Ahmad Al
Transcription
Ahmad Al
Determination and enhancement of X-Ray image quality using Shannon’s information capacity metric Ahmad Al-Nassre University of Maastricht September 6, 2009 Abstract With the technological progress digital X-Ray facilities have been established in the past few years. These facilities have to be calibrated and attended manually by specialists for the most part. The challenge is to assist those specialists among other things in calibrating X-Ray scanners and provide them with rationalised and automated tools and methods. By adjusting and improving of the here presented quality metric, the Doctor’s demands can be more efficiently taken into account. In this thesis the focus is made on quantifying X-Ray image quality using the Shannon’s information capacity metric. Keywords: Image quality assessment, XRay imaging, Shannon capacity 1 INTRODUCTION The storage plate technique was the first in the projection radiography procedure field. There has been an additional digital system besides the computed radiography (CR) with the flatdetector radiography for a few years now. Whereas a cassette with a specialized storage plate is needed for the storage plate technique, the flatdetector technique directly transmits the information to a computer system. Digital systems offer a seperated optimization for the individual components of the image generation: image data acquisition, image data processing, image rendition, image archiving and image transmission. Irrespectively of the applied digital system, the computer system is fed with image raw data, which is now ready for image processing. The aim of image processing is to adjust the contrast and the luminance and to emphasize particular details in the object. This can improve and maintain the image quality. However an improper use of image processing tools can decrease the image quality and lead to a false diagnosis. There is no direct correla- tion between the radiation dosis and the optical density. Hence the exposure latitude is significantly larger in comparison to the film-foil-combination. Therefore over- and underexposure can be extensively settled. Depending on the exposure, noise increase can restrict the image quality. The adoption of digital radiography in the medicine, poses the question of image quality and in its resulting diagnosis raliability compared to the conventinal filmfoil-system. Image postprocessing tools offer the facility to adjust the luminance, contrast and sharpness. Furthermore certain structures can be emphazised in dependence of specific problems. But IP also has some drawbacks, because it implicates the risk of suppressing diagnostical relevant structures, which could lead to a misinterpretation. These developments shows the prime importance of an objective image quality assessment for X-Ray imaging. This thesis evaluated an information theoretic metric that calculates the Information Capacity (IC) proposed by Papalazarou [2]. Here the quality of an image is described in terms of its IC. Shannon has laid the foundation for this capacity in his approach about noisy channel coding [7]. Since the noise in X-Ray images is signal dependent and non-additive the X-Ray image signal has to be adjusted first to fit the Shannon theorem. This approach is structured as follows : Section two will give an introduction into the Determination criteria of digital images. In Section three the full-reference objective image quality metric will be illuminated. The distortions used in the experiments are described in Section four. Shannon’s entropy is explicitly presented in chapter five. The experiments performed in this research are described in Section six, before finally the results are specified in Section seven. 2 Determination criteria of digital images Both analog and digital images have the same quality distinguishing properties, namely noise, sharpness, Ahmad Al-Nassre University of Maastricht Determination and enhancement of X-Ray image quality using Shannon’s information capacity metric contrast and artifacts. However those quality varieties have different reasons. In this approach the quality of digital X-Ray images is of particular importance. Therefore the focus is put on the properties of digital images, whereby their quality is influenced and the way it is affected. 2.1 pretentions of the beholder. Sharpness for one may be inacceptable for the professional beholder’s pretentions. Thus image sharpness is, at least by observing an image, a subjective quality criterion since in particular the observed images finest observable structure and sharpness is up to the individual beholder’s acuity as well. Noise Figure 2: sharpness Figure 1: Example of quality properties in a signal With the appearence of modern techniques, digital imaging has been widely spread. In these acquired digital images different types of noise may occur, like photon noise, thermal noise, on-chip electronical noise, KTC noise, amplifier noise and quantization noise. In this approach digital X-Ray images are of importance and photon noise mostly contributes of the noise which ammounts about 70 - 80%. The remaining 20-30% occur due to distortions created by the X-Ray equipment and measuring