Classification of Brain MRI in Wavelet Domain

Transcription

1 International Journal of Electronics and Computer Science Engineering 879 Available Online at ISSN Classification of Brain MRI in Wavelet Domain Prof. Bhupal Singh 1 2, Jai Singh 1 2 Department of Electronics and Communication Engineering 1 Ajay Kumar Garg Engineering College, Ghaziabad 2 Ajay Kumar Garg Engineering College, Ghaziabad 1 - er.jaisingh@gmail.com Abstract- The automatic classification of brain MRI images of a patient is an important task in clinical diagnostic for the detection of tumor/cancer or any kind of brain related disease; subsequently it will reduce the subjectivity of physician in decision making. In order to design and implement, A MRI image classification technique of Three-Stage approach, consisting of 2 nd level wavelet decomposition [1] in various non-overlapping bands, and extraction of corresponding feature set vectors employing first order statistics and principal component analysis is required. By using these approaches we have to train a Support Vector Machine [8] for the Final classification of brain MRI image. The proposed approach is expected to give better performance than the previous approaches used in the brain MRI classification. The MRI image data of a normal and abnormal person are utilized here from available resources and the problem will be carried out in MATLAB 7.12 Version by using Image processing, Wavelet & bioinformatics Toolboxes. The comparison will also be carried out at with existing conventional techniques to establish its superiority Keywords MR images, MRI, brain-tumor classifications, wavelet transform, artificial neural networks, support vector machines I. INTRODUCTION The automated classification of brain MRI of normal\pathological subject s is one of the toughest job in clinical diagnostic of Biomedical Imaging field.mri is an imaging technique used in radiology for visualizing the internal detailed structure of human soft-tissues anatomy. MRI provides a good contrast between the different soft-tissues of the human body which makes it useful in imaging brain muscles.the advantage is that it does not cause the harmful ionizing radiation to the patients. There are several types of approaches that are available in biomedical image processing but the most prominent approaches for analyzing the obtained image is in WAVELET DOMAIN [14][15].The wavelet transform represents a sort of compromise between Time and Frequency domains simultaneously. It provides information about both time and frequency event of signals. It could of any type of event e.g. similarity or discontinuity in a signal. In wavelet analysis, we often speak of approximations and details.the approximations are high-scale, low frequency components whereas the details are the low-scale, high frequency components. Wavelet Transform has been implemented in many fields of Computer Vision (Machine Vision) for image texture classification. After then for final classification there are several types of conventional statistical approaches available for image texture classification such as Artificial Neural Networks, Support Vector Machines, Fuzzy Measure, Genetic Algorithms, Fuzzy Support Vector Machines and Genetic Algorithms with Neural Networks. We are using here support vector machines because it proves very efficient when the data have more input variables. for instance,the author [1] used a slantlet transform[9] with Error-back Propagation Neural Networks using normal brain images and image of Alzheimer s disease with exact six features there system achieve 100% classification accuracy. The features that are basically used in previous studies were either textual statistics that were obtained from the original image and/or statistics extracted from approximations sub-band. Both means had disadvantages. Computing statistics obtained from the original image only characterized the overall distribution of the image. In this paper We Proposes a classification techniques for Brain MRI that will use the features obtained from the horizontal (LH) and vertical details (HL) instead of utilizing approximations (LL) sub-band. This paper basically uses a hypothesis that the details feature obtained from images are more efficient at characterizing the feature that are changing in normal/pathological Brain MRI.

2 IJECSE,Volume1,Number 3 Prof. Bhupal Singh et al. II. METHODOLOGY The digital transformation of any image for computer processing requires digitization of images so we are using here 2-D discrete wavelet transform which decomposes an image into several sub-bands according to a repetitive process known as dyadic (scale and position) filter bank decomposition. This decomposition will generate wavelet coefficients.these coefficients will include various approximation & detail coefficients at various levels. This leads at each level to 4 different sub-bands HH, HL, LH and LL. The role of each decomposition in the analysis of an N N image can be explained as follows: Initial low pass filtering of the rows blurs the image values along each row followed by low pass filtering along the columns which result in a low pass approximation of the whole image. Low pass filtering of the rows followed by high pass filtering of the columns highlights the changes that occur between the rows - horizontal details Initial high pass filtering of the original rows of the image highlights the changes between elements in any given low. Subsequent low pass filtering of the columns blurs the changes that may occur between the rows thus providing the vertical details High pass filtering of the rows followed by high pass filtering of the columns only changes that are neither horizontal are emphasized. This sequence gives the diagonal details of the original image. The obtained coefficients are useful for texture categorization of image for instance, very small wavelets can be used for isolating finer details whereas the large wavelets can identify coarse details and there approximation may be obtained by using only few coefficients. The 2D-Discrete wavelet transform decompose an image into spectral/frequency component that describes the image textures. In case of two-dimensional image, after a DWT transform, the image is divided into four corners, upper left corner of the original image, lower left corner of the vertical details, upper right corner of the horizontal details, lower right corner of the component of the original image detail (high frequency). You can then continue to the low frequency components of the same upper left corner of the 2nd, 3rd inferior wavelet transform. LL2 LH2 HL2 HH2 HL1 LH1 HH1 Figure1.two level 2D-DWT of an image It is obtained by convolving the image with a bank of filters, after which selected image features are extracted from the obtained image representation for further processing. More formally, the 2D-DWT of a function f(x) is defined by: =,,, (1), =, (2) Where.(x) stands for the wavelet function and, are the 2D-DWT coefficients of f(x). A mother wavelet ψ(x) is used to generate the wavelet basis = 2 2 (3) where j and i are the translation and dilation parameters, respectively functions by using translation and dilation operations. We are considering here a second level decomposition because it has shown its efficiency in

