MEDICAL IMAGE SEGMENTATION USING CELLULAR NEURAL NETWORK Oloruntoyin Sefiu Taiwo Department of Computer Science & Engineering Faculty of Engineering & Technology Ladoke Akintola University of Technology, Nigeria E-mail: email@example.comABSTRACT: Cellular neural network is an indispensable in the segmentation of an image. Over the last decade,numerous methods for the segmentation of images has been developed within different theoretical frameworks. Ourmain aim in this paper is to make use of the method of cellular neural network in image segmentation. We wouldclassify the images and show each step of how the images is been segmented. Furthermore, methods like theCompression based method, histogram based method, region growing method and spilt and merge method havebeen used for image segmentation in this project, histogram based method is the most efficient method of all. Also,in this work, cellular neural network algorithm and flowchart are also used to show analyze how the imagesegmentation are performed.KEYWORDS: Medical Image, Segmentation, Cellular Neural Network, Artificial Neural Network. I. INTRODUCTION Among these, image segmentation is more important as it is a critical step for high-level processing such as The analysis of image and processing of image object recognition. Multi-Layer Perceptron (MLP),tends to focus on 3-d images, how to transform one Radial Basis Function (RBF), Hopfield, Cellular, andimage to another. Like for instance the pixel-wise Pulse-Coupled neural networks have been used foroperation such as operations such as the contrast image segmentation. These networks can be categorizedenhancement, local operations or the geometrical into feed-forward (associative) and feedback (auto-transformations like the image rotation. associative) networks. MLP, Self-Organized Map (SOM), and RBF neural networks belong to the feed- Segmentation of tissues and structures from medical forward networks while Hopfield, Cellular, and Pulse-images is the first step in image analysis applications Coupled neural networks belong to the feedbackdeveloped for medical diagnosis. Development of networks.treatment plans and evaluation of disease progression areother applications. These applications stem from the fact Segmentation is refers to as the process ofthat diseases affect specific tissues or structures, lead to partitioning a digital image into many regions also, heloss, atrophy (volume loss), and abnormalities. goal of segmentation is to simplify and change heConsequently, an accurate, reliable, and automatic representation of an image into something that issegmentation of these tissues and structures can improve meaningful and definitely easier to analyze and process.diagnosis and treatment of diseases. Manual So classically, segmentation of image is thereby definedsegmentation, although prone to rater drift and bias, is as the partitioning of an image into non-overlapping andusually accurate but is impractical for large datasets constituent regions which are homogeneous with respectbecause it is tedious and time consuming. Automatic to some features such as intensity and texture.segmentation methods can be useful for clinical Therefore, the segmentation of image plays a vital role inapplications if they have: the medical image through the facilitating of the 1) Ability to segment like an expert; delineation of regions interest. It is leant that the 2) Excellent performance for diverse datasets; and clinicians gained knowledge from a good segmentation3) Reasonable processing speed. Artificial Neural as they provide important information for surgicalNetworks (ANNs) have been developed for a wide range planning, 3-D visualization and early disease detection,of applications such as function approximation, feature MR is generally more sensitive in detecting brainextraction, optimization, and classification. In particular, abnormalities during the early stages of disease, and isthey have been developed for image enhancement, excellent in early detection of cases of cerebralsegmentation, registration, feature extraction, and object infarction, brain tumors, or infections. MR is particularlyrecognition. useful in detecting white matter disease, such as multiple
