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

## TL;DR

A new deep learning model called SWAU-Net improves the prediction of geographic atrophy progression in age-related macular degeneration, helping with clinical trials and patient monitoring.

## Contribution

SWAU-Net introduces a novel hybrid architecture combining Transformer-based temporal modeling with U-Net spatial modeling, enhanced by consistency priors to improve prediction accuracy in low-data scenarios.

## Key findings

- SWAU-Net achieved a Growth Mask Dice Similarity Coefficient (DSC) of 0.66, outperforming baseline models.
- The model's structural constraints prevent overfitting to imaging noise, improving prediction robustness.
- The framework supports more efficient clinical trial designs and personalized patient monitoring for geographic atrophy.

## Abstract

Age-related macular degeneration (AMD) is the leading cause of central vision loss in aging populations. Geographic atrophy (GA) is the advanced, non-neovascular form of AMD. Predicting the longitudinal progression of GA remains a critical challenge in ophthalmic clinical practice and clinical trial design. Forecasting the trajectory of GA is complicated by highly variable growth rates and the inherent scarcity of long-term, high-quality imaging data. To address these challenges, we introduce the Sliding Window Attention U-Net (SWAU-Net), a hybrid architecture that integrates Transformer-based temporal modeling of GA growth with precise spatial modeling of GA location with a U-Net convolutional neural network (CNN). To ensure generalization in the low-data regime, SWAU-Net embeds explicit temporal and geometric consistency priors via a weight-shared Sliding Window Attention core and feature-level regularization that preserves sparse, high-frequency lesion boundaries across frames. Experimental results demonstrate that these structural constraints prevent the model from overfitting to imaging noise, achieving a Growth Mask Dice Similarity Coefficient (DSC) of 0.66 (representing the spatial overlap between the predicted and ground truth lesion expansion regions), a significant improvement over unregularized Transformer and standard recurrent baseline models. Our framework provides a robust tool for predicting GA lesion trajectories, potentially supporting more efficient clinical trial designs and personalized patient monitoring.