# Enhancement of chest radiographs with gradient operators

Post on 11-Mar-2017

214 views

Embed Size (px)

TRANSCRIPT

<ul><li><p>IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 7, NO. 2 , JUNE 1988 109 </p><p>Enhancement of Chest Radiographs with Gradient Operators </p><p>JOHN S . DAPONTE AND MARTIN D. FOX, MEMBER, IEEE </p><p>Abstract-The use of gradient operators for image enhancement has been widely reported in the literature, but they have not been used routinely in the medical arena, particularly in the most common ra- diographic plain film procedure, chest radiographs. Gradient opera- tors such as Sobel and Roberts operators, not only enhance image edges but also tend to enhance noise. Overall, the Sobel operator was found to be superior to the Roberts operator in edge enhancement. A theo- retical explanation for the superior performance of the Sobel operator was developed based on the concept of analyzing the x and y Sobel masks as linear filters. By applying pill box, Gaussian, or median fil- tering prior to applying a gradient operator, noise was reduced, but the pill box and Gaussian filters were much more computationally ef- ficient than the median filter with approximately equal effectiveness in noise reduction. </p><p>INTRODUCTION N image processing an edge is defined as a difference I in image characteristics within a local region [ 13. Edge </p><p>detection, a technique that has been used for the purpose of enhancing radiographic images, is carried out by com- puting the rate of change with respect to gray levels for two adjacent regions within an image [2]-[4]. Gradient operators are a well-known form of edge detection which are based on digital approximations of partial differential equations [ 5 ] . Alternative techniques for estimating derivatives of a digital image include differentiation of least-square fitting polynomials and Wiener filter approx- imations 161, but these techniques tend to require more computation than the gradient operator approach. </p><p>The ability to emphasize the fine details of a radio- graphic plain film image while suppressing noise was demonstrated qualitatively for cephalometric radiographs by employing a median filter prior to applying a gradient operator 171. Although the ability of a median filter to suppress noise has been previously established 181, [9] our results indicate that approximations to the pill-box blur and the Gaussian low-pass filter yield a similar amount of noise reduction with much better computational efficiency and subjectively no loss of useful image detail. </p><p>Edge detection can either provide an enhanced image for a human observer, or allow significant data reduction prior to further computer processing. A potential appli- cation of this is the early detection of lung nodules [ 101, </p><p>Manuscript received July 26, 1987; revised January 20, 1988. This work was supported in part by grants from the Whitaker Foundation and the State of Connecticut Department of Higher Education. </p><p>J . S . DaPonte is with the Department of Computer Science, Southern Connecticut State University, New Haven, C T 06515. </p><p>M. D. Fox is with the Department of Electrical and Systems Engineer- ing and Radiology, University of Connecticut, Storrs, CT 06268. </p><p>IEEE Log Number 8820192. </p><p>[ 111. Previously developed computer controlled lung nod- ule searches have often used edge detection as a prelimi- nary processing step [ 121, [ 131. </p><p>The primary objective of our study was to investigate the application of gradient operators to the most common radiographic procedure, plain film chest X-rays [ 141 using a frequency domain formulation to explain the effects of Sobel and Roberts gradient operators on the image. Im- ages were digitized using a video frame grabber on a mi- crocomputer based system [ 151, [ 161. In addition, we have quantitated the advantages of combining various prefilter- ing techniques with the Sobel gradient operator, which proved to be the most efficient, and compare the signal- to-noise ratio as well as the amount of time required for each combination. </p><p>THEORY </p><p>A common method for computing the rate of change with respect to gray level in an image is to use gradient operators which in their simplest form consist of partial difference operators. While these partial difference oper- ators lead to relatively simple calculations, they suffer from a dependence on directionality. Thus, more nonlin- ear techniques derived from two-dimensional filtering may be necessary to effectively display localized changes in gray level about a pixel. </p><p>Two-dimensional linear digital filtering of images is ac- complished by storage of values, multiplication by con- stants, and addition. For example, consider the M X N input image X ( mp, nq) which has the two-dimensional z- transform [ 171. </p><p>M N </p><p>X(Z,, Z,,) = c c x ( m p , nq)Z,"Z," (1) m = O n = O </p><p>where 2, = and Z,, = eJ2Tz'4, U and v are spatial frequencies and p and q are pixel spacings in the x and y directions, respectively. If we apply a filter to the image X(Z,, Z , ) , the input and output images can be related through the expression </p><p>where H(Z,, Z,,) is the