Combined Multi-Atlas and Multi-Layer Perception for Alzheimer's Disease Classification

Area of Science:

• Neuroscience
• Medical Imaging
• Computer Vision

Background:

• Alzheimer's disease (AD) is a progressive neurodegenerative disorder.
• Accurate classification of AD stages is crucial for patient management.
• Existing methods using Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) have limitations with discrete morphological features.

Purpose of the Study:

• To develop an effective model for Alzheimer's disease (AD) classification using multi-atlas morphological features.
• To address the challenge of classifying discrete feature values derived from MRI data.
• To improve upon traditional CNN and RNN approaches for AD staging.

Main Methods:

• A Multi-Atlas Multi-Layer Perceptron (MAMLP) model was proposed.
• Representative atlases were selected to preserve brain feature diversity.
• Individual Multi-Layer Perceptron (MLP) modules processed features from each atlas.
• A voting system combined results from multiple MLPs for final classification.

Main Results:

• The MAMLP approach ranked 10th out of 373 teams in the PRCV 2021 challenge.
• Atlas group selection reduced feature requirements without compromising accuracy.
• MLP architecture outperformed CNN and RNN models for morphological features.
• The combined MLP network demonstrated approximately 40% faster convergence and enhanced classification stability.

Conclusions:

• The MAMLP model offers a robust and efficient method for Alzheimer's disease (AD) classification.
• MLP networks are well-suited for analyzing discrete morphological features in neuroimaging.
• Combining multiple MLPs enhances classification stability and speeds up convergence, outperforming traditional methods.