sclerosis, progressive multifocal leukoencephalopathy,leukodystrophy, and post-infectious encephalitis. The II. RELATED WORKfollowing artifacts are present in MR imaging: MEDICAL IMAGE SEGMENTATION Partial Volume RF Noise TECHNIQUES Intensity inhomogeneity Gradient Motion Due to the fact that there is the increase in the number Wrap Around and size of medical image, there is necessity in the use of Gibbs Ringing computer in facilitating their processing and analysis. Susceptibility Most especially, computer algorithm for the delineation of anatomical structures and other regions of interest are There are different in the method of performing a known to be the key component in assisting andgood segmentation and this depends on the specific automating specific radiological tasks. The imageapplication, Medical imaging is performed in various segmentation algorithm really play a vital role in themodalities, such as MRI, CT, ultrasound, positron biomedical imaging applications like the location ofemission tomography (PET), etc. In the present review, pathology, study of anatomical structure, treatmentwe are focusing primarily on the segmentation of MR planning, quantitative of tissue volumes etc.and CT images only image modality and other notedfactors. Like for instance the method of segmentation of With the development of complex medicalkidney is definitely different from that of the Liver, each imaging modalities which is known to be capable ofimage modality has its own idiosyncrasies with which to producing a large quantity three dimensional images,contend. It is therefore concluded based on the research segmentation is known to be a challenging field of imageso far there is currently no single segmentation methods analysis. As it has been said earlier, the process ofthat yield an acceptable result for every medical image. image segmentation is very important for most doctorsHowever the methods that are hereby specialization to and patients in the regard of the provision of importantparticular applications can often achieved better information for some medical acts of practices like inperformance by taking into consideration via knowledge. surgical planning, 3-D visualization and early timeThere may arise difficulties in the selection of disease detection, image segmentation so this byappropriate approach to a segmentation problem. facilitating delineation of regions of interest.Techniques that are used in image segmentation are asfollows: METHODS OF IMAGE SEGMENTATIONi. Artificial Neural Networkii. Classifiersiii. Histogram-based methodiv. Region Growing Approaches There are various methods by which image segmentationv. Deformation Models can be carried out, these methods has its individual wayvi. Markov Random Field Model of operations, level of usefulness, implementations andvii. Threshold Approaches limitations. These methods will be discussed in the part.Each of these techniques has its own ways operation inthe image segmentation and each also has its own merit COMPRESSION-BASED METHODSand demerit. Compression based methods postulate that the optimal segmentation is the one that minimizes, over all possible segmentations, the coding length of the data.The connection between these two concepts is that segmentation tries to find patterns in an image and any regularity in the image can be used to compress it. The method describes each segment by its texture and boundary shape. Each of these components is modeled by a probability distribution function and its coding length is computed as follows:
1. The boundary encoding leverages the fact that image classification distance metric and integrated regions in natural images tend to have a smooth region matching are familiar. contour. This prior is used by Huffman to encode the difference chain code of the Histogram-based approaches can also be contours in an image. Thus, the smoother a quickly adapted to occur over multiple frames, while boundary is, the shorter coding length it attains. maintaining their single pass efficiency. The histogram 2. Texture is encoded by glossy compression in a can be done in multiple fashions when multiple frames way similar to minimum description length are considered. The same approach that is taken with one (MDL) principle, but here the length of the data frame can be applied to multiple, and after the results are given the model is approximated by the number merged, peaks and valleys that were previously difficult of samples times the entropy of the model. The to identify are more likely to be distinguishable. The texture in each region is modeled by a histogram can also be applied on a per pixel basis where multivariate normal distribution whose entropy the information results are used to determine the most has closed form expression. An interesting frequent color for the pixel location. This