tools. Although the later mentioned proportion is humble and may have less influence on the X-Ray image compared to the other proportion, it is of prime importance for the patience’s weal to improve these equipments and measuring tools. Because by enhancing image processing tools it is possible to reduce the amount of X-Ray radiation or contrast medium, which patients are exposed to. 2.2 Sharpness Sharp images are characterised by a high contrast between individual objects (areas, contours) in an image. But when is it possible to say that a certain image is sharp or unsharp? On the face of it, this question seems to be redundant, because an image can immediately be judged as sharp or not by the first view. In this regard, first problems occur by enlarging the image. An image could prove to be keenly unsharp by enlarging it a certain factor. The question if the image as such is sharp or not is up to the preferences and 2.3 Contrast The contrast characterises the difference between the pale and the dark areas in a visual representation. It is a distinction criteria of an images pale process or between two pixels. Furthermore the contrast range describes luminosity discrepancy between the palest and darkest pixel of an image. 2.4 Artifacts In the domain of computer graphics artifacts are described as a visible negative representation of a rendered image which is not caused by the original input data. There are several reasons for such inaccurate representations. Visual artificats may occur in image compression (JPEG) or in video compression (MPEG) due to missing information. These for the human eye hardly recognisable artifacts are unavoidable in compression, but the improved image processing algorithms try to minimize them. Grafical artifacts are also caused by screening an object into the two-dimensional display resolution. This disturbance is minimized by Antialiasing. In X-Ray images artificats may also occur caused among other things by X-Ray scattering. The figure 3 exemplifies those artifacts marked by the white arrow. 3 SHANNON’S INFORMATION CAPACITY METRIC There are several Objective Image Quality metrics (OIQ) which can be mainly classified into three cate- (v. September 6, 2009, p.2) Ahmad Al-Nassre University of Maastricht Determination and enhancement of X-Ray image quality using Shannon’s information capacity metric Figure 3: Artifact in the interpetrous region (white arrow) gories: no-reference, reduced-reference and full-reference metrics. Shannon’s information capacty metric belongs to neither of the three categories. It only exhibits some similitudes to the full-reference model.The full-reference model requires a perfect reference image to accurately obtain the quality of images. In former researches the full-reference IQ metric was applied to determine the qualitiys decline of compressed images. Therefore the compressed image was compared to the original image and a quality score was evaluated. It is nearly impossible to get a perfect reference X-Ray image hands-on. The full-reference model has been improved and adjusted in further researches to get an approximated perfect reference image. To obtain this image, a certain amount of X-Ray images of the same object in the same situation has to be made. Subsequently, the average of the X-Ray image set which serves as a reference image is taken in this metric. On the other hand, in Shannon’s information capacity metric this averaged image is regarded as a noiseless version of the objective instead of a reference image. In figure 4 the process is simplifiedly illustrated. The noiseless version mentioned above is subtracted from the input image to seperate the noise from the input signal. Accordingly a fourrier transformation of the signal and the noise of the image is done in order to subsequently figure out the Information Capacity. Figure 5 shows this approach apparitional. Figure 4: Information Capacity Model Image Set (1...M) (m >> 1) Average Input Image "I" Noiseless Image "S" Make Noise Independent mapping function A(x) = ∫(1/σ(x) ) dx Noise Image A(I) - A(S) = "N" Noise Power Spectrum math In the context of research into communication through telephone wires, Shannon (1948) defined the first historical metric of information - the Shannon entropy: H(X) = − n X i=1 p(xi ) ∗ log(p(xi ). (1) Signal Power Spectrum Calculate Information Capacity Figure 5: Scheme Information Capacity (v. September 6, 2009, p.3) Ahmad Al-Nassre University of Maastricht Determination and enhancement of X-Ray image quality using Shannon’s information capacity metric The random variable X can be expressed in terms of the underlying distribution p(xi ), while n is the number of states that the (discrete) signal can assume. The information term of Shannon does not evaluate the semantical content or the meaning of a message, but its unpredictability (contingency, disarrangement). The more information a message has, the more insecure one is about the syntactical content. The Shannon entropy is the classical measure for information.The name ’Entropy’ goes back to the analogue meaning of the formula in the thermodynamics. Mutual information (MI), sometimes also called transinformation, is a quantity of the information theory, which