Image Pre-processing and Feature Extraction Techniques for Magnetic Resonance Brain Image Analysis

Size: px

Start display at page:

Download "Image Pre-processing and Feature Extraction Techniques for Magnetic Resonance Brain Image Analysis"

Olivia Lang
6 years ago
Views:

1 Image Pre-processing and Feature Extraction Techniques or Magnetic Resonance Brain Image Analysis D. Jude Hemanth and J. Anitha Department o ECE, Karunya University, Coimbatore, India {jude_hemanth,rajivee1}@redimail.com Abstract. Image pre-processing and eature extraction techniques are mandatory or any image based applications. The accuracy and convergence rate o such techniques must be signiicantly high in order to ensure the success o the subsequent steps. But, most o the time, the signiicance o these techniques remain unnoticed which results in inerior results. In this work, the importance o such approaches is highlighted in the context o Magnetic Resonance (MR) brain image classiication and segmentation. In this work, suitable pre-processing techniques are developed to remove the skull portion surrounding the brain tissues. Also, texture based eature extraction techniques are also illustrated in this paper. The experimental results are analyzed in terms o segmentation eiciency or pre-processing and distance measure or eature extraction techniques. The convergence rate o these approaches is also discussed in this work. Experimental results show promising results or the proposed approaches. Keywords: Pre-processing, MR brain images, eature extraction, segmentation eiciency and convergence rate. 1 Introduction Generally, real-time images collected rom scan centre and simulated images collected rom publicly available database are used or image classiication and segmentation. These are raw images which are unsuitable or analysis due to the various types o noises present in the images. Hence, suitable pre-processing methodologies must be used to enhance the quality o the images. Literature survey reveals the availability o several pre-processing and eature extraction techniques or MR brain image analysis. The raw MR images normally consist o many artiacts such as intensity inhomogenities, extra cranial tissues, etc. which reduces the overall accuracy. Several researches are reported in the literature to minimize the eects o artiacts in the MR images. An analysis on iltering techniques such as Gabor & QMF ilters or noise reduction is perormed by Nicu et al (2000). Fuzzy connectedness based intensity non uniormity correction has been implemented by Yongxin et al (2006). A sequential approach with uzzy connectedness, atlas registration and bias ield correction is used in this approach. The conclusions revealed that the proposed technique can be used only i the intensity variations between the images are o a limited range. Marianne et al (2006) have T.-h. Kim et al. (Eds.): FGCN/DCA 2012, CCIS 350, pp , Springer-Verlag Berlin Heidelberg 2012

2 350 D. Jude Hemanth and J. Anitha minimized the eects o inter-slice intensity variation with the weighted least square estimation method. The selection o weights or the least square method is the major disadvantage o this approach. Bo et al (2008) have proposed the noise removal technique using wavelets and curvelets. Hybrid approaches involving Variance Stabilizing Transorm (VST) are also used in this work. But this technique is applicable or images with Poisson noise. Tracking algorithm based de-noising technique is perormed by Jaya et al (2009). Since the seed point or tracking is random in nature, this technique is not much eicient. A contrast agent accumulation model based contrast enhancement is implemented by Marcel et al (2009). This improves only the contrast o the image and the unwanted tissues are not eliminated. Rajeev et al (2009) have used the wiener iltering technique or noise removal in MR brain images. Apart rom noise removal, several other preprocessing steps are also reported in the literature. This includes image ormat conversion, image type conversion etc. Rajeev et al (2009) also have used the combination o three modalities o MR images or urther processing. All the above mentioned techniques remove only speciic artiacts which is not suicient or high classiication accuracy and segmentation eiciency. The next step in the automated diagnosis process is eature extraction. Feature extraction is the technique o extracting speciic eatures rom the pre-processed images o dierent abnormal categories in such a way that the within - class similarity is maximized and between - class similarity is minimized. Earlier research works report many eature extraction techniques employed or medical image processing. Arivazhagan et al (2003) have used 2D wavelet transorm based textural eatures or classiication. In this report, initially basic statistical eatures are used and then cooccurrence based textural eatures are used to improve the accuracy. But the eects o usage o dierent wavelets are not dealt in the report. A comparison o 2D wavelet transorm based textural eatures and 3D wavelet transorm based textural eatures is perormed by Kourosh et al (2004). This work concluded that the combination o 2D and 3D wavelet based textural eatures yield better results than the 2D wavelet eatures. Hiremath et al (2006) have presented a eature extraction technique using the complimentary wavelet transormed image. The report claimed that the eatures extracted rom all the our sub-bands are more eicient than the eatures rom the only the approximation sub-band. All these techniques used the basic Discrete Wavelet Transorm (DWT) which does not yield superior results. An improved version based on wavelet packet decomposition is implemented by Hiremath et al (2006). The results revealed that the packet decomposition technique is more eicient than the DWT technique. Apart rom extracting the eatures rom the whole image, eatures are also extracted rom local regions which are used or image segmentation applications. One such work is reported by Ryszard (2007). Pantelis et al (2007) have described a novel eature set which comprises the eatures such as short run emphasis, run length non-uniormity, etc. which are based on run length matrices. The drawback o this work is the low classiication accuracy which shows that these eatures do not guarantee superior results. Ke et al (2008) have explored the merits o wavelet eatures or image classiication. The technique o dimensionality reduction based on sub band grouping and selection have also been implemented in this work. A comparative analysis with the conventional algorithms is presented in this report.