3 Classification of Brain MRI in Wavelet Domain 881 previous studies. The daubechies -4 wavelet is our choice for mother wavelet because of it s fine resolution for smoothly changing signals. In addition, first order statistics are extracted from the coefficients of second level horizontal and vertical levels sub-bands which is our main concern here. Especially the feature set we are using here consists of six components Mean, Standard Deviation, Smoothness, Third- Moment, Uniformity and Entropy. The obtained result comprised of twelve component vectors, six from horizontal and six from vertical. The principal component analysis is now applied to reduce the number of dimension to 6. Principal component analysis is a variable reduction procedure. It is useful when you have obtained data on a number of variables (possibly a large number of variables), and believe that there is some redundancy in those variables. In this case, redundancy means that some of the variables are correlated with one another, possibly because they are measuring the same construct. Because of this redundancy, you believe that it should be possible to reduce the observed variables into a smaller number of principal components (artificial variables) that will account for most of the variance in the observed variables. Finally a support vector machine is deployed for the classification of the images. The SVM was chosen for the ability to avoid local minima and its scalability and generalization capabilities. We are considering here the twoclass SVM which is a very dominant approach for doing a single OR multiple class problems into multiple binary class problems into multiple binary classifications. Multiclass SVM classifier distinguish between (i) one of the label and the rest one (one versus all) or (ii) between every pair of class one-versus-one There are different way to establish a classifier performance graphical analysis of lift charts, and ROC Curves, and the use of measure function. We use here ROC-curve because graphically it describes the performance without the requirement of threshold measured by AUC (area under curve). The area under curve (AUC) [23] was used to assess classifier rule quality (AUC=0.5 for a non-discriminating rule, AUC=1 for a perfectly discriminating one) [24]. The AUC is chosen due to its independence of the decision threshold and its usefulness to measure the ranking quality of a classifier in two-class classification problems [24-28]. To test our model, we compared our approach to the standard one used for texture classification, where six statistics are computed from the LL2 sub-band to form the pattern vector [8][10].In particular the following performance measure are computed: correct classification rate, sensitivity, and specificity.!"!# $!!% &$ Sensitivity = ' $!!% &$ (4)!"!#()!% &$ Specificity = ' ()!% &$ (5) On the basis of such considerations, the algorithm uses a different color image multiplied by the weighting coefficients of different ways to solve the visual distortion, and by embedding the watermark, wavelet coefficients of many ways, enhance the robustness of the watermark.

4 IJECSE,Volume1,Number 3 Prof. Bhupal Singh et al. Figure 2.Flow Diagram of proposed Approach After that we select the ordered coefficient from 1 to N to get N coefficient. the formulae of watermark embedding are as follows. III. DATA AND RESULT A collection of 56 axial, T2-weighted, MR brain images of size was downloaded from the Harvard Medical School webpage ( See appendix for example images). The set included 5 images of normal brains and 51 of abnormal brains. The images were chosen based on visual sharpness and their class distribution was as listed in Table 1. The Mat-lab wavelet and bioinformatics toolboxes were used to implement the image analysis and recognition algorithms. We randomly selected 50% of the images of each class for training and the other half or testing (rounded to the nearest integer). As mentioned before, PCA was used to reduce the number of input variables to six. The principal components were identified by using the Kaiser criterion as a stopping rule (i.e. 70% threshold of cumulative explained variance) [29]. The selected statistics1were the mean, third moment and entropy for the LH sub-band and the standard deviation, smoothness, and third moment for the HL sub-band. The fact that the third moment is selected in both sub-bands appears to indicate that skewness is important to characterize the distribution of detail coefficients. Table -1 Number of images for each class Normal 5 Alzheimer s disease, visual agnosia 9 Glioma with tumor 13 Herpes encephalitis with a tour 7 Metastatic bronchogenic carcinoma 8 Multiple scelorsis with a tour 14 TOTAL 56 Support vector machines with a polynomial kernel2 were chosen to classify the images. We varied the order of the polynomial from 2 to 5 and found that the polynomial of order 2 gives the best classification rate. Table 4 summarizes the performance measures obtained when the inputs to the SVM are LH2 and HL2 features, and when they are LL2 features. For each approach, six one-against-all experiments were conducted (e.g. the SVM classifies a normal image against all other types). For instance, it is more appropriate to use one-against-all SVM rather than multi-class SVM since the number of samples of each pathology and normal images is small. Figure 4 provides the

