
For FullText PDF, please login, if you are a member of IEICE,
or go to Pay Per View on menu list, if you are a nonmember of IEICE.

Annealed Hopfield Neural Network with Moment and Entropy Constraints for Magnetic Resonance Image Classification
JzauSheng LIN
Publication
IEICE TRANSACTIONS on Information and Systems
Vol.E83D
No.1
pp.100108 Publication Date: 2000/01/25
Online ISSN:
DOI:
Print ISSN: 09168532 Type of Manuscript: PAPER Category: Biocybernetics, Neurocomputing Keyword: simulated annealing, mean field annealing, annealed Hopfield neural network, magnetic resonance image,
Full Text: PDF>>
Summary:
This paper describes the application of an unsupervised parallel approach called the Annealed Hopfield Neural Network (AHNN) using a modified cost function with moment and entropy preservation for magnetic resonance image (MRI) classification. In the AHNN, the neural network architecture is same as the original 2D Hopfield net. And a new cooling schedule is embedded in order to make the modified energy function to converge to an equilibrium state. The idea is to formulate a clustering problem where the criterion for the optimum classification is chosen as the minimization of the Euclidean distance between training vectors and clustercenter vectors. In this article, the intensity of a pixel in an original image, the first moment combined with its neighbors, and their graylevel entropy are used to construct a 3component training vector to map a neuron into a twodimensional annealed Hopfield net. Although the simulated annealing method can yield the global minimum, it is very timeconsuming with asymptotic iterations. In addition, to resolve the optimal problem using Hopfield or simulated annealing neural networks, the weighting factors to combine the penalty terms must be determined. The quality of final result is very sensitive to these weighting factors, and feasible values for them are difficult to find. Using the AHNN for magnetic resonance image classification, the need of finding weighting factors in the energy function can be eliminated and the rate of convergence is much faster than that of simulated annealing. The experimental results show that better and more valid solutions can be obtained using the AHNN than the previous approach in classification of the computer generated images. Promising solutions of MRI segmentation can be obtained using the proposed method. In addition, the convergence rates with different cooling schedules in the test phantom will be discussed.