Image Pre-processing and Feature Extraction Techniques or MR Brain Image Analysis 351 In this work, emphasis is given to remove the extra-cranial (skull) tissues which oten interere with the brain

Next, an extensive eature set is extracted rom these images and supplied as input to the automated system.

These eature extraction techniques are perormed in a dierent manner or the classiication and segmentation techniques. 2 Brain Image Database The image database used in this work is collected rom M/s.

3 Image Pre-processing and Feature Extraction Techniques or MR Brain Image Analysis 351 In this work, emphasis is given to remove the extra-cranial (skull) tissues which oten interere with the brain tissues leading to the perormance reduction o the system. Suitable morphology based techniques are used to remove the skull tissues. Next, an extensive eature set is extracted rom these images and supplied as input to the automated system. The characteristics o the input images are relected by these eatures which are necessary to enhance the eiciency o the system. These eature extraction techniques are perormed in a dierent manner or the classiication and segmentation techniques. 2 Brain Image Database The image database used in this work is collected rom M/s. Devaki scan centre in Madurai, Tamilnadu, India. The total number o images is 540 with representations rom our categories such as Meningioma, Glioma, Metastase and Astrocytoma. The images are urther categorized into training dataset and testing dataset. The images are gray level images with intensity value ranges rom (0 to 255). These images are used or both classiication and segmentation applications. Some samples o the MRI database have been displayed in Figure 1. (a) (b) (c) (d) Fig. 1. Sample data set: (a) Metastase (b) Glioma (c) Astrocytoma (d) Meningioma The ground truth tumor image is also available or all the 540 images. These ground truth are used as reerence or the tumor segmentation applications. 3 Image Pre-processing The raw images collected rom the scan centre and the websites are not suitable or direct processing due to the various noises present in these images. In this work, emphasis is laid on the removal o extra-cranial (skull) tissues which oten interere with the brain tissues. The presence o these skull tissues has reduced the perormance o the automated system. Hence, suitable morphology based techniques are devised in this work to eliminate these extra-cranial tissues. 3.1 Framework o the Proposed Technique The various steps ollowed in the extra-cranial tissue extraction are shown in this section. Dierent sequential steps with dierent parameters are implemented in this work to eliminate the skull tissues which are usually surrounding the brain tissues.

4 352 D. Jude Hemanth and J. Anitha The procedual low o this algorithm is detailed in the ollowing three steps. Step 1: Generation o mask The irst step ater reading the image is the generation o a mask. Two basic morphological operations namely erosion and illing are used in this work to generate the mask. a) Erosion: The main unction o this operation is to remove the pixels on object boundaries. The rule given below deines the operation o the erosion operation. I every pixel in the input pixel's neighborhood is on, the output pixel is on. Otherwise, the output pixel is o. A speciied neighborhood is used in this operation. The neighborhood or an erosion operation can be o arbitrary shape and size. The neighborhood is represented by a structuring element, which is a matrix consisting o only 0's and 1's. The shape o the structuring element used in this work is ball and the size o the structuring element is 5-by-5. b) Gray to binary image conversion: A suitable threshold (40-50) is selected and the pixels above the threshold value are made equal to 255 and the pixels below the threshold value are made equal to 0. These thresholds are selected based on trial and error method. c) Connected component analysis: It is a process that "ills" a region o interest by interpolating the pixel values rom the borders o the region. This process can be used to make objects in an image seem to disappear as they are replaced with values that blend in with the background area. This unction is useul or removal o extraneous details or artiacts. The resultant image yields the mask or the corresponding input image. Step 2: Logical conversion o the mask The indices o the binary mask image are changed rom (0-255) to (0-1). Step 3: Masking The original input image is masked with the binary mask. The masking is achieved by perorming the multiplying operation between the original image and the mask. The resultant image is ree rom the extra-cranial tissues which is evident rom the output images. 4 Feature Extraction The purpose o eature extraction is to reduce the original data set by measuring certain properties or eatures that distinguish one input pattern rom another pattern. The extracted eature should provide the characteristics o the input type to the classiier by considering the description o the relevant properties o the image into a eature space. Eight dierent textural eatures are used in this work or image analysis. But, these eatures are estimated in a dierent manner or the classiication and segmentation applications. Since image classiication is perormed between images, the eatures are estimated rom the whole image. On the other hand, image segmentation is perormed within the image and hence the eatures are estimated or each pixel. A neighborhood window o size 3 3 is chosen with the center pixel being the pixel o interest. The procedure is repeated or all the pixels to estimate the eatures.

