-

In this paper , a method of dimensionality reduction for MR image sequence data OF brain to a three-dimensional space is presented.
Reduction of dimensions is motivated by the need to visualize the distribution of data and interactively segment the MR image sequence.
New feature space is extracted using a non-linear neural network.
A genetic algorithm is use to search for the neural network parameters such that the transformed data in the network output optimized a specific objective function.
Three objective functions of proposed based on Sommon’s cost functions were in to of them constraint is added to the cost function . The data in the three-dimensional(3-D) output feature space is visualized using the perspective image of the 3-D histogram.
MR images are segmented interactively by determining the cluster centers in the perspective images.
The results of the proposed methods are compared with those of linear transformation and back-propagation neural network (BPNN) methods.
For simulated MR images , to of the proposed criterions product better results in terms
of cluster separation.
Based on 10 real MR image sequence data the same to criterions result in lower segmentation error rates.