A NEW METHOD FOR COLOURIZING OF MULTICHANNEL MR IMAGES BASED ON REAL COLOUR OF HUMAN BRAIN

Transcription

1 A NEW METHOD FOR OLOURIZING OF MULTIHANNEL MR IMAGES BASED ON REAL OLOUR OF HUMAN BRAIN Mohaad Hossein Kadbi; Eadoddin Fateizaheh; Abouzar Eslai; and arya khosroshahi Electrical Engineering Dept., Sharif University of Tech.,Tehran, Iran ABSTRAT In this paper, we propose a novel approach for colourizing of ultichannel MR iages based on real colour of huan brain. At first, we use independent coponent analysis (IA) to transfer three MR odalities into different representations, which enhance the physical characteristics of tissues. Then, Statistical Features are extracted fro three independent coponents and their wavelet coefficients to generate a feature vector for each pixel of iages. Huge aount of features are reduced by using soe feature reduction ethods such as rinciple oponent Analysis (A) to classify the brain into priary classes. The colour space of Visible Huan roject (VH) dataset is quantized to colours by using a partially supervised algorith included selforganization ap and Linear Vector Quantization (LVQ) algoriths. Acquired colours are allotted to each pixel based on texture and intensity of iage in two different steps. onsequently, a fast-autoated algorith is proposed, which has low sensitivity to variation of pixel intensity and precise result regarding to VH iages.. INTRODUTION Medical iages fro different odalities can provide copleentary inforation about anatoy or physiology in a synergistic anner. T-weighted, T-weighted and protondensity-weighted (D), which is obtained fro MRI systes, are popular odalities in practical edical diagnosis and these are widely used in iage processing. Although these iages show the anatoical structure of the patient with an accuracy of illietres, the intensity value in the iages only represents the tie constant of spin relaxation in the case of MRI. The intensity value does not represent the colour. In recent years, there has been considerable research effort to the colouring of edical iages. However, these ethods cannot represent a realistic colour due to eployent of linear apping to transfer the colour to tissues. In, Muraki et.all introduced an eleentary non-linear apping ethod using Radial Basis Function (RBF) network []. They used IA to extract independent coponents of MR odalities. However, that ethod is based on intensity colour apping that ake it sensitive to noise and iage artefact as well as needing a long tie to perfor a colour result. In this paper, we present a new approach, which has a fast and reliable perforance and fine colour in coparison with Visible Huan Data (VHF), distributed by national Library of Medicine (NLM). IA is applied to three odalities to separate the coponents of interest, which are due to taskrelated brain activation, fro other coponents []. Soe robust statistical features are extracted fro three independent iages and their wavelet coefficients, which contain textural and spatial inforation about iages. Best features are selected using soe feature reduction ethods such as A and used for segenting brain into soe classes and eventually these classes are apped into colour space regarding a new colour space quantization algorith based on Linear Vector Quantization (LVQ).This algorith has copletely described in separate article []. The reainder of the paper is laid out thus: Section discusses the Independent oponent Analysis (IA) which obtains independent iages. Next, Section describes features, which are used in texture description and iage segentation. Segentation and classification of iages using fuzzy c-eans (FM) and etc. will be described in Section 4. In Section 5 colour quantization of brain iages applying a new ethod is detailed. Eventually, proposed ethod and experiental results of this paper is discussed, followed by soe conclusions in Section 5 and 6.. INDEENDENT OMONENT ANALYSIS (IA) In agnetic resonance iages (MRI) observed data is a superposition of soe underlying unknown sources such as White atter, Gray atter, cerebro-spinal fluid (SF) and etc. In addition, three utilized odalities, which are used in this approach, are highly correlated with each other and correspond to ixture of physically independent coponents. Therefore, we need to obtain three independent iages, each of the represent one of reinded tissues. This separation of independent sources is reachable by using of IA. Three odalities of T-weighted, T-weighted and proton-density-weighted (D) fro Visible Huan Data (VHD) has been selected as an observed vector T X [ X, X, X ] and IA exerted to achieve three independent sources S S, S, S ], T as [ X A S () IA is considered to be the proble of finding a atrix A and the data sets S that ake each eleent of S as independent as possible. In practice, finding A and S siultaneously 7 EURASI 449

