A new feature clustering method based on crocodiles hunting strategy optimization algorithm for classification of MRI images

Abstract

In complex data with high dimensions, the dimension reduction methods are used to increase accuracy and speed in the classification algorithms. Feature clustering methods have had a good performance in the selection of important features of data due to using clustering methods. The process of selecting important features of data is a challenge in feature clustering methods which has led to the creation of different algorithms with different performances. The combination of the clustering methods and metaheuristic algorithms, especially the kind of population-based algorithms, have had good results in most cases. In this paper, a new feature clustering method is proposed which is used as a dimension reduction in the classification of brain tumors in 900 magnetic resonance images (MRI). The classification algorithm includes three main steps: in the first step, the Google-Net and ResNet-18 methods have been used for feature extraction of MRI images. Due to the creation of many features using the Google-Net and ResNet-18 methods, a new proposed feature clustering is introduced to reduce the feature dimensions in the second step. In designing the feature clustering algorithm, a new metaheuristic algorithm is introduced which is called the crocodiles hunting strategy optimization algorithm (CHS) that simulates crocodiles’ behavior in hunting. Also, the feature clustering algorithm introduced the new chromosome encoding for feature clustering which is called feature clustering based on the crocodiles hunting strategy optimization algorithm (FC-CHS). Finally, in the third step, the support vector machine (SVM) algorithm is used for classification. According to the results of classification on the MRI images, the proposed algorithm has achieved high accuracy in Google-Net and ResNet features based on confusion matrices. For comparing the performance of the FC-CHS, this algorithm is compared with five well-known dimension reduction algorithms. Also, real data are used to further investigate the performance of the FC-CHS algorithm. The results show that the combination of the FC-CHS and SVM algorithms have been reached high accuracy in Iris, and Wine data, and in other real data, the proposed algorithm is outperformed compared to other dimension reduction methods in most cases.