Improved Performance of t-SNE Feature-Based Convolutional Neural Network for Multi-Stage Alzheimer’s Disease Detection

Abstract

Alzheimer's is one of the most common causes of dementia, causing memory loss, impaired problem-solving skills, and other cognitive impairments that can interfere with daily functioning. Early detection is critical for timely intervention and for improving the patient's quality of life. Our proposed approach, structured into three distinct phases, has the potential to improve the detection of multi-stage Alzheimer’s disease. In the first phase, to improve the quality and consistency of MRI images, various pre-processing steps are applied; furthermore, a deep synthetic oversampling technique is used to address class imbalance, thereby reducing bias and stabilizing variance across the dataset. This technique ensures a more uniform distribution of balanced samples across various Alzheimer’s classes. In the second phase, t-distributed stochastic neighbor embedding is used to map high-dimensional class-sample representations into a lower-dimensional space, enabling clear visual examination of class separability. This helps to generate well-distributed clusters. In the third phase, the proposed convolutional neural network comprises four blocks to effectively extract meaningful anatomical structural features associated with grey-white matter organization, followed by a dense layer, which is critical for detecting subtle and advanced neurodegenerative disease. As a result of this proposed model evaluation, the accuracy improved to 97.97%, with corresponding improvements in precision, recall, and F1-score.