Analysis And Diagnosis of Different Types of Dementia Using Neural Networks

Abstract

Dementia is the key term used to define any brain related issues that affect memory. newlineIt is caused by physical changes in the human brain. This severely affects thinking, newlineremembering activities and daily routines of a human being. In the span of the last newlinethree decades, many studies conducted on designing highly accurate systems for the newlineclassification of dementia. But this is limited to the diagnosis of a single type of dementia newline(either Alzheimer s or Parkinson s but not both). The proposed system focuses newlineon revealing the multiple dementias with a single type of dataset that is FDG-PET (Fluorodeoxyglucose newline- Positron Emission Tomography) brain scans. This research work newlineconcentrates on three models for the diagnosis of different types of dementia s namely newlineAlzheimer s disease (AD), FrontoTemporal Dementia (FTD) and Parkinson s disease newline(PD). newlineFirst model proposes a diagnostic method to classify the three dementia using an newlineArtificial Neural Network (ANN). The novelty in this work is the texture based feature newlineextraction and the selected combination of features which is not used in any existing newlineworks. The ANN used for the classification is the combination of feed forward and back newlinepropagation algorithms. The diagnostic results are generated by comparing the input newlineimage with the trained samples in the FDG-PET image database. The classification newlineincludes AD, PD and Healthy brain with an accuracy of 93.14% for modified k-means newlinesegmentation and 94% for fuzzy based gray matter segmentation. The accuracy of newlinedetection of FTD is very minimal due to data shortage which led to the proposal of newlinemodel 2. newlineSecond model proposes a Deep Neural Network (DNN) for the diagnosis of multiple newlinedementias. Since the first model is purely based on image, for better accuracy, newlinethe images were converted to values and so the second model demonstrates a statistical newlineparameter mapping method for feature extraction and Goodness based Sequential newlineFloating Forward Selection algorithm for feature selection. The classification includes newlineAD, PD and Healthy brain