SEGMENTATION WITH NEURAL NETWORK

B.Prasanna Rahul Radhakrishnan
Valliammai Engineering College Valliammai Engineering College prakrish_2001@yahoo.com krish_rahul_1812@yahoo.com

Abstract: Our paper work is on Segmentation by Neural networks. Neural networks computation offers a wide range of different algorithms for both unsupervised clustering (UC) and supervised classification (SC). In this paper we approached an algorithmic method that aims to combine UC and SC, where the information obtained during UC is not discarded, but is used as an initial step toward subsequent SC. Thus, the power of both image analysis strategies can be combined in an integrative computational procedure. This is achieved by applying “Hyper-BF network”. Here we worked a different procedures for the training, preprocessing and vector quantization in the application to medical image segmentation and also present the segmentation results for multispectral 3D MRI data sets of the human brain with respect to the tissue classes “ Gray matter”, “ White matter” and “ Cerebrospinal fluid”. We correlate manual and semi automatic methods with the results.

Keywords: Image analysis, Hebbian learning rule, Euclidean metric, multi spectral image segmentation, contour tracing.

Introduction: Segmentation can be defined as the identification of meaningful image components. It is a fundamental task in image processing providing the basis for any kind of further highlevel image analysis. In medical image processing, a wide range of application is based on segmentation. A possible realization of high-level image analysis principle is the acquisition and processing of multisprectral image data sets, which forms the basis of the segmentation approach. A good survey is provided by the list of citations published in [1] that may serve as a good starting point for further reference. Different segmentation methods range from simple histogram-based thresholding or region growing algorithms, to more sophisticated techniques such as active contours or watershed transformation. Appropriate preprocessing steps comprise the anatomically correct registration of the data sets and masking a region of interest in which the segmentation should be performed. Data analysis may be performed by two different strategies. The first one tries to identify characteristic properties of the multidimensional data distribution of unlabeled feature vectors. This approach is called as unsupervised clustering (UC). The second strategy involves labeled data, i.e., the learning algorithm requires both the feature vector itself and a target function defining its interpretation with regard to segmentation class membership. We call it supervised classification (SC).

Structure and function of the Hyber-BF network: The general architecture of a
Hyper-BF network is shown in the figure It consists of three layers of neurons:
Input layer, hidden layer, and outer layer

[pic] n” input neurons with activations xi , i €{1,……,n}, the activation pattern of the input layer is represented by an n-dimensional vector x in the feature space R. This activation is propagated to the N neurons of the hidden layer by directed connections with synaptic weights Wji. The synaptic weights Wj € R, j € {1,……, N}, Are computed as a set of prototypical vectors that represent the data set in the feature space. The activation aj of the hidden layer neuron j is chosen as a function of distance d=||x-wj|| of the data vector x with respect to the virtual position wj of the hidden layer neuron j. d hereby defines an arbitrary metric in the feature space, e.g., the Euclidean metric. The term virtual position is based on the idea that the activation aj of the hidden layer neuron should take its maximum value xmax, which can be looked at as a specification of the neuron j with respect to the position xmax. It is obviously reasonable to choose aj as a monotonically declining function of d, i.e., the activation aj of the hidden layer neuron should decline with increasing distance between x and the virtual position wj . A simple choice is an isotropically decreasing function aj,i.e., he declining behavior does not depend on the direction of the difference vector (x-wj). From this results a symmetry with respect to rotation, i.e., a radial decline of aj(x) in the neighborhood of wj: Therefore, we refer to the activation function aj(x) as a Hyper-BF network.With respect to favorable properties regarding function approximation, F.Girosi and T.poggio [2] .The Gaussian activation functions.
[pic]
Moody and Darken [3] propose a global normalization of the hidden layer activation by
[pic]

Which results in a hidden layer activation of
[pic]

The radial symmetry of the activation function is obviously lost by the normalization.
In [2] such a system is called “Hyper-BF Network. In the final step, a linear signal propagation of the hidden layer activation is performed to the m neurons of an output layer by weighted summation.
[pic]

Or in terms of single components,

[pic] In neural network computing this mapping corresponds to a structure called the perceptron (Rosenblatt [4]).

Training procedure: The memorized “knowledge” of the network is represented by the parameter P={(wj,sj,pj)} ,i.e., the weights wj and sj as well as the range pj of the Gaussians in the hidden layer .Here, The two aspects of should be considered : On one hand , the network should reproduce the given training data as well as possible ; on the hand , it should be able to provide some reasonable generalization with respect to unknown test data. An obvious training strategy is the minimization of the error.

[pic]

An attempt to simultaneous optimization of all the free parameters by gradient descend on the cost function. A proposal by Moody and Darken [3] is to split up the parameter space p={(wj,sj,pj)} into three smaller partitions {wj},{sj},{pj} that may be optimized separately. The resulting dimensionality reduction involves an increased convergence rate for the optimization process within the different partitions. This concept of separate optimization within each of the three partitions of p is the basis of the Hyper-BF architecture.

