EFFECTIVENESS OF LEARNING RATE IN DEMENTIA SEVERITY PREDICTION USING VGG16

Abstract

Alzheimer’s disease (AD) is the leading worldwide cause of dementia. It is a common brain disorder that significantly impacts daily life and slowly progresses from moderate to severe. Due to inaccuracy, lack of sensitivity, and imprecision, existing classification techniques are not yet a standard clinical approach. This paper proposes utilizing the Convolutional Neural Network (CNN) architecture to classify AD based on MRI images. Our primary objective is to use the capabilities of pre-trained CNNs to classify and predict dementia severity and to serve as an effective decision support system for physicians in predicting the severity of AD based on the degree of dementia. The standard Kaggle dataset is used to train and evaluate the classification model of dementia. Synthetic Minority Oversampling Technique (SMOTE) tackles the primary problem with the dataset, which is a disparity across classes. VGGNet16 with ReduceLROnPlateau is fine-tuned and assessed using testing data consisting of four stages of dementia and achieves an overall accuracy of 98.61% and a specificity of 99% for a multiclass classification, which is superior to current approaches. By selecting appropriate Initial Learning Rate (ILR) and scheduling it during the training phase, the proposed method has the benefit of causing the model to converge on local optimums with better performance.