A novel discriminant feature selection–based mutual information extraction from MR brain images for Alzheimer's stages detection and prediction. Issue 4 (16th December 2021)

• Record Type: Journal Article
• Authors: Shankar, Venkatesh Gauri; Sisodia, Dilip Singh; Chandrakar, Preeti
• Is Part Of: International journal of imaging systems and technology. Volume 32:Issue 4(2022)
• Page Start: 1172
• Page End: 1191
• Publication Date: 2021-12-16
• Subjects: Alzheimer's disease -- classification -- feature selection -- machine learning -- medical imaging system -- neurodegenerative disorder; Imaging systems -- Periodicals; Image processing -- Periodicals; 621.367
• DOI: 10.1002/ima.22685
• Languages: English
• ISSNs: 0899-9457

Abstract

Abstract: Alzheimer's disease (AD), a neurodegenerative disorder, is a form of dementia. Quick or early diagnosis of AD is essential, but most of the available studies have focused on clinical or survey‐based data, leading to data inconsistency as many people feel hesitant or hide information due to disease and societal stigma. Nowadays, current computer‐aided support and techniques are mainly based on feature extraction, but due to redundant or similarly extracted features, any specific model is not producing best performance. This paper proposes a novel mutual relationship–based feature selection model with high‐altitude acute response like features for the brain‐predefined feature areas using magnetic resonance imaging. The mutual relationship between features is used to avoid similar features or create discrimination between different features due to which accuracy gets affected. In this work, four machine learning classifiers, namely decision tree, support vector machine (SVM), k‐nearest neighbor, and Naive Bayes are used for developing predictive model and algorithms. To avoid overfitting in model selection, we have used a 10‐fold cross‐validation for model training and testing. We collected data set from Alzheimer's disease Neuroimaging Initiative repository. The performance of supervised classifiers was estimated with area under the receiver operating characteristic curve, accuracy, precision, recall, and F1 score assessment measures. The results indicate SVM to be the best performing classifier for prediction (SVM recording the highest average area under curve score of 0.936, accuracy of 94.1%, precision of 96.9%, recall of 96.6%, and F1‐score of 96.8%) with 37 discriminant features.