indicates the statistical correlation of two random variables. Here it indicates the correlation between the distribution X and an output Y . M I(X; Y ) = H(X) + H(Y ) − H(X; Y ) = X p(x, y) H(Y ) − H(Y | X) = p(x, y) log2 p(x) ∗ p(y) 4 Distortions Since it was impossible to inverse the image processing tools integrated in the X-Ray image facility, for instance noise reduction or deblurring techniques, artificial distortions were added. Therefore the Information Capacity Model was extended with a distortion module, shown in figure 6. This distortion module consists of noise addi- (2) X,Y For this approach another aspect of the Shannon theory is of quite importance and will be considered in respect of coding. The channel capacity is equal to the already introduced mutual information, maximized over all input distributions. For the case of a one-dimensional, zero-mean signal of bandwidth B, which is corrupted by independent additive white Gaussian noise, the maximum information capacity of the channel CShannon can be written in the simple form [1] : CShannon = max(H(Y )−H(Y | X) = B ∗log2 (1+( S 2 ) ) N (3) The Shannon-theorem deals with the relation between the available bandwidth, the Signal to Noise ratio and the outcome of this is the maximum data transmission rate. In this approach a generalization of the equation above for two-dimensional signals, where the additive noise is not white and both the signal and noise are zero-mean and bandwidth-limited, is given in equation 4: fN y fN x Z Z C0 = log2 (1 + P (fx , fy ) )dfx dfy N (fx , fy ) (4) −fN x −fN y Where C0 is the Channel capacity in bits per second, P (fx , fy ) and N (fx , fy ) are the signal and noise power spectra, f is the frequency in Hz and the integration is taken up the Nyquist frequency. Figure 6: Extended Information Capacity Model tion and blur filters which were seperately added to the input images. Gaussian noise is supposed to simulate the photon noise in X-Ray images and is increasingly added to the input images. Details of obtaining the experimental X-Ray image set will be explained in Chapter 5. Figure 7 shows an X-Ray image before and after adding gaussian noise. An odd gaussian blurring filter, illustrated in figure 8, is apllied on the input image by increasing the window size to intensify the blurring. An extension is shown in figure 9. 5 EXPERIMENTS The experiments were all performed on an iMac 24” with a 2.93 GHz Intel Core 2 Duo processor and 4 GB DDR3 RAM under Mac OS X using Matlab 7.4 (R2007a). In Chapter 3 it was already mentioned that the Information Capacity metric only exhibits some similitudes to the full-reference model. Its experimental design differs in various aspects, which will be illuminated in the following part. Since there were some technical difficulties to accomplish a similar design of experiment performed by C. Papalazarou in her research [1], an alternative experimental setup obtaining the required X-Ray set was performed. Different X-Ray images were made using a Provatec X-Ray facility. For this approach it was assumed that these images are noiseless and they were (v. September 6, 2009, p.4) Determination and enhancement of X-Ray image quality using Shannon’s information capacity metric Ahmad Al-Nassre University of Maastricht • • Figure 9: X-Ray of knee before and after gaussian blur addition • • • Figure 7: X-Ray of shoulder before and after gaussian noise addition Figure 8: Odd blurring filter with a window size of 10 and a sigma of 5 seen as the signal in the original image. The objective of these experiments is to figure out how Shannon’s IC responds to miscellaneous types of distortions. In this research the distortions gaussian noise and blurring are of particular interest. The gaussian noise was added increasingly in one hundred steps to obtain a representative X-Ray set with the power of hundred images. In each step the noise was seperated by subtracting the noiseless image from the noisy image. Accordingly a fourrier transformation of the signal and noise image is done in order to subsequently figure out the signal to noise ratio. In the context of imaging it can be enunciated that the greater the signal to noise ratio, the more information the image contains and hence it indicates an image of a high quality. It can be definetly expected to lose quality by adding noise. In the second part of the experiments several blurring filters with different parameter settings were applied to the X-Ray images. Only two of them will be presented in this approach. The Gaussian and Laplace-filter. For the Gaussian-filter the parameters, σ i.e. the standard deviation and the window size are important, whereas for the Laplace filter the value of the diffusion coefficient and the number of steps influence the blur of the image. The Gaussian-filter was applied by increasing the window size in fifty steps, whereas the parameter sigma took the value of half of the respective window size. Furthermore the Laplace-filter was used with the diffusion coefficient α in a similar as aforementioned experimental setup. For both blurring filters, in each step the noise was sper- (v. September 6, 2009, p.5) Ahmad Al-Nassre University of Maastricht Determination and