approach property of this model is that the estimated segments based on active objects and a static entropy bounds the true entropy of the data environment, resulting in a different type of from above. This is because among all segmentation useful in Video tracking. distributions with a given mean and covariance, normal distribution has the largest entropy. Thus, the true coding length cannot be more REGION-GROWING METHODS than what the algorithm tries to minimize. For any given segmentation of an image, this The first region-growing method was thescheme yields the number of bits required to encode that seeded region growing method. This method takes a setimage based on the given segmentation. Thus, among all of seeds as input along with the image. The seeds markpossible segmentations of an image, the goal is to find each of the objects to be segmented. The regions arethe segmentation which produces the shortest coding iteratively grown by comparing all unallocatedlength. This can be achieved by a simple agglomerative neighboring pixels to the regions. The differenceclustering method. The distortion in the lossy between a pixels intensity value and the regions mean,compression determines the coarseness of the , is used as a measure of similarity. The pixel with thesegmentation and its optimal value may differ for each smallest difference measured this way is allocated to theimage. This parameter can be estimated heuristically respective region. This process continues until all pixelsfrom the contrast of textures in an image. For example, are allocated to a region.when the textures in an image are similar, such as incamouflage images, stronger sensitivity and thus lower Seeded region growing requires seeds asquantization is required. additional input. The segmentation results are dependent on the choice of seeds. Noise in the image can cause the HISTOGRAM-BASED METHOD seeds to be poorly placed. Unseeded region growing is a modified algorithm that doesnt require explicit seeds. It starts off with a single region – the pixel chosen here Histogram- based methods are very efficient does not significantly influence final segmentation. Atwhen compared to other image segmentation methods each iteration it considers the neighboring pixels in thebecause they typically require only one pass through the same way as seeded region growing. It differs frompixels. In this technique, a histogram is computed from seeded region growing in that if the minimum is lessall of the pixels in the image, and the peaks and valleys than a predefined threshold then it is added to thein the histogram are used to locate the clusters in theimage. Color or intensity can be used as the measure. respective region . If not, then the pixel is considered significantly different from all current regions and a A refinement of this technique is to recursively new region is created with this pixel.apply the histogram-seeking method to clusters in theimage in order to divide them into smaller clusters. This One variant of this technique, proposed byis repeated with smaller and smaller clusters until no Haralick and Shapiro (1985), is based on pixelmore clusters are formed. intensities. The mean and scatter of the region and the intensity of the candidate pixel is used to compute a test One disadvantage of the histogram-seeking statistic. If the test statistic is sufficiently small, the pixelmethod is that it may be difficult to identify significant is added to the region, and the region’s mean and scatterpeaks and valleys in the image. In this technique of
are recomputed. Otherwise, the pixel is rejected, and is structures has led to very efficient implementations ofused to form a new region. this method. A special region-growing method is called - ARTIFICIAL NEURAL NETWORKSconnected segmentation (see also lambda-connectedness). It is based on pixel intensities andneighborhood-linking paths. A degree of connectivity Artificial neural networks (ANNs) are massively parallel(connectedness) will be calculated based on a path that is networks of processing elements or nodes that simulateformed by pixels. For a certain value of , two pixels biological learning. Each mode in an ANN is capable ofare called - connected if there is a path linking those performing elementary computation. Through thetwo pixels and the connectedness of this path is at least adaptation of weights assigned to the connections . -connectedness is an equivalence relation. between nodes, learning can be achieved. ANNs represent a paradigm for