Integrative Spectral-Graph Learning With CNN Features for the Classification of Circle of Willis Anatomical Variants

Area of Science:

• Medical Imaging
• Neuroscience
• Computer Science

Background:

• Cerebrovascular diseases are linked to morphological variations in the Circle of Willis (CoW).
• Early detection of CoW abnormalities is crucial for effective treatment and disease prevention.
• Automated identification of CoW variants requires robust computer-aided models.

Purpose of the Study:

• To develop a computer-aided model for automatic identification of Circle of Willis anatomical variants.
• To apply the Lippert and Pabst classification system in a computer-assisted analysis context.
• To address challenges posed by small, imbalanced medical datasets for Convolutional Neural Networks (CNNs).

Main Methods:

• Developed a novel graph-based method integrating spectral analysis with a hybrid CNN and Graph Neural Network (GNN) architecture.
• Utilized the Lippert and Pabst classification system for CoW variants.
• Compared performance across various VGG and ResNet network configurations.

Main Results:

• The proposed graph-based hybrid method achieved a balanced accuracy of 0.69 for anterior CoW variants.
• The framework attained a balanced accuracy of 0.71 for posterior CoW variants.
• Demonstrated significant improvement in CoW classification performance for both anterior and posterior classes.

Conclusions:

• The novel graph-based hybrid CNN-GNN framework effectively captures complex CoW morphological structures.
• The proposed method offers a significant advancement in computer-assisted analysis of Circle of Willis variants.
• This approach enhances the potential for early detection and management of cerebrovascular diseases.