Enhancing magnetic resonance imaging-driven Alzheimer's disease classification performance using generative...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Generative adversarial networks (GAN) show potential for improving image quality.
• The application of GANs to augment image-based classification, particularly in neuroimaging for Alzheimer's disease (AD), is not fully explored.
• Magnetic resonance imaging (MRI) at different field strengths presents unique characteristics for diagnostic analysis.

Purpose of the Study:

• To evaluate if a modified GAN can enhance Alzheimer's disease (AD) classification performance.
• To investigate the ability of GANs to learn from multi-field strength MRI scans.
• To improve the quality of MRI scans for better diagnostic outcomes.

Main Methods:

• T1-weighted brain MRI scans from 151 ADNI participants (1.5-T and 3-T) were used to construct a GAN model.
• A three-dimensional fully convolutional network (FCN) was trained using GAN-generated 3-T images (3T*) to predict AD status.
• Image quality was assessed using signal-to-noise ratio (SNR), BRISQUE, and NIQE; validation was performed on AIBL and NACC datasets.

Main Results:

• The 3T*-based FCN classifier demonstrated improved performance compared to the 1.5-T FCN model across ADNI, AIBL, and NACC datasets.
• Mean area under the curve (AUC) increased from 0.907 to 0.932 (ADNI), 0.934 to 0.940 (AIBL), and 0.870 to 0.907 (NACC).
• Generated 3T* images showed consistently higher quality than 1.5-T images based on SNR, BRISQUE, and NIQE metrics.

Conclusions:

• GAN frameworks can be constructed to augment AD classification performance.
• This study provides a proof of principle for using GANs to improve MRI image quality for AD diagnosis.
• The findings suggest GANs are a promising tool for enhancing neuroimaging-based disease classification.