Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more
Download
Standard view
Full view
of .
Look up keyword
Like this
3Activity
0 of .
Results for:
No results containing your search query
P. 1
Heschl's Gyrus Auditory Cortex Slice Registration Using Echo State Neural Network (ESNN)

Heschl's Gyrus Auditory Cortex Slice Registration Using Echo State Neural Network (ESNN)

Ratings: (0)|Views: 76|Likes:
Published by ijcsis
This paper presents Herschel’s gyrus auditory cortex slice registration using Echo state neural network (ESNN). Training the network is done with translation and rotational values of the selective points (feature points) from two images at a time (source and target images). The input layer is given with coordinates of the selective points of the source image and in the output layer; the labeling is the translation
and rotational values of the selective points of the target image. ESNN is an estimation network which estimates the required registration information from the selective points of target and source image. The output of ESNN is compared with radial basis function (RBF).
This paper presents Herschel’s gyrus auditory cortex slice registration using Echo state neural network (ESNN). Training the network is done with translation and rotational values of the selective points (feature points) from two images at a time (source and target images). The input layer is given with coordinates of the selective points of the source image and in the output layer; the labeling is the translation
and rotational values of the selective points of the target image. ESNN is an estimation network which estimates the required registration information from the selective points of target and source image. The output of ESNN is compared with radial basis function (RBF).

More info:

Published by: ijcsis on Mar 08, 2011
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

03/19/2011

pdf

text

original

 
HESCHL'S GYRUS AUDITORY CORTEXSLICE REGISTRATION USINGECHO STATE NEURAL NETWORK (ESNN)
1
R.Rajeswari,
Research Scholar,Department of Computer ScienceMother Theresa Women’s University,Kodaikanal, India.Email:rajeswaripuru@gmail.com
2
Dr.Anthony Irudhayaraj
Dean, Information Technology,Arupadai Veedu Institute of TechnologyPaiyanoor-603104, India.E-mail: <anto_irud@hotmail.com>,
Abstract
This paper presents Herschel’s gyrusauditory cortex slice registration using Echo stateneural network (ESNN). Training the network is donewith translation and rotational values of the selectivepoints (feature points) from two images at a time(source and target images). The input layer is given withcoordinates of the selective points of the source imageand in the output layer; the labeling is the translationand rotational values of the selective points of the targetimage. ESNN is an estimation network which estimatesthe required registration information from the selectivepoints of target and source image. The output of ESNNis compared with radial basis function (RBF).
Keywords-Echo state neural network, functional magnetic resonance imaging (fMRI), Heschl's gyrus,auditory cortex
I.
 
INTRODUCTION
 
The image registration [1] aims to find atransformation that aligns images of the same scenetaken at different times, from different viewpoints. Ithas been studied in various contexts due to itssignificance in a wide range of areas, includingmedical image fusion, remote sensing, and computervision. Medical image acquisition systems generatedigital images that can be processed by a computerand transferred over computer networks. Digitalimaging allows extracting objective, quantitativeparameters from the images by image analysis.Medical image analysis exploits the numericalrepresentation of digital images to develop imageprocessing techniques that facilitate computer-aidedinterpretation of medical images. The continuingadvancement of image acquisition technology and theresulting improvement of radiological image qualityhave led to an increasing clinical need andphysician’s demand for quantitative imageinterpretation in routine practice, imposing new andmore challenging requirements for medical imageanalysis[2][3]. A fundamental problem in medicalimage analysis is the integration of information frommultiple images of the same subject, acquired usingthe same or different imaging modalities and possiblyat different time points. One essential aspect thereof is image registration, i.e., recovering the geometricbetween corresponding points in multiple images of the same scene. While various more or lessautomated approaches for image registration havebeen proposed in the field of medical imaging andimage analysis, one strategy in particular, namelymaximization of mutual information[4][5] , has beenextremely successful at automatically computing theregistration of 3-D multimodal medical images of various organs from the image content itself.
II.
 
MATERIALS AND METHODS
A.
 
Neural Network Structures
The Echo state neural network is used forlearning the images. The number of neurons in theinput layer is 4, and the number of neurons in theoutput layer is 6.
Input layer description
Node 1 = x coordinate of point in image 2(targetimage)Node 2 = y coordinate of point in image 2(targetimage)Node 3 = x coordinate of point in image 1(imageto be registered with target image)Node 4 = y coordinate of point in image 1(imageto be registered with target image)
Output layer description
Node 1= vertical shift
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 9, No. 2, February 2011204 http://sites.google.com/site/ijcsis/ISSN 1947-5500
 
