An efficient multiclass classifier for classification of Alzheimer's disease/mild cognitive impairment/Normal subjects

Abstract

Typically, in sparse representation‐based classifiers, the weight associated with each training sample is ignored, resulting in reduced accuracy. Moreover, individual binary classifiers solved a multiclass problem. It requires more time as multiple runs are needed to compute the accuracy. In this paper, we propose a novel optimal sparse representation‐based classifier. It solves the ternary classification problem with improved accuracy in a single run. The ternary classification considers Alzheimer's disease versus mild cognitive impairment versus normal control in a single run. A two‐stage sparse representation model is used to design the proposed classifier. To update the weight coefficients, we suggest a regularized Levenberg–Marquardt learning. It allows selecting a subset of significant training samples. To determine the appropriate subset size, we investigate an objective function in terms of classification accuracy. For optimization, we suggest a hybrid particle swarm optimization–squirrel search technique. The experiment conducted on the Alzheimer's Disease Neuroimaging Initiative database shows our method outperforms other state‐of‐the‐art methods in terms of computation time and accuracy. The use of different training–testing partition ratios makes the proposed method immune to biased results, overfitting, and underfitting difficulties. Moreover, results are obtained from 100 iterations to confirm its stability. The suggested model may be helpful for further research in medical image analysis.