A new neural unsupervised classification approach using amended competitive Hebbian learning: PET image segmentation insights

Abstract

This paper proposes a new classification procedure based on the competitive concept, divided into three processing stages. It begins by the estimation of the Probability Density Function (pdf), followed by a competitive training neural network with the Mahalanobis distance as an activation function. This stage allows detecting the local maxima of the pdf. Then, we use the competitive Hebbian learning to analyse the connectivity between the detected maxima of the pdf upon the Mahalanobis distance. The so detected groups of maxima are then used for the classification process. Compared to the K-means clustering or the clustering approaches based on the different competitive learning schemes, the proposed approach has proven, under a number of real (positron emission tomography image) and synthetic data samples, that it does not pass by any thresholding and does not require any prior information on the number of classes or on the structure of their distributions in the data set.Keywords: PET image; positron emission tomography; probability density function; competitive neural networks; Mahalanobis distance; competitive Hebbian learning; topology preserving feature; K-means; classification; competitive concept; thresholding. 2 M. Timouyas, S. Eddarouich and A. HammouchReference to this paper should be made as follows: Timouyas, M., Eddarouich, S. and Hammouch, A. (2017) 'A new neural unsupervised classification approach using amended competitive Hebbian learning: pet image segmentation insights', Int.