Exceptional performance with minimal data using a generative adversarial network for alzheimer's disease...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Neuroscience

Background:

• Deep learning for Alzheimer's disease (AD) classification is limited by scarce medical imaging data due to privacy regulations.
• Open-access datasets like OASIS often present imbalanced class distributions, further complicating model training.

Purpose of the Study:

• To propose and evaluate a generative adversarial network (GAN) integrated approach for enhancing Alzheimer's disease classification with limited data.
• To address data scarcity and class imbalance issues in medical image datasets for AD research.

Main Methods:

• Utilized a generative adversarial network (GAN) for data augmentation, generating synthetic MRI data.
• Trained the GAN model using experimental data from the Open Access Series of Imaging Studies (OASIS) database.
• Integrated synthetic data into a pretrained convolutional neural network (CNN) for multistage AD classification.

Main Results:

• Achieved a multistage Alzheimer's disease classification accuracy exceeding 80% with a reduced dataset.
• Demonstrated the effectiveness of GANs in overcoming data scarcity for medical image analysis.
• Showcased comparable accuracy to models trained on larger, more balanced datasets.

Conclusions:

• Generative adversarial networks (GANs) offer a viable solution to the challenge of insufficient data in Alzheimer's disease classification.
• GAN-based data augmentation can significantly improve deep learning model performance in medical imaging research.
• This approach holds promise for advancing AD diagnosis and research despite data access limitations.