2 fro X is ipossible. Usually, we start fro an appropriate atrix W to define Y W X () and odify W iteratively so that each eleent of Y becoes as independent as possible. If the iteration converges, Y is considered equivalent to S apart fro the scale and the perutation []. For this purpose, the X vector has been centred and whitened using A, at first. Then, using the FastIA algorith, atrix W has been coputed. Once vector W has becoe available, an IA vector S can be coputed fro S W T X () Therefore, we can separate physically independent tissue coponents fro the ultichannel MRI data by using IA. These coponents are suitable for deterining the correspondence with the tissue colour.. FEATURE EXTRATION AROAH The feature extraction in a MR iages is of interest because ight contain essential inforation. The detection of these features is an iportant step in segenting an iage for colouring purposes. When ultispectral iages are used, the result of IA on intensity vectors or MR tissue paraeters T, T and D can be used as input features. The reason of using IA of iage odalities as an input feature is that, it carries inforation rendered by all MR characteristics of the tissue class. Although we use independent coponent of signal intensities at the pixel sites as input features to the segentation process, it can be easily extended when other features such as statistical and spatial features of the intensities are. In this paper, the following three feature sets have been used: -First-order statistical features obtaining fro IA approach. -Wavelet packet based feature containing spatial and highly textured inforation -obination of statistical features and wavelet packet based features... First order statistics Sooth regions can be segented using local grey level statistics such as ean, variance, axiu, iniu, energy, skewness, kurtosis and etc. in a neighborhood of each pixel. First-order statistics are on of the ost significant statistical features which are quite straightforward. They are coputed fro a function that easures the probability of a certain pixel occurring in an iage. This function which is also known as histogra is defined as follows: ng H ( g) ; g,,..., G (4) N where N is the nuber of all pixels in an iage, G is the nuber of gray levels and n is the nuber of pixels of value g g in an iage. The following easures could be extracted by using first-order probability distribution. Mean: Variance: Skewness: Kurtosis: Entropy: Energy: gh ( g) (5) g g g ( g ) H ( ) (6) g g ( g ) H ( g) 4 4 (7) ( g ) H ( g) 4 (8) 6 H ( g) () g In addition to entioned features which include coarse inforation about texture, spatial frequency-based features can be used for highly textured regions. The ain reason we eploy wavelet decoposition as a feature set is that, the textured region has ore high-frequency coponent and the soothing region has ore low-frequency coponents. It has been deonstrated that the soothed iages in the high frequency ebody the textural inforation of the iage, where horizontal, vertical and diagonal sub-iages coprise of the directional inforation of texture. For extracting textural inforation, we used wavelet packet, which has been described below... Wavelet acket Wavelet transfor could be explained as that the original signal is filtered by a series of band-pass filters in different scales. Therefore, we can decopose the signal to a series of frequency bands for analyzing and processing. DWT decoposes a signal in a set of channels whose bandwidth is saller in the low frequency regions. This turns out to be sufficient for regular iages whose inforation is concentrated at low frequencies, but not for textures. In the case of the DWT, only the low frequency block is redecoposed. To avoid this drawback, we can use wavelet packet transfor. Wavelet packet transfor is a generalization of discrete wavelet transfor that leads to over-coplete set of basis functions. Wavelet packets have been described by the following collection of basis functions [4]: () where j is a scale index, k is a translation index, h is a lowpass filter, h is a high-pass filter. The function W can be identified with the scaling function f and W with the other wavelet y. For discrete signal, wavelet packet coefficients ay be coputed efficiently as follows: (9) 7 EURASI 45