transfer function of the filter. The two-dimensional digital convolution operation y = x ** h can be defined [ 171 </p><p>w m </p><p>Y(m, n ) = c c h(a, b ) X ( m - a , n - b ) ( 3 ) a = - w h = --m </p><p>0278-0062/88/0600-0109$01 .OO @ 1988 IEEE </p></li><li><p>110 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 7 . NO 1, I U N E 1988 </p><p>which is the corresponding operation to the filtering op- eration of (2) in the image domain where in this context m, n and a , b are pixel indexes. Note that by substitution for the complex variables Z , = e Z , , = e '**li4 in ( 2 ) we can determine the transfer function H ( u , U ) in terms of spatial frequency. </p><p>Gradient operators can be distinguished from simple convolutional filters as defined by ( 3 ) in that they require the use of two masks, one to obtain the x-direction gra- dient, the second to obtain the y-direction gradient, the results of which are then combined to obtain the orthog- onal components of a vector quantity whose magnitude represents the strength of the gradient or edge at a point in the image and whose angle represents the gradient an- gle. In general, while gradient operators are thus nonlin- ear operators, which are not necessarily quantifiable, through linear filter analysis, the x and y masks can be separately understood as linear filters. </p><p>A. Sobel Operator The Sobel operator can be defined by the two kernels </p><p>[21, [31 </p><p>- 1 0 1 -1 -2 -1 </p><p>S, = - 2 @ 2 and S, = 0 @ 0 -1 0 1 1 2 1 </p><p>which are sequentially convolved with the original image file. The result is stored in the center or circled pixel and can be represented as a vector quantity with magnitude </p><p>lY(mP, n q ) ( = (x(mp, nq) ** + I X h P ? nq) ** s,jl (4) </p><p>and angle </p><p>Arg [ Y b P , n s ) ] </p><p>= arctan [x(mp, nq) ** ~ , / x ( m p , nq) ** s,] ( 5 ) where ** represents two-dimensional convolution and i a n d j are unit vectors in the x and y directions, respec- tively. The two-dimensional z transform of (4) is </p><p>Y ( z u , Zt , ) = X(Zu, Zr,) sx(Zu, Zrl)i </p><p>+ X(ZU> Z , ) S,(ZU> Z , ) j ( 6 ) </p><p>and the transfer function of the Sobel operator can be ex- pressed </p><p>HS(ZU, ZV) = WU9 Z , , > / X ( Z u , Z , , ) = S,(Z, , ZV)i + q z , , Z , ) j . (7) </p><p>The z transform of the Sobel operator S, will be Z , , ) = '(1 + z,' + Z ? ' ) </p><p>- zb(1 + z,' + z;') + z,' - z,+' ( 8 ) </p><p>T </p><p>- 5 - 5 </p><p>(b) </p><p>Fig. 1 . Three-dimensional mesh plots of the magnitude of the spatial fre- qlrency response of a single component of the (a) Sobel and (b) Roberts operators. Results are plotted out to the sampling theorem limit in each case. Note that the Sobel response is better controlled orthogonal to the direction of differencing. </p><p>which can be simplified to obtain </p><p>&(U, U ) = 2j sin (2xup)[2 + 2 cos (27rvq)I. (9) Similarly, </p><p>sY(u, U ) = 2 j sin (27ruq)[2 + 2 cos TUP)]. (10) Note that fo rp= q, &(U, U ) = &(U, U ) aftera 90 degree rotation. A three-dimensional plot of 1 S., ( U , U ) 1 is shown in Fig. 1. </p><p>B. Roberts Operator The Roberts operator convolves the kernals </p><p>0 + 1 - 1 0 r-45 = </p><p>O Q r4s = -1 @ with the image to obtain orthogonal gradients along the 5 4 5 degree axes. These operators are convolved with the input image x(mp, n q ) to obtain the complex modified image </p><p>Note that any orthogonal pair of axes, in this case k45 can provide an estimate of the strength and angle of an </p></li><li><p>DA PONTE AND FOX: CHEST RADIOGRAPHS WITH GRADIENT OPERATORS 111 </p><p>Gradient </p><p>F i l t e r Magnl tude </p><p>3 '4 Phase </p><p>G = Origlnal Image M = F i l t e r Mask </p><p>M = Second Gradient Mask M 1 = Fi rs t Gradient Mask 2 3 </p><p>Fig. 2 . Block diagram of experimental schema </p><p>edge. The more important issue is how well the operator can detect edges, particularly in the presence of noise. </p><p>blur operator which is a convolution type filter with the following mask: </p><p>1 1 1 1 1 The two-dimensional 2 transform of the Roberts operator r45 1s </p><p>2,) = 2;' - 2," (12) and along the 45 degree line if p = (U' + U * ) is the radical coordinate i t t h e frequency domain and p = q, we </p><p>1 2 2 2 1 </p><p>1 2 0 2 1 </p><p>1 2 2 2 1 </p><p>1 1 1 1 1 have U = v = p / J 2 and we obtain </p><p>~ 4 5 ( p , 450) = -2 j sin (2apq/&) -( 13) </p><p>(14) The second mask is a 3 x 3 approximation to a Gauss- </p><p>~ - 4 5 ( p , 450) = -2j sin (27rpq/Jz e-*j*q/fi. ian low-pass convolution type filter defined as </p><p>A three-dimensional plot of I r45(Zu, Z,,) 1 is shown in Fig. 1 for comparison to the Sobel operation. </p><p>Inspection of the frequency response plots shown in Fig. 1 reveals that the Sobel operator tends to filter out spatial </p><p>1 2 1 </p><p>2 @ 2 </p><p>1 2 1 frequencies orthogonal to the direction of differencing (edge detection) better than does the Roberts gradient op- erator. This behavior in the frequency domain thus pro- </p><p>The third filter was a 3 x 3 median filter with the fol- lowing mask: </p><p>vides a possible explanation for the superior performance of the Sobel operator in detecting edges in a noisy envi- ronment. </p><p>While gradient operators are