Node 2= upward (1) or downward (2)Node 3=horizontal shiftNode 4= left (1) or right (2)Node 5= angle with respect to axis passingthrough centre of the imageNode 6= left (1) or right (2)The hidden layer has been trained with differentnumber of nodes increasing from 2 neurons.The target values corresponding to x, y values of image 1 and image2 are calculated as followsTS=size (Directions, 1)for i=1:TS-1%1I=Directions(i,:);F=Directions(i+1,:);X=F(1,1)-I(1,1);Y=F(1,2)-I(1,2);if X==0 & Y==1D(i)=1;elseif X==0 & Y==1D(i)=2;elseif X==-1 & Y==0D(i)=3;elseif X==1 & Y==0D(i)=4;elseif X==-1 & Y==-1D(i)=5;elseif X==1 & Y==1D(i)=6;elseif X==-1 & Y==1D(i)=7;elseif X==1 & Y==-1D(i)=8;endendTable 1 shows the direction of rotation amongpixel coordinates of source and target image. The sizeof the image considered is 63 rows by 63 columns.The term ‘T’ refers to target image and ‘S’ refers tosource image. Curved arrow to the right is theclockwise direction and the curved arrow to the left isthe counter clockwise direction. Table 1 shows thepossible rotation of the pixel of source image todifferent location in target image.Table 2 presents 10 sample pixel coordinates thatis used for training the network. For testing thenetwork, the same sample points with another 10points (total 20 points) are presented.The description of Table is as follows.Column 1 = pattern numberColumn 2= x coordinate of points in target imageColumn 3= y coordinate of points in target imageColumn 4= x coordinate of points in source imageColumn 5= y coordinate of points in source imageColumn 6= shift in rowsColumn 7= Upward or downward translationColumn 8= shift in columnsColumn 9 = Horizontal translationColumn 10= Rotation of source pixel coordinatewith respect to corresponding target pixel coordinateColumn 11= Clock wise or counterclockwiserotation
Table 1 Rotation of source coordinates fromTarget image coordinates
12345
Table 2 Patterns used for training and testing ESNN
Input pattern Target patternTarget(actual) Source(distorted) Translation (pixel) Rotation (degrees)
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 9, No. 2, February 2011205 http://sites.google.com/site/ijcsis/ISSN 1947-5500
 
Patternnumberx y x y Vertical shiftUpward(1)Downward(2)Horizontal shiftLeft(1)Right(2)Anglerotated DirectionCW(2) / CCW(1)
1 3 14 1 17 2 1 1 2 3.05 22 5 41 3 42 2 1 1 2 0.59 13 22 47 19 48 3 1 1 2 5.4 14 34 47 32 48 2 1 1 2 7.59 25 38 18 36 18 2 1 0 0 7.25 16 28 6 27 7 1 1 1 2 2.56 27 48 14 47 15 1 1 1 2 0.2 28 49 45 48 45 1 1 0 0 1.68 19 36 62 35 62 1 1 0 0 1.88 110 13 57 12 58 1 1 1 2 0.33 1
B ECHO STATE NEURAL NETWORK (ESNN)
An Artificial Neural Network (ANN) is anabstract stimulation of a real nervous system thatcontains a collection of neuron units, communicatingwith each other via axon connections. Artificialneural networks are computing elements which arebased on the structure and function of the biologicalneurons. These networks have nodes or neuronswhich are described by difference or differentialequations.Dynamic computational models require the abilityto store and access the time history of their inputs andoutputs. The most common dynamic neuralarchitecture is the Time-Delay Neural Network (TDNN) that couples delay lines with a nonlinearstatic architecture where all the parameters (weights)are adapted with the back propagation algorithm.Recurrent Neural Networks (RNNs) implement adifferent type of embedding that is largelyunexplored. RNNs are perhaps the most biologicallyplausible of the Artificial Neural Network (ANN)models. One of the main practical problems withRNNs is the difficulty to adapt the system weights.Various algorithms, such as back propagationthrough time and real-time recurrent learning, havebeen proposed to train RNNs; these algorithms sufferfrom computational complexity, resulting in slowtraining, complex performance surfaces, thepossibility of instability, and the decay of gradientsthrough the topology and time. The problem of decaying gradients has been addressed with specialprocessing elements (PEs). ESNN possesses a highlyinterconnected and recurrent topology of nonlinearPEs that constitutes a reservoir of rich dynamics andcontains information about the history of input andoutput patterns. The outputs of this internal PEs (echostates) are fed to a memory less but adaptive readoutnetwork (generally linear) that produces the network output. The interesting property of ESNN is that onlythe memory less readout is trained, whereas therecurrent topology has fixed connection weights. Thisreduces the complexity of RNN training to simplelinear regression while preserving a recurrenttopology, but obviously places important constraintsin the overall architecture that have not yet been fullystudied.The echo state condition is defined in terms of thespectral radius (the largest among the absolute valuesof the eigenvalues of a matrix, denoted by (|| || ) of thereservoir’s weight matrix (|| W || < 1). This conditionstates that the dynamics of the ESNN is uniquelycontrolled by the input, and the effect of the initialstates vanishes. The current design of ESNNparameters relies on the selection of spectral radius.There are many possible weight matrices with thesame spectral radius, and unfortunately they do notperform at the same level of mean square error(MSE) for functional approximation.
Figure 1 Echo State Network (ESNN)
ALGORITHM
1.Read data2.Separate into inputs (datain) and target outputs(dataout)3.Initialize number of reservoirs
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 9, No. 2, February 2011206 http://sites.google.com/site/ijcsis/ISSN 1947-5500

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->