Multi-resolution 3D-HOG feature learning method for Alzheimer's Disease diagnosis.

Area of Science:

• Neuroimaging analysis
• Medical image processing
• Biomedical engineering

Background:

• Alzheimer's Disease (AD) is a progressive neurodegenerative disorder.
• Timely identification of AD is critical for managing disease progression.
• Current diagnostic methods require improvement for early and explainable detection.

Purpose of the Study:

• To develop a discriminative feature extraction and selection strategy for explainable AD identification.
• To propose a spatial pyramid based 3D-HOG (SPHOG) feature learning method.
• To enhance the accuracy and interpretability of AD diagnosis.

Main Methods:

• Utilized spatial pyramid based 3D-HOG (SPHOG) for feature extraction, capturing global and local texture changes.
• Employed a modified wrapper-based feature selection algorithm to identify discriminative features and reduce dimensionality.
• Validated the approach on both common and clinical datasets.

Main Results:

• Selected discriminative SPHOG histograms at various resolutions effectively represent cerebral cortex atrophy.
• The identified subareas align with clinical expertise, emphasizing the method's explainability.
• Demonstrated promising performance in AD identification using the proposed feature learning strategy.

Conclusions:

• The proposed SPHOG feature learning method is effective for AD identification.
• The method's explainability is highlighted, particularly with reference to the Hippocampus.
• This approach holds potential for early AD diagnosis and further medical analysis of other brain ROIs.