3 j n, k j n+, k h ( k). j n, h ( k). j n, () We have applied wavelet packets (coeflet8) to independent iages for levels and extracted soe statistical paraeters such as ean, energy, variance, skewness and kurtosis of each pixel in a specified neighborhood window, in soe sub-bands of second level which coprises directional inforation of texture. 4. UNSUERVISED SEGMENTATION Unsupervised techniques, which are usually called clustering, autoatically find the structure in the iages. A cluster is an area in feature space with a high density. Soe clustering ethods include fuzzy c-eans, Kohonen selforganizing ap and Hopfield neural network. The fuzzy c-eans algorith generalizes the K-eans algorith [5] allowing for soft segentations based on fuzzy set theory. Unlabeled data are labeled by iniizing a generalized within group su of squared error objective function. Kohonen s self-organizing ap generates nodes on a two-diensional lattice in which the distribution of these nodes corresponds to the proxiity of their associated node patterns in the feature space [6,7,8,9]. The opetitive Hopfield neural network (HNN) is an artificial neural network that has been used in [] for the classification of MR iages via energy iniization. The energy function for iniization is defined as the ean of the squared distance easures of the gray levels within each class. 5. OLOUR QUANTIZATION olour quantization is preliinary and necessary process of colourizing iages. It is practically ipossible to generate a colour ap with too any colours. In reducing colours, those ore probable and ore valuable should survive and others be oitted. These two properties do not occur siultaneously for all colours. In brain iages, there are rare colours representing iportant anatoic tissues. In this case, unsupervised clustering Figure - Three stiulated MR odalities contained T- weighted, T-weighted and proton density cannot reach into satisfying results. Rare but valuable colours can easily be oitted during fuzzy clustering, K-eans, neural gas and siilar approach. We have proposed to use seisupervised vector quantization for colour reduction in brain iages. In our study, quantization has been perfored on occasionally labelled colour vectors. The labelled vectors include those rare colours which quantization should preserve the and produce prototypes for the. Our quantization strategy copletely differs fro independent LVQ and self-organization ap (SOM). Interaction between two groups of colour saples enhances colour quantization quality. Our experiental exaples of brain reduced colour iages are so satisfying []. Hence, brain-quantized colours have been eployed here for colour ap instruction. Further these colours, a bilinear interpolation of the is also included in final colour ap to prevent fro discontinuity in quantized iages. For such interpolation, colour arrangeent is iportant; so the visually arranged version of colours has been eployed here. The colour ap for brain iages using sei-supervised algorith is depicted in figure.(d). It s clear that rare colour is presented in this colour ap but discarded in other ethods. 6. ROESED METHOD In the current research, two different databases are used to copare the result on different iages. The first one is siulated MR iages fro BrainWeb institution which contains siulated brain MRI data based on noral anatoical odels, includes T-weighted, T-weighted and proton density odalities and the label of each pixel. Figure shows a saple slice of these database odalities. As we know the label of each pixel so utilize of siulated iages ake the task of validating a segentation ethod easy. MR iages fro Harvard Medical School, is another dataset to exaine the result of proposed ethod on in vivo iages. Figure depicts three odalities of the database. Figure 5 depicts block diagra of proposed ethod to colour MR iages.at first step, we used FastIA to perfor IA on the entioned datasets and obtained three independent coponents, which enhance the physical characteristics of tissues. Feature extraction process has been used to obtain best feature set to segent the iages, as Figure - Obtained colour ap by using soe different ethod: (a).kohonen SOM, (b).fs-som, (c).fm and (d). roposed ethod Figure - Three odalities fro Harvard edical school database 7 EURASI 45