commonly used for edge </p><p>1 1 1 </p><p>1 0 1 </p><p>1 1 1 detection, they have the drawback of enhancing image noise. One way of reducing this problem is to filter high The fourth was a median with the spatial frequencies containing noise prior to applying the mask: gradient operators. This process can best be described with the aid of a functional block diagram as depicted in Fig. 2. If we let G denote the gray levels of the original image and let MI denote the filter mask, then the first step is to filter the image. Next the filtered image is processed by two gradient masks, M2 and M 3 , producing orthogonal components which are passed to a magnitude operator which takes the absolute value of each pixel in the or- thogonal matrices and sums them. Finally, the output magnitude is compressed to fit in the available gray lev- els. </p><p>Four types of noise reduction filters were investigated in this study for use prior to the application of gradient operators. The first was an approximation to the pill-box </p><p>1 1 1 1 1 </p><p>1 1 1 1 1 </p><p>1 1 0 1 1 </p><p>1 1 1 1 1 </p><p>1 1 1 1 1 </p><p>For each of these filters the pixel of interest located in the center of the mask has been circled. </p><p>The pill-box and the Gaussian operators smooth the im- age by convolving the corresponding mask, while the me- dian filters are a nonlinear operator that sort each pixel in its neighborhood according to intensity and replaces the </p></li><li><p>112 </p><p>center pixel by the median of the sort. For example in a 5 X 5 region of interest there are 25 points that must be sorted when using a median filter. Thus, as the mask is translated about the region of interest, a large number of pixels must be sorted over and over again resulting in a very time consuming process. </p><p>METHODS The system used to acquire and process the images pre- </p><p>sented in this paper consists of a commercial frame grabber' installed on a personal compute? [ 151 containing 640 Kbytes of memory. A functional block diagram of a typical frame grabber system is shown in Fig. 3 . </p><p>The frame grabber board is memory mapped into the address space of the host computer and has the ability to store two images at a time in its on-board memory. Ap- plication software resides in the host computer's memory and interacts with the data via the bus interface. </p><p>Radiographic images were captured from the teaching files at the University of Connecticut Health Center using a video camera which adheres to the RS-170 video stan- dard [IS]. Once the board had digitized an image, it was passed through an input look-up table which could per- form real time gray scale manipulations before being stored in the frame memory. Then the various software algorithms in the host memory were applied to obtain the processed image which was passed through an output look-up table and converted to analog before being dis- played on an external video monitor. Hard copies of out- put images were produced using a thermal video ~ r i n t e r . ~ </p><p>RESULTS A variety of chest radiographs were processed using the </p><p>techniques described in this paper. We found that the So- bel operator subjectively produced much better results than did the Roberts operator for the same image. The output edges of the Roberts operator were so faintly dis- played that histogram equalization was necessary to make them visible which resulted in the unacceptable output de- picted in Fig. 4. Thus, the resulting sequence of photo- graphs are limited to demonstrating the effects of the So- bel gradient operator on the same image combined with various types of filtering. </p><p>Fig. 5 contains three unprocessed chest radiographs, a normal (a), a subject with a hamartoma (b), and subject with a squamous cell carcinoma (c). The results of apply- ing the Sobel operator to these images without prefiltering is illustrated in Fig. 6. Note the occurrence of noise in these images demonstrating the amplification of noise caused by applying the Sobel gradient operator. The im- ages in Fig. 7 were obtained by first applying the pill-box blur and then the Sobel operator to the chest radiographs. In these cases, the image noise subjectively appears to have been reduced while the edges, containing signal, can be distinguished. </p><p>'Imaging Technology, Inc 'IBM PC/AT Series 'Mitsubishi Model P60U </p><p>IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 1, NO. 2. JUNE 198X </p><p>[ I 1lel D m </p><p>Fig. 3 . Block diagram of a typical frame grabberiimage procesaor system. </p><p>Fig. 4. Typical chest radiograph after Roberts operator and histogram equalization. </p><p>Following a similar pattern the results of prefiltering with a 3 X 3 Gaussian, a 3 X 3 median, and a 5 X 5 median to each of the original chest radiographs are pre- sented in Figs. 8, 9, and 10. Subjectively, all of the pre- filtered images appear to have reduced noise when com- pared to cases in which the Sobel is applied without prior filtering. </p><p>In order to compare these prefiltering techniques, a quantitative approach was used in which both the signal- to-noise ratio and computational time of each filter was considered. The signal-to-no...</p></li></ul>