An ensemble-based 3D residual network for the classification of Alzheimer's disease

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Imaging

Background:

• Alzheimer's disease (AD) is a leading cause of dementia, with mild cognitive impairment (MCI) as a critical precursor.
• Early diagnosis of MCI is vital for managing AD progression, but differentiating MCI from normal controls (NC) and distinguishing early MCI (EMCI) from late MCI (LMCI) presents diagnostic challenges due to subtle neuroimaging variations.

Purpose of the Study:

• To develop and evaluate a novel deep learning approach for accurate Alzheimer's disease diagnosis.
• To improve the differentiation between normal cognition, MCI, and different stages of MCI (EMCI and LMCI) using neuroimaging data.

Main Methods:

• Implementation of an ensemble learning strategy integrating multiple 3D ResNet architectures (ResNet-18, ResNet-34, ResNet-50).
• Incorporation of the Convolutional Block Attention Module (CBAM) to enhance model focus on relevant image features.
• Application of data augmentation techniques to address data limitations and improve model robustness.
• Utilization of a weighted probability-based ensemble method to combine predictions from individual 3D CNN models.

Main Results:

• Achieved high diagnostic accuracy: 94.87% for MCI vs. NC, 92.31% for MCI vs. AD.
• Demonstrated strong performance in differentiating MCI stages: 95.49% for EMCI vs. LMCI.
• Obtained excellent accuracy in a four-class classification: 95.97% for NC vs. EMCI vs. LMCI vs. AD.

Conclusions:

• The proposed deep learning ensemble method effectively diagnoses Alzheimer's disease and its precursor stages.
• The integration of 3D ResNet, CBAM, and ensemble learning offers a promising approach for early and accurate AD detection using neuroimaging.