4 Table. ercentage of correctly classified pixel to three general tissues based on different feature sets Seg. ethod Tissue First-order Statistic(stat.) Waveletacket (wavelet) obination of stat. and wavelet Table. ercentage of correctly classified pixel to three general tissues based on third feature set Tissue Seg. ethod Gray atter White atter SF Gray atter Fuzzy c-eans White atter SOM SF Hopfield it has been described previously. The ain idea of using reinded features is to classify pixels based on their statistical and spatial inforation rather than their intensities only. We constructed three sets of features: statistical features with *8 diensions for three iages, wavelet packet based features with * diensions and cobination of statistical and wavelet packet based features with *8 diensions. Therefore, we need a syste to select the best set of features regarding to the result of segentation of these feature space. We avoid the effect of segenting algorith in selecting the best feature space, whereas we have used three vigorous algoriths, eployed forerly. These algoriths included of fuzzy c-ean, kohonen self organization ap and Hopfield neural network. In order to reduce coputational coplexity and iprove the classification perforance those features that ay be irrelevant or redundant have been discarded by Fisher s linear discriinant analysis (LDA). LDA finds that feature space which axiizes the ratio of between-class to withinclass variation jointly for each feature (diension). A could be further applied to find a copact subspace to reduce the feature diensionality. After eploying feature selection and reduction ethods, we have had five features for statistical feature set, eight features for wavelet based feature set and eight features for cobination of these two feature sets. The result of applying these feature sets to segent the iages into three general classes (white atter, gray atter and SF.) has been shown in table. Three different segentation ethods subsequently have been applied to the best features obtained previously. Segentation results for each dataset have been copared to the specified labels in siulated iages and the average nuber of pixels that classified correctly has been calculated. Table depicts the percentage of these correct classifications for each algorith. In fact, in current research, we have not focused Table. Result of different ethod of segentation on different feature set Feature set Seg. ethod Fuzzy c-eans First-order Statistic(stat.) 8.6 Wavelet acket (wavelet) 9. obination of stat. and wavelet 97.6 on segentation approach, so we have selected soe usual unsupervised segentation ethods and have used the best paraeters, obtained in previous works. To read ore details about these algoriths see [,,,4]. Achieved result has shown that fuzzy c-ean algorith has the best perforance to segent the iage, when third feature set (cobination of statistical and wavelet based features) is used. The percentage of correctly classified pixels of siulated MR iages for brainweb dataset to three general tissues has been indicated in table. Result shows that all the tissues especially SF has been classified better than other ethods [5]. Above processes has been perfored on real MR iages and siilar result has been shown for this dataset. Lower values in coparison with siulated MR iages consequences are probable, due to available noises and paraeter in MRI systes. According to quantization of colour space to ajor colours, we can specify the colours regarding to each three significant tissues, experientally. For instance, orange or livid colour can just be appeared in Gray atter. This issue has been accoplished by an expert in MRI field and looking to all the iages in Visual Huan Feale (VHF) atlas. onsequently, we have had liited colours in each tissue which do not appear in other tissues. In regard to result of segentation, each pixel has been appeared in one of the three classes and also we have known the colours of each tissue. Therefore, a siple apping ethod can assign existing colours to specified pixel in that tissue. This apping has been constructed by coparing the intensity of each pixel of class with the intensities corresponding to each colour. Figure 4 (a) (b) (c) Figure 4 - (a): olour iage fro visible huan project (VH) atlas. (b), (c): Result of colouring siulated MR iages using this ethod on two different slices. SOM Hopfield Available in 7 EURASI 45