marching learning and can be used in a variety of ways for image segmentation. SPLIT-AND-MERGE METHODS ANNs as a classifier is most widely used where the weight are determine using the training data and later used to segment new data. Though ANNs are inherently Split-and-merge segmentation is based on a parallel, their processing is usually simulated on aquad tree partition of an image. It is sometimes called standard serial computer, thus reducing its potentialquad tree segmentation. This method starts at the root of computational advantage. The ANNS can as well bethe tree that represents the whole image. If it is found used is an unsupervised fashion as a clustering methodnon-uniform (not homogeneous), then it is split into four and as a deformable models.son-squares (the splitting process), and so on so forth.Conversely, if four son-squares are homogeneous, theycan be merged as several connected components (the RESEARCH METHODOLOGYmerging process). The node in the tree is a segmentednode. This process continues recursively until no further CELLULAR NEURAL METHODsplits or merges are possible. When a special datastructure is involved in the implementation of the ALGORITHM AND FLOWCHARTalgorithm of the method, its time complexity can reach , an optimal algorithm of the method. The algorithm comprises of the following steps: Step 1: start Step 2: select image LEVEL SET METHODS Step 3: read image matrix Step 4: perform laplacian filter on the image to remove noise The level set method was initially proposed to Step 5: smoothen image using Gaussian filtertrack moving interfaces by Osher and Sethian in 1988 Step 6: find the global image threshold using otu’sand has spread across various imaging domains in the methodlate nineties. It can be used to efficiently address the Step 7: convert image to grayscale i.e. 2 dimensionalproblem of curve/surface/etc. propagation in an implicit Step 8: convert image to binary (i.e. black and white)manner. The central idea is to represent the evolving Step 9: detect the edges or regions above threshold pointcontour using a signed function, where its zero level Step 10: find the complement of the detected regioncorresponds to the actual contour. Then, according to the Step 11: stopmotion equation of the contour, one can easily derive asimilar flow for the implicit surface that when applied tothe zero-level will reflect the propagation of the contour.The level set method encodes numerous advantages: it isimplicit, parameter free, provides a direct way toestimate the geometric properties of the evolvingstructure, can change the topology and is intrinsic.Furthermore, they can be used to define an optimizationframework as proposed by Zhao, Merriman and Osher in1996. Therefore, one can conclude that it is a veryconvenient framework to address numerous applicationsof computer vision and medical imageanalysis.Furthermore, research into various level set data
FIG. 1 : This is the original image of the part of the brain Fig 3: This shows the complement of the segmentedbefore any segmentation take place, the part in the image. This image demonstrate what happen during aftermiddle of the brain is to be segmented. the second and last segmentation, when the image is further segmented, the part of the brain that has been segmented which is the middle part becomes white and the rest part appear to be black. METHOD TIME NUMBER OF USED TAKEN ITERATION CELLULAR 4.68 3 NEURAL NETWORK TABLE 1 :THE ANALYSIS OF IMAGE 1 The table tells more about the analysis of the third image. i.e how image 3 is formed in terms of the method being used, the time take for segmentation to take place and the number of iteration involved.FIG 2 : Binary masking of the segmented image, thisshows the first step of segmentation where the focus partof the brain which is the middle part which has beensegmented becomes black while the other part appears tobe white during the first segmentation. This is just thewhat the above diagram in Fig2 is demonstrating.
Fig 4: This shows the original image beforesegmentation take place. Fig. 4 The Complement of the segmented image, this shows what happen to the image after its undergoes the final segmentation where the affected part finally becomes white while the rest part which are unaffected appears black. TABLE 4; THE ANALYSIS OF IMAGE 2 The table below tells more about the analysis of the third image i.e how image 3 is formed in terms of the method used for segmentation, the time take for segmentation to take place and the number of iteration involved. NUMBER METHOD TIME OFFig 4; Binary masking of the segmented image, this TAKEN ITERATIONdemonstrate how the image looks like after its undergoesfirst segmentation where the part that has been CELLULARsegmented appear to be black differentiating it fromother part of the brain that is not affected. NEURAL 4.69 3 NETWORK
