GLCM and statistical features extraction technique with Extra-Tree Classifier in Macular Oedema risk diagnosis

Abstract

When eye images are properly classified, it makes the diagnosis of Macular Oedema (MAOE) easier seeing that human errors have hampered the diagnostic accuracy of this debilitating condition. This study relies on statistical and Grey Level Co-occurrence Matrix (GLCM) extracted features from 6000 eye images for characterizing the risk acuity and susceptibility of the eye to MAOE. After trying different predominant algorithms used for classifying MAOE by numerous researchers, Extra Tree Classifier (ETC) was adopted for this analysis because of the better accuracy. The hyperparameters, which are features used for controlling the performance of an algorithm was tuned to determine the best set for classifying MAOE. After using Synthetic Minority Oversampling Technique (SMOTE) to balance the data and performing 10-fold cross-validation and grid-search, the optimal hyperparameter set for enhanced accuracy of MAOE classification was obtained. The optimal hyperparameter set is later used to compute the MAOE status of the eyes utilizing 75% of the dataset for training and 25% for testing. The accuracies obtained from the classification for MAOE risk acuity are {accuracy: 95.02%, AUC: 99.45%) and MAOE susceptibility {accuracy: 92.94%, AUC:97.99%}. This result makes the strategy developed to standout, hence, can improve MAOE management if implemented in an automated eye diagnostic framework.