Classifying retinal diseases via pyramid vision graph convolutional network for optical coherence tomography images

Area of Science:

• Medical imaging analysis
• Deep learning for ophthalmology

Background:

• Deep neural networks (DNNs) are widely used for retinal disease classification in optical coherence tomography (OCT) images.
• Traditional DNNs struggle with the irregular structures in OCT images due to their grid-based processing, leading to suboptimal performance.

Purpose of the Study:

• To introduce a novel visual neural network model, the pyramid vision graph convolutional network (PVGCN), for improved retinal disease classification in OCT images.
• To address the limitations of traditional DNNs in handling complex retinal structures.

Main Methods:

• The PVGCN model utilizes a vision graph block to segment images into nodes, representing them as graph data for better structural correlation.
• A pyramid structure is employed to integrate multi-scale features, capturing hierarchical information and reducing model parameters.
• Graph convolution and feed-forward networks are used to model relationships between image regions and mitigate over-smoothing.

Main Results:

• The PVGCN model achieved high accuracies of 0.9954 and 0.9787 on two independent datasets.
• The model demonstrated superior performance compared to existing methods for retinal disease classification.
• The PVGCN's diagnostic capabilities were found to be comparable to those of human experts.

Conclusions:

• The proposed PVGCN model offers a flexible and effective approach for classifying retinal diseases from OCT images.
• PVGCN's graph-based and multi-scale feature integration overcomes limitations of traditional DNNs, enhancing diagnostic accuracy.
• This novel architecture holds significant promise for advancing automated detection of eye diseases.