5 Image Pre-processing and Feature Extraction Techniques or MR Brain Image Analysis 353 In this work, eight textural eatures such as angular second moment, contrast, correlation, variance, entropy, Inverse Dierence Moment, skewness and kurtosis are used or image classiication and segmentation. The textural eatures are extracted rom the pre-processed image which guarantees high success rate or the subsequent steps. The eatures used in this work are estimated rom the ollowing ormulae. Angular Second Moment (ASM): The summation o the squares o gray levels o the image is known as angular second moment. It is also named as energy (or) uniormity. The energy is usually high when the intensity values are unequal. = { p( i j) } 2 (1) 1, i j where p(i,j) represents the input image. Contrast: The local contrast o an image is measured by this eature. It is expected to be low i the intensity values o the pixel are similar. N g 1 N g N g 2 2 = n p( i, j) (2) n= 0 i= 1 j= 1 where N g is the number o gray levels in the original image. Correlation: The linear dependency o grey levels on the neighboring pixels is represented by the correlation eature. The statistical relationship between the two variables is denoted by this eature. = i j ( ij) p( i, j) μ x σ σ x y μ 3 (3) where μ x, μ y, σ x, σ y are the means and standard deviations o the input image in the row-wise and column-wise order. Variance: Variance is a measure o the dispersion o the values around the mean. The variance can also be deined as the measure o how ar the gray values are spread out in the input image. 2 ( i ) p( i, ). 4 j i j y = μ (4) where µ is the mean o the whole image. Inverse Dierence Moment (IDM): The details o the smoothness o the image are deined by the Inverse dierence moment. The IDM is expected to be high i the gray levels o the pixel are similar.

354 D. Jude Hemanth and J. Anitha 1+ 1 ( i j) p( i, ).

The entropy is expected to be high i the gray levels are distributed randomly throughout the image. = p( i, j) log( p( i, )).

The value is usually zero i the image is exactly symmetrical.

distribution. 1 4 8 = 4 σ i j ( p( i, j) μ) 3 These eatures are commonly used or brain image analysis.

5 Experimental Results and Discussions The algorithms are implemented using MATLAB sotware with the processor speed o 1.

6 354 D. Jude Hemanth and J. Anitha 1+ 1 ( i j) p( i, ). = (5) 5 j 2 i j Entropy: The randomness o a gray level distribution is denoted by entropy. The entropy is expected to be high i the gray levels are distributed randomly throughout the image. = p( i, j) log( p( i, )). (6) 6 j i j Skewness: The measure o symmetry is given by this textural eature. This value can be either negative or positive. The value is usually zero i the image is exactly symmetrical. 1 7 =, 3 σ i j ( p( i j) μ) 3 Kurtosis: Kurtosis is a measure o whether the data are peaked or lat relative to the normal distribution = 4 σ i j ( p( i, j) μ) 3 These eatures are commonly used or brain image analysis. These eatures are speciically ound to be superior or image classiication and segmentation applications. 5 Experimental Results and Discussions The algorithms are implemented using MATLAB sotware with the processor speed o 1.66GHz and 1GB RAM. Sample results o the various stages o skull removal technique or the real-time dataset are shown in Figure 2. (7) (8) (a) (b) (c) (d) (e) (a) (b) (c) (d) (e) Fig. 2. Sample pre-processed images: (a) Input images, (b) Eroded images, (c) Binary images, (d) Images ater connected component analysis, (e) Masked images

