MixWT-Net: A MixStyle and Wavelet-Guided Transformer Network for Generalized Glaucoma Classification

Abstract

Early diagnosis of glaucoma is crucial to prevent permanent blindness. However, it remains challenging due to subtle structural changes in the lesion regions, overlapping features, and variation in imaging conditions. While convolutional neural networks (CNNs) achieve significant improvements for automated glaucoma screening through fundus images, they exhibit limited capability in extracting critical lesion features and correlations among them, leading to suboptimal generalization performance. To address these issues, in this paper, we propose a MixStyle and wavelet transform guided transformer network, MixWT-Net, for effective glaucoma classification. Specifically, we introduce a wavelet-based efficient self-attention module (W-ESAM) to capture relevant global contextual dependencies across the spatial and channel dimensions, thereby facilitating the capture of salient lesion cues. The W-ESAM leverages wavelet transform to obtain low-dimensional feature map while preserving contextual information and ensuring efficient attention computation. A MixStyle module is adopted to mix feature statistics and enhance adaptation and generalization ability. We perform extensive experiments on four benchmark datasets: LAG, Drishti-GS, REFUGE, and ORIGA, and the results demonstrate that our method achieves superior classification and generalization performance compared to other state-of-the-art approaches, making the MixWT-Net a method of choice for real-time glaucoma screening. © 2017 IEEE.