5 reprocessing Feature Extraction Feature cobination Segentation Fine colourization oloured iage T T D oarse olourization Figure 5. Block diagra of proposed ethod depicts the result of colouring real MR iages using this ethod. Results prove that this ethod has ore realistic colour coparing with previous ethods and rare colour presented in result. In addition this ethod is ore robust to noise than other ethods, because we have used both textural and intesity inforation of iage instead of only intensity inforation. This fact helps us to have a powerful ethod against usual paraeters of MRI syste such as intensity inheogenity and peripheral noises. 7. ONLUSION In this paper a new ethod for colouring brain MR iages is proposed based on two scale processes. Initially, the input iage is segented into three ajor classes. Selecting of these ajor classes is perfored by clustering on spatial and statistical features of independent iage. Indeed each iage pixels is classified to one of these three classes regarding to its textural inforation. There is a subset of quantized colours associated with each ajor class. The texture segentation idea rises fro the anatoical fact that presence of colours in each class of texture is restricted. Hence, a physician deterines the divisions of colour space for each ajor class. The low scale segentation deterines a colour set for each pixel (regarding to its textural class) and does not deterine exact colour. The final step is fine colour selection using intensity colour ap, against previous processes that deal with texture. The final step is the process on individual pixel intensity and enhances final colour iage quality by introducing high frequency details. Although intensity colour ap is copletely unusual, but in this study three aps are predefined. Therefore, that pixel class picks one of the. onsequently, eployed colour ap is constrains by texture which increase certainly for edical colouring. REFERENES [] M. H. Kadbi, E. Fateizadeh, oloring of Magnetic resonance iage (MRI) of Huan brain, Technical report, shairf unversity of technology, 5, pp [4] A. Laine, and J. Fan, Texture classification by wavelet packet signatures, IEEE Trans AMI, Nov. 99, vol.5, p [5] D. L. ha and J. L. rince, An adaptive fuzzy c-eans algorith for iage segentation in the presence of intensity inhoogeneities, att. Rec., pp , 999. [6] J. hou,. hen, and W. Lin, Segentation of dual-echo MR Iages using neural networks, In TO. SIE, Medical Iaging, 99, vol. 898, pp. -7. [7] J. Alirezaie, M. E. Jernigan, and. Nahias, Autoatic Segentation of erebral MR Iages using Artificial Neural Networks, IEEE Trans. on nuclear science, August 998, Vol. 45, No. 4. [8] T. Kohonen, Self-Organization and Associative Meory, Springer-Verlag, New York, 989. [9] M. W. Vannier,. M. Speidel, and D. L Rickan., Magnetic resonance iaging ultispectral tissue classification, News hysiol. Sci., 988, Vol., pp [] K. S. heng, J. S. Lin, and. W. Mao, The application of copetitive Hopfield neural network to edical iage segentation, IEEE Trans., Med. Iaging, 996, pp [] A. K. Jain and R.. Dubes, Algoriths for clustering data, rentice Hall, 988. [] G. Gerig, J. Martin, R. Kikinis, O. Kubler, M. Shenton, and F. A. Jolesz, Unsupervised tissue type segentation of D dual-echo MR head data, Iage Vision oputing, 99, Vol., pp [] W.E. Reddick, J. O. Glass, E. N. ook, T. D. Elkin, and R. J. Deaton, Autoated Segentation and lassification of Multispectral Magnetic Resonance Iages of Brain Using Artificial Neural Networks, IEEE Transactions on Medical Iaging, Dec. 997, Vol. 6, No. 6, pp [4] M. N. Ahed, S. M. Yaany, A. A. Farag, and T. Moriarty, Bias Field Estiation and Adaptive Segentation of MRI Data Using a Modified Fuzzy -Means Algorith, IEEE Trans, 999, pp [5] L. Mona, F. Laberti,. Deartini, A Neural Network Approach to Unsupervised Segentation of Single-hannel MR Iages, st IEEE EMBS onf. on Neural Engineering, Mar., pp [] S. Muraki, T. Nakai, and Y. Kita, Basic Research for olouring Multichannel MRI Data, ETL Technical Report TR--,, pp [] M. J. McKebwn, S. Makeig, G. Brown, T.. Jung, S. Kinderann, A. Bell, and T. J. Sejnowski, Analysis of fri data by blind separation into independent spatial coponents, Hu.Brain Mapping, 998, vol. 6, pp EURASI 45