Siamese Capsule Network (SNNCap): Cognitive Analysis for Alzheimer's Disease Classification From MRI Data

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Alzheimer's Disease (AD) detection is crucial for effective patient care and treatment.
• Magnetic Resonance Imaging (MRI) provides high-resolution brain structure data vital for diagnosing neurological disorders.
• Existing deep learning methods for AD detection using MRI, such as Siamese Convolutional Neural Networks (SCNN) with ResNet-34, face challenges in preserving spatial information due to pooling operations.

Purpose of the Study:

• To develop a cognitively inspired deep learning approach for accurate classification of Alzheimer's Disease severity from MRI scans.
• To address limitations of previous methods in capturing spatial relationships and part-whole hierarchies within brain MRI data.
• To improve the generalizability and accuracy of automated AD detection systems.

Main Methods:

• Proposed a Siamese Capsule Network (SNNCap) model for classifying MRI images into dementia severity categories: Non Demented, Very Mild Demented, Mild Demented, and Moderate Demented.
• Utilized ResNet-18 for feature extraction and capsule layers within SNNCap to preserve crucial spatial and part-whole relationships.
• Implemented a reference-based validation strategy where test images are compared against known examples, mimicking human cognitive reasoning.

Main Results:

• The SNNCap model demonstrated strong performance in classifying unseen MRI data.
• The system effectively preserved spatial and part-whole relationships, overcoming limitations of pooling operations in traditional CNNs.
• Classification reports and confusion matrices confirmed the model's effectiveness in differentiating dementia severity levels.

Conclusions:

• The proposed SNNCap offers a promising cognitively inspired approach for Alzheimer's Disease detection using MRI.
• Preserving spatial and part-whole relationships through capsule networks enhances diagnostic accuracy and generalizability.
• This method represents an advancement in automated analysis of neuroimaging data for neurological disorder diagnosis.