the algorithm after the extensive study of the methods that can be of assistance. This will aid more accuracy will. REFERENCES Amartur, S. C.; Piraino, D. &Takefuji, Y. (1992). Optimization neural networks for the segmentation of magnetic resonance images. IEEE Transactions on Medical Imaging, 11(2): 215-220. Angelini, E. D.; Jin, Y. &Laine, A. (2005). State-of-the-art of levelset methods in segmentation and registration of medical imaging modalities. The Handbook of Medical Image Analysis 3. Amadieu, O., Debreuuve, E., Barlaud, M., and Aubert, G. 1999. InProcessing ICIP, Japan, pp. 188-192Fig 4: the graphics user interface for the implementation Ambrosio, L. 1989. A compactness theorem for a special class of functions of bounded variation. Boll Un. Mat. It., 3(B) : 857cellular neural network method – 881 HOW ITS IMPLEMENTED WORKS FOR Ambrosio, L., and Tortorelli, V.M. 1990. Approximation of the functional depending on jumps by elliptic functional THE SEGMENTATION OF IMAGE. via_convergence. Comm. Pure Appl. Math., 43:999-1036 Ambrossio, L., and Tortorelli V.M. 1992. On the approximation of free discontuinity problems Bollettino U.M.I., 7(6-B): 105-i. First browse for the image for segmentation 123ii. Then press detect Alzheimer region to perform Aubert, G., and Kornprobst, P. 2001. Mathematical Problems insegmentation image Processing.Partial Differential Equation and the Calculus of Variations. Applied Mathematical Sciences,iii. After then we shall wait for few seconds for the vol. 147,Springer: New York.image to undergo segmentation Bishop, C. M. (2006). Pattern Recognition and Machine Learning,iv. The segmented image will then appear Springerv. Then press exit to close the user interface. Chang, C. Y. & Chung, P. C. (2001).Medical image segmentation using a contextualconstraint-based Hopfield neural cube.Image and Vision Computing, 19(9-10): 669- 678. CONCLUSION AND FUTURE WORK Chang, P. L. and W. G. Teng (2007).Exploiting the self-organizing map for medical imagesegmentation, 20th IEEEMedical image segmentation is regard to as a case study international Symposium on Computer-Based Medicalthat is interesting and as well very important. Some Systems(CBMS07), Maribor, 281:288.methods like histogram based, level set and compressionbased method has been seen to be so effective in the Cheng, K.-S.; Lin, J.-S.& Mao, C.-W.(1996). Application ofimage segmentation process, they are also seen to be competitive Hopfield neural network to medical imageeasy and convenient in term of time, accuracy and segmentation.IEEE Transactions on Medical Imagingiterations.We recommend that there methods of image 15(4): 560-561.segmentation should still be checked so closely andthoroughly so that any mistakes and lapses will be Chua, L. O. & Yang, L. (1988a). Cellular neural networks, IEEEcorrected. The correction can be done by reorganizing international Symposium, Circuits and Systems.
Chua, L. O. & Yang, L. (1988b). Cellular neural networks: Luan, Q.X.; Li, X., Kang, J.Y. &Zhi, Z. (2007).Analysis of dental Applications. IEEE transactions on circuits and systems, plaque by using cellular neural network-based image 35(10): 1273-1290. segmentation.Chinese journal of stomatology 42(12): 720- 722.Chua, L. O. & Yang, L. (1988c). Cellular neural networks: Theory. IEEE transactions on circuits and systems, 35(10): 1257- Ma, Y. D.; Liu, Q. &Zhi-Bai, Q. (2004). Automated image 1272. segmentation using improved PCNN model based on cross- entropy, Proceeding of 2004 International Symposium onColodro, F. &Torralba, A. (1996). Cellular neuro-fuzzy networks Intelligent Multimedia, Video and Speech Processing, Hong (CNFNs), a new class of cellular networks, Proceeding of Kong, China. 