enhancement of X-Ray image quality using Shannon’s information capacity metric ated by subtracting the noiseless image from the blurred image. Accordingly a fourrier transformation of the signal and blurred image is done in order to subsequently figure out the signal to noise ratio. In the IP, blurring is also used as a denoising technique. Consequently it is expected of a small quality increase. RESULTS By contemplating this metric the results of the experiments were in some respects surprising and unexpected. The noise addition test yield as expected in a decrease of the image quality score. Figure 11 shows this result numerical and figure 10 visual. The IC of the noisy image already indicated the quality loss compared to the noiseless image, and figure 12 shows explicitly that no medical relevant information can be determined anymore. However, the blur addition test results are surprising and raise some questions. In respect to the IC the blurring tests yield an increase of the quality score of the blurred image??. The initial improvement can be explained by the fact that every image contains some certain type of noise, which has been removed by blurring the image . Actual the quality score of the blurred image should have been drop down by increasing the intense of the blur factor. But according to the IC the opposite effect occurs. The quality score increases continuously even at a gaussian blur windows size of 50 units. The reason for this is that the IC determines the information content e.g. entropy of an image. Blurring an image has the effect of increasing the information content of it . Despite the high quality score of the blurred image, it is impossible to achieve any medical relevant and reliable information out of it. A mathematical technical motivation could be that in Shannon’s signal to noise ratio, the noise and signal strength are weighted differently. A possible explanation for this is, that the signal and noise strength are averaged over the frequency bins that the fourier transform calculates. The noise reduction is felt in the entire spectrum and the decrease will be in every frequency bin [2]. 7 Conclusion Figure 10: Hip X-Ray before and after 3 steps of noise addition 4 12 Shannons IC of noisy hip x−rays x 10 11 10 9 Quality score of x−ray 6 8 7 6 5 4 In this approach an image quality metric was proposed, the Shannon’s information capacity metric. It emerges with some insignificant flaws as a reliable and applicable image quality metric. In the noise experiments its performance was sufficient. On the other hand the adjustments of the parameters of the blur filter experiments were difficult. This issue keeps a challenge in future researches. Summing up, once the optimal parameters found it, it obtains a substantial indicator for comparison to other specific metrics. 3 2 0 10 20 30 40 50 60 noisy hip x−rays 70 80 90 100 Figure 11: Basic noise addition test results of Shannon’s IC of hip X-Ray (v. September 6, 2009, p.6) Determination and enhancement of X-Ray image quality using Shannon’s information capacity metric Ahmad Al-Nassre University of Maastricht Figure 12: Knee X-Ray before and after 3 steps of noise addition Figure 14: Spine X-Ray before and after gaussian blurring with windows size 25 4 10 Shannons IC of noisy knee 2 x−rays x 10 6 16 8 14 7 12 6 Quality score Quality score of x−ray 9 5 Shannons IC of blurred Spine x−ray x 10 10 8 4 6 3 4 2 0 10 20 30 40 50 60 noisy knee 2 x−rays 70 80 90 100 2 Figure 13: Basic noise addition test results of Shannon’s IC of knee2 X-Ray 0 5 10 15 20 25 30 Blur filter window size 35 40 45 Figure 15: Basic blur addition test results of Shannon’s IC of spine X-Ray (v. September 6, 2009, p.7) 50 Ahmad Al-Nassre University of Maastricht Determination and enhancement of X-Ray image quality using Shannon’s information capacity metric References [1] Chrysi Papalazarou, Rudolph M.Snoeren,Frans M.J. Willems, Peter H.N. de With,Han Kroon and Peter Rongen Towards a full-reference, informationtheoretic quality assessment method for X-Ray images 2008 [2] Jop Daalmans Objective Image Quality Assessment for X-Ray imaging May 2008 [3] Blanchet, Grard, Maurice Charbit Digital signal and image processing using Matlab [4] Robert M. Gray and Lee D. Davisson An Introduction to Statistical Signal Processing 2003 [5] Ting-Chung Poon Partha P. Banerjee Contemporary Optical Image Processing with MATLAB 2001 [6] Martin Werner Digitale Signalverarbeitung mit MATLAB-Praktikum [7] C. E. Shannon, A mathematical theory of communication, Bel l Syst. Tech. J. 27, pp. 379423, 1948. [8] The Mathematical Theory Of Communication, Shannon & Weaver September 1949 Figure 16: Shoulder X-Ray before and after gaussian blurring with windows size 25 [9] Multidimensional Signal, Image, and Video Processing and Coding, John W.Woods 2006 [10] Discrete Fourier Analysis And Wavelets , S.Allen Broughton, Kurt Bryan 2009 [11] Digital Filters Design for Signal and Image Processing, Mohamed Najim 2006 6 4.5 Shannons IC of blurred Shoulder x−ray x 10 4 Quality score 3.5 3 2.5 2 1.5 1 0 5 10 15 20 25 30 Blur filter window size 35 40 45 50 Figure 17: Basic blur addition test results of Shannon’s IC of shoulder X-Ray (v. September 6, 2009, p.8)