5 Classification of Brain MRI in Wavelet Domain 883 average values of the computed performance measures. It shows that the proposed approach (LH+HL features) leads to higher performance indicators on average. In the few cases where they are lower, the individual results listed in Table 4 reveal that these indicators track their LL-based counterparts with less than 10% deviation, except for specificity in the case of Glioma with tumor. Figure 5 illustrates this in the case of the correct classification rate. Finally, the AUC results on discrimination power show that the proposed approach offers a much better discrimination power than the standard one for all classes (see Table 4), except Glioma with tumor. Table -2 Summary of obtained results Correct Rate Error Rate Sensitivity Specificity AUC LH+HL Features Alzheimer Glioma Herpes Meastatic Multiple Normal LL Features Alzheimer Glioma Herpes Meastatic Multiple Normal Correct classification rate AUC Sensitivity Specificity LH+HL LL Figure 3. Performance comparison of LH+HL versus LL based classification

6 IJECSE,Volume1,Number 3 Prof. Bhupal Singh et al LH+HL Features LL Features 0 Figure 4. Correct Classification rate as a function of image class IV.CONCLUSION The obtained results demonstrate the utility of extracting statistical features from the LH and HL wavelet sub-bands for the classification of brain MR images with support vector machines, as opposed to using statistical features extracted from the LL sub-band. The discrete wavelet transform was applied to both normal and six different types of pathological images, and statistical features from the LH and HL wavelet sub-bands were computed to form the input to the support vector machines after preprocessing. The performance of the classification system was studied using receiver operating characteristic analysis. The experimental results show that, in general, the proposed approach leads to higher correct classification rates than the standard approach. However, the standard approach is more suitable to detect Glioma images since its correct classification rate is perfect. In addition, the proposed approach gives a sensitivity of one (100%) when binary support vector machines are used to detect Glioma, Herpes, Multiple scelorsis, and normal images, while the standard approach leads to a sensitivity of one when the binary support vector machines are used to detect Alzheimer s disease, Glioma and Herpes. In sum, the information extracted from horizontal and vertical sub-bands does help characterizing textures of normal and pathological images. Moreover, it was found that the third moment largely contributes to the variance proportion of the statistical features extracted from the LH and HL sub-bands and, therefore, is an important feature to include in the input variables to the SVM classifier. This indicates that information about the asymmetry of detail images is important for the classification task. Finally, while our approach is more computationally intensive than when simply using the LL coefficients, it is still within the reach of a 16-bit microcontroller, DSP or FPGA, let alone a workstation. Thus, it can easily be included in the processing chain of a MRI machine as an embedded module. Future work will focus on validating the approach with a larger image database, on using wavelet packets for more detailed image decomposition, and on performing cross-validation to obtain more reliable values for the average and standard deviation of the obtained classification rates. In addition, a hybrid system that combines LH, HL and LL sub-band information will be investigated, potentially leading to more powerful discrimination capabilities for the resulting system. In this respect, the features of the HH sub-bands might contain all the desired information. Finally, a system on chip (SoC) implementation of the system will be realized. V. REFERENCE [1] M. Maitra and A. Chatterjee, "A Slantlet Transform Based Intelligent System for Magnetic Resonance Brain Image Classification," Biomedical Signal Processing and Control, vol. 1, p , [2] H. Selvaraj, S. Thamarai Selvi, D. Selvathi, and R. Ramkumar, "Support Vector Machine Based Automatic Classification of Human Brain using MR Image Features," International Journal of Computational Intelligence and Applications, vol. 6, no. 3, pp: , 2006.

