SWAU-Net: Longitudinal Prediction of Geographic Atrophy via Sliding-Window Attention

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Age-related macular degeneration (AMD) is a primary cause of vision loss in older adults.
• Geographic atrophy (GA), the advanced form of AMD, presents challenges in predicting its progression due to variable growth rates and limited data.
• Accurate GA progression prediction is crucial for clinical practice and designing effective clinical trials.

Purpose of the Study:

• To introduce a novel hybrid AI model, Sliding Window Attention U-Net (SWAU-Net), for predicting the longitudinal progression of geographic atrophy (GA).
• To address the challenges of variable growth rates and data scarcity in forecasting GA trajectories.
• To improve the accuracy and reliability of GA progression prediction for clinical applications.

Main Methods:

• Developed SWAU-Net, a hybrid architecture combining Transformer-based temporal modeling with U-Net convolutional neural network (CNN) for spatial modeling.
• Integrated explicit temporal and geometric consistency priors through a weight-shared Sliding Window Attention core.
• Employed feature-level regularization to preserve lesion boundaries and ensure generalization in low-data scenarios.

Main Results:

• SWAU-Net achieved a Growth Mask Dice Similarity Coefficient (DSC) of 0.66, indicating significant improvement in predicting GA lesion expansion.
• The model demonstrated superior performance compared to unregularized Transformer and standard recurrent baseline models.
• Structural constraints in SWAU-Net effectively prevented overfitting to imaging noise.

Conclusions:

• SWAU-Net offers a robust framework for predicting GA lesion trajectories, outperforming existing methods.
• The model's ability to generalize in low-data regimes enhances its clinical utility.
• This advancement holds potential for optimizing clinical trial designs and enabling personalized patient monitoring for AMD.