5th IEEE on Fuzzy Systems, New Orleans, LA, USA Magnotta, V. A.; D. Heckel, Andreasen, N.C., Corson, P.W.,Eckhorn, R.; Reitboeck, H. J., Arndt, M. &Dicke, P. (1989). Ehrhardt, J.C. &Yuh, W.T.C. (1999). Measurement of Feature linking via stimulusevoked oscillations: brain structures with artificial neural networks: Two- and Experimental results from cat visual cortex and functional three-dimensional applications. Radiology 211(3): 781-790. implications from a network model, International Joint Conference on Neural Network, Washington, DC, USA. Mangin, J. F.; F. Poupon, Duchesnay, E., Riviere, D., Cachia, A., Collins, D.L., Evans, A.C.,Fu, J. C.; Chen, C. C., Wong S.T.C. & Li, I.C. (2010). Image segmentation by EM-based adaptive pulse coupled neural Regis, J. (2004). Brain morphometry using 3D moment networks in brain magnetic resonance imaging. invariants.Medical Image Analysis, 8(3): 187-196. Computerized Medical Imaging and Graphics 34(4): 308- 320. Powell, S.; V. A. Magnotta, Johnson, H., Jammalamadaka, V.K., Pierson, R. &Andreasen,Gacsádi, A. &Szolgay, P. (2010).Variational computing based N.C. (2008). Registration and machine learning-based automated segmentation methods for medical imaging by using CNN, segmentation of subcortical and cerebellar brain structures. 12th international Workshop on Cellular Neural Networks NeuroImage39(1): 238-247. and Their Applicaiton (CNNA), 1:6.. Rickard, H. E.; Tourassi, G. D. &Elmaghraby, A.S. (2004).BreastGiménez-Martínez, V. (2000).A modified Hopfield auto-associative segmentation in screening mammograms using multiscale memory with improved capacity.IEEE Transactions on analysis and self-organizing maps, Proceedings of the 26th Neural Networks, 11(4): 867-878. Annual International Conference of the IEEE EMBS, San Francisco, CA.Grassi, G. and P. Vecchio (2006).Cellular neural networks for object- oriented segmentation.Research in Microelectronics and Selvathi, D.; Selvaraj, H. &Selvi, S.T. (2010). Hybrid approach for Electronics 2006, Otranto. brain tumor segmentation in magnetic resonance images using cellular neural networks and optimization techniques.Halkiotis, S.; Botsis, T. &Rangoussi, M. (2007). "Automatic International Journal of Computational Intelligence and detection of clustered microcalcifications in digital Applications 9(1): 17-31. mammograms using mathematical morphology and neural networks."Signal Processing, 87(7): 1559-1568. Shi, W.; Miao, Y., Zhanfang, C. &Hongbiao, Z. (2009). Research of automatic medical image segmentation algorithm based onHassanien, A. E. & Ali J.M. (2004). Digital mammogram Tsallis entropy and improved PCNN, segmentation algorithm using pulse coupled neural InternatioalConference on Mechatronics and Automation, networks, Third International Conference on Image and Changchun. Graphics (ICIG04), Hong Kong.Hopfield, J. J. (1982). Neural networks and physical systems with Shitong, W. & Min, W. (2006). A new detection algorithm (NDA) emergent collective computational abilities.Proceedings of based on fuzzy cellular neural networks for white blood cell the National Academy of Sciences of the United States of detection. IEEE Transactions on Information Technology in America, 79(8): 2554-2558. Biomedicine, 10(1): 5-10.JabaroutiMoghaddam, M.; Rahmani, R. &Soltanian-Zadeh, H. Sing, J. K., D. K. Basu, Nasipuri, M. &Kundu, M. (2005). Self- (2009).Automatic segmentation of putamen using geometric adaptive RBF neural networkbased segmentation of medical moment invariants.The 15th IranianConference on images of the brain, Proceedings of International Biomedical Engineering, Mashad, Iran. Conference on Intelligent Sensing and Information Processing.JabaroutiMoghaddam, M. &Soltanian-Zadeh, H. (2009). Automatic segmentation of brain structures using geometric moment Wang, S.; Fu, D., Xu, M. &Dewen, H.U. (2007). Advanced uzzy invariants and artificial neural networks. Information cellular neural network: Application to CT liver processing in medical imaging : 21st International images."Artificial Intelligence in Medicine 39(1): 65-77. Conference on Information Processing in Medical Imaging (IPMI’09), Williamsburg, VA, 21: 326-337.
Wu, Y. &Pados, D.A. (2000).A feedforward bidirectional associative memory.IEEE Transactions on Neural Networks, 11(4): 859-866.Xiao, Z.; Shi, J. & Chang, Q. (2009). Automatic image segmentation algorithm based on PCNN and fuzzy mutual information, IEEE internatioal Conference on Computer and Information Technology, 241:245.Yang, T.; Yang, L.-B., Chai, W.W. & Chua, L.O. (1996). Fuzzy cellular neural networks: theory, 4th IEEE International Workshop on Cellular Neural Network, Seville, Spain.Zhang, Z.; Xia, S.R. &Duan, H.L. (2007). Cellular neural network based urinary image segmentation. 3th International Conference on Natural Computation (ICNC), Haikou, Hainan.Zhang, Z. C.; Xia, S.R. &Duan, H.L. (2008). Cellular neural network based urinary sediment image segmentation. Zhejiang DaxueXuebao (Gongxue Ban)/Journal of Zhejiang University (Engineering Science) 42(12):z2139-2144.