Glaucoformer: Dual-Domain Global Transformer Network for Generalized Glaucoma Stage Classification

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma stage classification from fundus images is challenging due to subtle variations and inter-stage similarities.
• Existing Convolutional Neural Networks (CNNs) capture local features but miss global dependencies.
• Vision Transformers model global context but neglect local features and frequency domain analysis.

Purpose of the Study:

• To develop a novel dual-domain global transformer network, Glaucoformer, for accurate glaucoma stage classification.
• To enhance feature representation by integrating spatial and frequency domain analysis.

Main Methods:

• Introduced a dual-domain global transformer layer (DGTL) with dual-domain channel and spatial attention.
• Integrated a Fourier domain feature analyzer (FDFA) to process frequency domain information.
• Utilized deformable convolutions to handle complex lesion irregularities and a shared key-query scheme for efficiency.

Main Results:

• Glaucoformer demonstrated superior effectiveness in classifying glaucoma stages compared to existing CNN and vision transformer models.
• The dual-domain approach successfully exploited both local and global contextual features in spatial and frequency domains.
• The model showed strong generalizability on an unseen dataset.

Conclusions:

• The proposed Glaucoformer network offers a powerful new tool for glaucoma stage classification.
• Integrating dual-domain analysis significantly improves the model's ability to learn discriminant feature representations.
• The approach holds promise for advancing automated glaucoma diagnosis.