GLIM-Net: Chronic Glaucoma Forecast Transformer for Irregularly Sampled Sequential Fundus Images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness due to progressive optic nerve damage.
• Accurate glaucoma forecast is crucial for early detection and intervention to prevent vision loss.
• Analyzing historical fundus images is key for predicting future eye health.

Purpose of the Study:

• To propose GLIM-Net, a transformer-based model for glaucoma forecast using irregularly sampled fundus images.
• To address the challenge of irregular image sampling in capturing glaucoma progression.
• To develop a model capable of predicting future glaucoma probability, even for specific future time points.

Main Methods:

• Developed GLIM-Net, a transformer model leveraging irregularly sampled fundus images.
• Introduced novel time positional encoding and time-sensitive multi-head self-attention (MSA) modules.
• Proposed an extended model for time-conditioned glaucoma prediction.

Main Results:

• GLIM-Net achieved superior accuracy compared to state-of-the-art models on the SIGF benchmark dataset.
• Ablation experiments validated the effectiveness of the proposed time positional encoding and time-sensitive MSA modules.
• The model demonstrated strong performance in predicting future glaucoma development.

Conclusions:

• GLIM-Net offers a robust solution for glaucoma forecast with irregularly sampled data.
• The novel temporal encoding modules significantly enhance the model's ability to track glaucoma progression.
• This approach provides a valuable reference for optimizing transformer models in medical image analysis for eye diseases.