Alzheimer's Disease Classification Using 2D Convolutional Neural Networks | JoVE Visualize

Area of Science:

Medical Imaging
Artificial Intelligence
Neuroscience

Background:

Alzheimer's disease (AD) is a prevalent, irreversible neurodegenerative disorder affecting 6% of individuals aged 65 and older.
Brain magnetic resonance imaging (MRI) is crucial for AD diagnosis, but deep learning analysis using 3D Convolutional Neural Networks (CNNs) is computationally intensive and requires large datasets.
The limitations of 3D CNNs hinder their application in medical imaging, where training data is often scarce.

Purpose of the Study:

To propose and evaluate novel methods utilizing 2D CNNs for analyzing 3D MRI data in Alzheimer's disease diagnosis.
To address the computational and data requirements challenges associated with traditional 3D CNN approaches in medical imaging.
To establish a more efficient and effective deep learning baseline for AD detection using brain MRI.

Main Methods:

Development of three distinct approaches that adapt 2D CNN architectures for processing 3D MRI data.
Testing the proposed 2D CNN methods on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.
Comparison of the 2D CNN methods against a standard ResNet-based 3D CNN model.

Main Results:

The proposed 2D CNN methods demonstrated improved performance in AD diagnosis, achieving an 8.33% increase in accuracy and a 10.11% increase in area under the receiver operating characteristic curve (auROC).
A significant reduction in training time, exceeding 89%, was observed when using the 2D CNN approaches compared to the 3D CNN model.
The study identified potential reasons for the performance enhancements and acknowledged inherent limitations.

Conclusions:

Leveraging 2D CNNs on 3D MRI data offers a computationally efficient and effective strategy for Alzheimer's disease diagnosis.
The proposed methods provide a strong foundation and a practical baseline for future research in AI-driven medical image analysis for neurodegenerative diseases.
This work highlights the potential of optimizing deep learning models to overcome data and computational barriers in clinical applications.