7 Image Pre-processing and Feature Extraction Techniques or MR Brain Image Analysis 355 An analysis o the above results has clearly revealed the success rate o the preprocessing technique in removing the skull tissues. The dierence is visible in Figure 2 (e). A quantitative analysis on these results is shown in Table 1. Image type Real-time images Table 1. Quantitative analysis o image pre-processing technique No. o ground No. o skull pixels Segmentation truth skull pixels removed by the technique eiciency (%) The average values are shown in the results since there is a slight deviation in the number o skull pixels removed in each input images. The robustness o this technique is also revealed since this method is applicable or the simulated images and the real-time images. The convergence rate o this method is only 0.47 CPU seconds which is very much signiicant or real-time applications. The textural eatures mentioned above are also calculated using the MATLAB sotware. These eatures are calculated separately or the image classiication and segmentation process. The size o the eature set or the image classiication is 1 8 or an image o any input size. These eatures are estimated or all the our categories and are tabulated in Table 2. Table 2. Feature extraction results or image classiication Features Metastase Meningioma Glioma Astrocytoma ASM e Contrast Correlation e e e e+005 Variance e e e e+011 IDM e e e e+004 Skewness e e e e+010 Entropy e e e e+007 Kurtosis e e e e+012 The eatures extracted using this method is highly eicient which is validated rom the experimental results. The distance measure o these eature values within the class is very minimum or any individual eature and maximum between the classes. Thus, this work has suggested suitable methodologies or image pre-processing and eature extraction or MR brain image analysis. 6 Conclusion The signiicance o pre-processing and eature extraction methodologies is veriied in this work. The accuracy level is signiicantly high or the proposed techniques. The time requirement is also signiicantly low which suggests the easible nature o these techniques. Thus, these approaches can be used or practical medical applications where the accuracy and convergence rate are extremely important.

8 356 D. Jude Hemanth and J. Anitha Acknowledgment. The authors wish to thank M/s. Devaki Scan Centre or their help regarding database collection and validation. Reerences 1. Nicu, S., Michael, S.L.: Wavelet based texture classiication. In: 15th International Conerence on Pattern Recognition, vol. 3, pp (2000) 2. Yongxin, Z., Jing, B.: Atlas based uzzy connectedness segmentation and intensity nonuniormity correction applied to brain MRI. IEEE Transactions on Biomedical Engineering 54(1), (2006) 3. Marianne, M., Russell, G., Jorg, S., Albert, M., Mark, S.: Learning a classiication based glioma growth model using MRI data. Journal o Computers 1(7), (2006) 4. Bo, Z., Jalal, M.F., Jean-Luc, S.: Wavelets, ridgelets and curvelets or Poisson noise removal. IEEE Transactions on Image Processing 17(7), (2008) 5. Jaya, J., Thanushkodi, K.: Implementation o classiication system or medical images. European Journal o Scientiic Research 53(4), (2011) 6. Marcel, P., Elizabeth, B., Guido, G.: Simulation o brain tumors in MR images or evaluation o segmentation eicacy. Medical Image Analysis 13, (2009) 7. Rajeev, R., Sanjay, S., Sharma, S.K.: Brain tumor detection based on multi-parameter MR image analysis. Graphics, Vision and Image Processing Journal 9(3), 9 17 (2009) 8. Rajeev, R., Sanjay, S., Sharma, S.K.: Multiparamter segmentation and quantization o brain tumor rom MR images. Indian Journal o Science and Technology 2(2), (2009) 9. Arivazhagan, S., Ganesan, L.: Texture classiication using wavelet transorm. Pattern Recognition Letters 24, (2003) 10. Kourosh, J.K., Hamid, S., Kost, E., Suresh, P.: Comparison o 2D and 3D wavelet eatures or TLE lateralization. In: Proceedings o SPIE, vol. 5369, pp (2004) 11. Hiremath, P.S., Shivashankar, S.: Texture classiication using wavelet packet decomposition. Graphics, Vision and Image Processing Journal 6(2), (2006) 12. Ryszard, S.C.: Image eature extraction techniques and their applications or CBIR and biometrics system. International Journal o Biology and Biomedical Engineering 1(1), 6 16 (2007) 13. Georgiadis, P., Cavouras, D., Kalatzis, I., Daskalakis, A., Kagadis, G.C., Siaki, K., Malamas, M., Nikioridis, G.C., Solomou, E.: Non-linear Least Squares Features Transormation or Improving the Perormance o Probabilistic Neural Networks in Classiying Human Brain Tumors on MRI. In: Gervasi, O., Gavrilova, M.L. (eds.) ICCSA 2007, Part III. LNCS, vol. 4707, pp Springer, Heidelberg (2007) 14. Ke, H., Selin, A.: Wavelet eature selection or image classiication. IEEE Transactions on Image Processing 17(9), (2008)

CHAPTER 6 MODIFIED FUZZY TECHNIQUES BASED IMAGE SEGMENTATION

CHAPTER 6 MODIFIED FUZZY TECHNIQUES BASED IMAGE SEGMENTATION 6.1 INTRODUCTION Fuzzy logic based computational techniques are becoming increasingly important in the medical image analysis arena. The significant