7 Classification of Brain MRI in Wavelet Domain 885 [3] J. G. Webster, Medical Instrumentation: Application and Design, New York: John Wiley & Sons Inc, pp: , [4] E. M. Haacke, R.W. Brown, M.R. Thompson, and R. Venkatesan, Magnetic Resonance Imaging: Physical Principles and Sequence Design, Wiley-Liss, New York, [5] M. Sifuzzaman, M.R. Islam, and M.Z. Ali, "Application of Wavelet Transform and its Advantages Compared to Fourier Transform," Journal of Physical Sciences, vol. 13, pp , [6] E. Avci, "An Expert System Based on Wavelet Neural Network-Adaptive Norm Entropy for Scale Invariant Texture Classification," Expert Systems with Applications, vol. 32, issue.3, [7] D. Cvetkovic, E.D. Ubeyli, and I. Cosic, "Wavelet Transform Feature Extraction from Human PPG, ECG, and EEG Signal Responses to ELF PEMF Exposures: A pilot Study," Digital Signal Processing, vol. 18, pp , [8] S. Chaplot, L.M. Patnaik, and N.R. Jagannathan, "Classification of Magnetic Resonance Brain Images using Wavelets as Input to Support Vector Machine and Neural Network," Biomedical Signal Processing and Control, vol. 1,pp , [9] I.W. Selesnick, "The Slantlet Transform," IEEE Trans. Signal Process, vol. 47, no 5, pp: , [10] E.-S.A El-Dahshan, T. Hosny, and A.-B.M. Salem, "Hybrid Intelligent Techniques for MRI Brain Images Classification," Digital Signal Processing, vol 20,pp: , [11] C.E. Heil, and D.F. Walnut, "Continuous and Discrete Wavelet Transforms," SIAM Review, vol. 31, no. 4,pp , [12] A. Sengur, I. Turkoglu, and M.C. Ince, "Wavelet Packet Neural Networks for Texture Classification," Expert Systems with Applications, vol. 32, pp: , [13] Z.-Z. Wang, and J.-H. Yong, "Texture Analysis and Classification with Linear Regression Model Based on Wavelet Transform," IEEE Transactions On Image Processing, vol. 17, no. 8,pp: , [14] Ingrid Daubechies, Ten Lectures on Wavelets, SIAM, Philadelphia, PA, [15] Rafael C. Gonzalez, Richard Eugene Woods, Digital Image Processing. Third ed. Prentice Hall, [16] L. Kjaer, P. Ring, C. Thomsen, and O. Henriksen, "Texture Analysis in Quantitative MR Imaging," Acta Radiologica, vol. 36, no. 2, pp , [17] M.N. Mughal and W. Ikram, "Early Lung Cancer Detection by Classifying Chest CT Images: A Survey," Proceedings of the 8 th International Multitopic Conference, Proceedings of (INMIC 04), pp: 67 72, [18] H. Selvaraj, S.T Selvi, D. Selvathi, and L. Gewali, "Brain MRI Slices Classification Using Least Squares Support Vector Machine," International Journal of Intelligent Computing in Medical Sciences and Image Processing, (ICMED 07), vol. 1, no. 1, issue. 1, pp: 21-33, [19] G. Schaefer, M. Závišek and T. Nakashima, "Thermography Based Breast Cancer Analysis using Statistical Features and Fuzzy Classification," Pattern Recognition, vol. 47, pp: , [20] A. Faro, D. Giordano, C. Spampinato, and M. Pennisi, "Statistical Texture Analysis of MRI Images to Classify Patients Affected by Multiple Sclerosis," Proceedings of the XII Mediterranean Conference on Medical and Biological Engineering and Computing, vol. 29, Part 2, pp: ,2010. [21] V.N. Vapnik, The Nature of Statistical Learning Theory, Springer-Verlag, [22] J. Ma, M. N. Nguyen, and J.C. Rajapakse, "Gene Classification Using Codon Usage and Support Vector Machines," IEEE/ACM Transactions on Computational Biology and Bioinformatics, vol. 6, no. 1, [23] A.E. Bradley, "The Use of the Area Under the ROC Curve in the Evaluation of Machine Learning Algorithms," Pattern Recognition, vol. 30, no. 7, pp: , [24] [24] C. Marrocco, R.P.W. Duin, and F. Tortorella, "Maximizing The Area under The ROC Curve by Pairwise Feature Combination," Pattern Recognition, vol. 41,pp: , [25] [25] J.A. Hanley and B.J. McNeil, "The Meaning and Use of the Area under A Receiver Operating Characteristic (ROC) Curve," Radiology, vol. 143,pp:29 36, [26] [26] C. Cortes, M. Mohri, "AUC Optimization vs. Error Rate Minimization," Advances in Neural Information Processing Systems (NIPS 2003). [27] [27] J. Huang, C.X. Ling, "Using AUC and Accuracy in Evaluating Learning Algorithms," IEEE Trans. Knowledge Data Eng, vol. 17, pp: , [28] [28] D.J. Hand and R.J. Till, "A Simple Generalization of The Area under The ROC Curve to Multiple Class Classification Problems," Machine Learning, vol. 45, pp: , [29] [29] N. Brauner and M. Shacham, "Considering Precision of Data in Reduction of Dimensionality and PCA," Computers and Chemical Engineering, vol. 24, pp: , 2000.