Effects of Patchwise Sampling Strategy to Three-Dimensional Convolutional Neural Network-Based Alzheimer's Disease...

Area of Science:

• Neuroimaging analysis
• Artificial intelligence in medicine
• Machine learning for disease diagnosis

Background:

• Convolutional Neural Networks (CNNs) are increasingly used in neuroimaging.
• Three-dimensional (3D) CNNs offer advantages but face challenges like high dimensionality and overfitting.
• Patch-based CNNs improve generalization, yet optimal sampling strategies for Alzheimer's Disease (AD) classification remain unclear.

Purpose of the Study:

• To investigate the impact of patchwise sampling strategies on 3D CNN-based AD classification performance.
• To evaluate different patch partitioning methods and patch sizes.
• To analyze the interaction between patch size and training set size.

Main Methods:

• A 3D CNN framework with two-stage subnetworks was employed for AD classification.
• Patch-level subnetworks extracted features from local image patches.
• A subject-level subnetwork aggregated patch features for final classification.

Main Results:

• The 3D CNN model achieved the best AD classification performance (ACC = 89.6%) using 48x48x48 cubic image patches.
• Hippocampus-centered, region of interest (ROI)-based patches yielded suboptimal results.
• Cubic patches outperformed cuboidal patches, possibly due to AD's pathological distribution.

Conclusions:

• The optimal patch size and sampling strategy are critical for 3D CNN performance in AD classification.
• Patch size and training set size interact complexly, influencing classification accuracy.
• Findings provide guidance for developing superior 3D patch-based CNN models with appropriate sampling strategies.