Optimization: Initially the virtue positions wj of the hidden layer neurons are calculated in a procedure called vector quantization (VQ).This is based on an unsupervised clustering Algorithm VQ algorithm can be classified according to the cooperativity function aj .One can distinguish between so called hard clustering method that assign each feature vector x to exactly one cluster represented by the Codebook vector wj and soft clustering methods of fuzzy membership of feature vectors in several clusters. A simple method for hard clustering is the algorithm proposed by Linde, Buzo, and Gray (LBG) [5]. The procedure can be monitored by various control parameters such as the free energy, entropy and reconstruction error, which enable an easy detection of cluster splitting. This version is well suited for implementation on parallel computers .The output weights can be calculated in a procedure called supervised learning. They can be determined by two different ways: global training procedure and local training procedure based on Hebbian learning rule.

[pic]
Application to medical image segmentation: Hyper-BF neural network provides the basis for the application to automatic segmentation of magnetic resonance imaging data sets of the human brain. The four acquisition sequences are T1 weighted, T2 weighted, proton density weighted and inversion recovery sequences. Segmentation aims at classifying each voxel of the multispectra data set as belonging to a specific tissue type, thus obtaining information about structure and volume of the tissue classes. A classical problem with numerous clinical applications is the segmentation of brain imaging data.
With respect to the tissue classes gray matter, white matter, and cerebrospinal fluid. A threefold classification of brain tissue may be sufficient for numerous clinical applications, it should be emphasized that the concept presented in
This paper can be extended to an arbitrary number of tissue classes.
Especially one may think of introducing additional classes for the identification of pathological tissue.

Image Data
[pic]
The image data sets were obtained on 1.5T whole body MRI scanner (siemens, Magneton vision). Four persons were examined in a prospective study employing a standardized MRI sequence protocol: T1 weighted MP-RAGE, T2 weighted and proton density (PD) weighted spin echo, and Inversion-Recovery (IR) sequences. The MRI acquisition parameter of each sequence are listed in table.

[pic]

PREPROCESSING The concept of voxel- based multispectral image segmentation requires anatomically correct alignment of the data sets acquired within different image acquisition procedures. This is known as image registration. All the extracerebral structure those are not required for the tissue classification task should be excluded from the data set. This is pre-segmentation. There is a wide scope of methods for pre-segmentation ranging from manual contour tracing to semiautomatic or fully automatic procedures. The latter exist in numerous implementations and are available on a commercial basis. They are frequently based on filter operations or region-growing algorithms

[pic]

Classification Vector quantization can determine a set of prototypical codebook vectors wj representing the feature space, a set of feature vectors x in a grey-level feature space G. This provides the basis for segmentation of the imaging data set with respect to different tissue classes. Two alternative approaches are Interactive assignment of codebook vectors and supervised classification .The comparison of above approaches are discussed in figure.

[pic]

Results
The paper represents the results for the segmentation of 3D MRI data sets of the human brain. Multispectral MRI data were acquired in four healthy human volunteers. Initially, the four data sets were preprocessed according to the procedure explained before.
After preprocessing, the data sets were segmented by applying two different strategies:
Manual interactive assignment of codebook vectors to tissue classes. At first the result for the vector quantization of the gray value feature space with manual interactive assignment of codebook vector of tissue classes will be discussed .The average manual processing time for the assignment of codebook vectors to tissue classes are 6 minutes per data set, i.e., the procedure requires only a little human intervention. However this approach yields suboptimal results, especially in case where a code book vector cannot be attributed to a single tissue class without ambiguity. Sometimes a codebook vector is positioned at the border between two tissue classes, which makes a clear assignment difficult or even impossible.
Semiautomatic classification by Hyper-BF network
In the following, the results for classification by a Hyper-BF network after preceding vector quantization of the gray level feature space are discussed. The training data comprise 1%of the whole data set each. This approach yields best segmentation result than unsupervised clustering with respect to subjective evaluation by a neuroradiologist. The improvement of segmentation quality was accompanied by a considerable increase of human intervention. Manual processing time for interactive labeling of the training data was approximately 30 minutes per data set. Thus supervised classification requires more human intervention than unsupervised clustering.

Discussion In this paper we have presented a neural network approach to automatic segmentation of MRI data sets of the human brain by neural network computation. The Hyper-BF network computation. The Hyper-BF architecture enables an effective combination of unsupervised and supervised learning procedures. Although the results mentioned above are very encouraging, several critical issues still need to be discussed with regard to the data, the aim of segmentation and the validation of the results. In summary, a wide range of further research topics has to be covered in future studies based on the results presented in this article. Such investigations can help to introduce automatic neural network segmentation as a cost-effective and reliable tool for routine medical image processing according to the growing importance of quantitative image analysis techniques for clinical decision making.

References
1.A.J.Worth . http://demonmac.mgh.harvard.edu/cma/seg_f/seg_refs.html. 2.F.Girosi and T.Poggio, Networks and the best approximation property. A.I.Memo 1164,Massachusetts Institute of Technology, 10,1989.
3.J.Moody and C.Darken. Fast learning in networks of locally tuned processing units.
4. F.Rosenblatt.Principles of neurometrics. Spartans books,1962.
5.Y.Linde,A.Buzo,and R.M.Gray . An algorithm for vector quantizer design.
6.D.O.Hebb.The organization of behaviour.

[pic]