) is well suited for chest processing because it covers most of the edge detail within the chest. We found that this kernel size produced natural and acceptable images in our study. Filter kernels of 10 x 10 and 20 x 20 pixels enhanced more fine edges but were found less effective for improving the visibility of mediastinal structures. We did not attempt to enhance the local contrast of the lung area with this method because it may produce false positives [12]. However, because the dynamic range of the mediastinum is suppressed and the image data were projected onto the full effective dynamic range of the CRT, the lung area showed enhanced image contrast and was displayed on the most sensitive gray range of the CRT. Consequently, the lung area and the mediastinum were visualized better in the processed images. The improved detection of both lung nodules and mediastinal nodules in the observer performance study seems to reflect this result.

IMPROVED

VISUALIZATION

OF SIMULATED

NODULES

It is important in digital radiography and PACSs to determine whether CRT displays can provide adequate observer performance for diagnosis in radiographic applications. The encouraging results of the observer performance test in this study seem to indicate a positive answer to this question, provided that care is taken with the viewing environment and an adequate processing technique is applied to the image display, REFERENCES 1. Fraser RG, Sanders C, Barns GT, et al. Digital imaging of the chest. Radiology 1989; 171:297-307. 2. McAdams HP, Johnson GA, Suddarth SA, Sherrier RH, Ravin CE. Implementation of adaptive filtration for digital chest imaging. Opt Eng 1987;26: 669-674. 3. Abe K, Katsuragawa S, Sasaki Y, Yanagisawa T. A fully automated adaptive unsharp masking technique in digital chest radiograph. Invest Radiol 1992;27:64-70. 4. Sherder RH, Johnson GA. Regionally adaptive histogram equalization of the chest, IEEE Trans Med Imaging 1987;MI-6:1-7. 5. Dwyer SJ Ill, Stewart BK, Sayre JW, et aL Performance characteristics and image fidelity of gray-scale monitors. RadioGraphics1992; 12:765-772. 6. Baxter B, Ravindra H, Normann RA. Changes in lesion detectability caused by light adaptation in retinal photoreceptors. Invest Radio11982; 17:394-401. 7. Bronskill MJ. Experience with the SMPTE test pattern in quality control of magnetic resonance images. Proc SPIE 1984;486:180-184. 8. Green DM, Swets JA. Signal detection theory and psychophysics. New York: Krieger, 1974. 9, Lee TS, Min BG. Image restoration in digital radiography using dual sensor Wiener filter. Med Phys 1991; 18:1132-1140. 10. Sherrier RH, Johnson GA, Suddarth SA, Chiles C, Hulka C, Ravin CE. Digital synthesis of lung nodules. Invest Radio11985;20:933-937. 11. Metz CE. Some practical issues of experimental design and data analysis in radiological ROC studies. Invest Radio11989;24:234-245. 12. Goodman LR, Foley WD, Wilson CR, Tikofsky RS, Hoffman RG. Pneumothorax and other lung diseases: effect of altered resolution and edge enhancement on diagnosis with digitized radiographs. Radiology 1988; 167:83-88. 13. Goodman LR, Foley WD, Wilson CR, Rimm AA, Lawson TL. Digital and conventional chest images: observer performance with film digital radiography system. Radiology 1986; 158:27-33. 14. Plewes DB, Vogelstein E. A scanning system for chest radiography with regional exposure control: practical implementation. Med Phys 1983;10: 655-663.

99