Optic cup and disc segmentation based on multi-level wavelet subband-assisted learning and hybrid convolution-based dual self-attention

Abstract

Glaucoma is one of the three leading causes of blindness globally, with over 21 million patients affected in China alone. In clinical practice, accurate segmentation of the optic cup and optic disc is crucial for ophthalmologists in diagnosing glaucoma. Although significant progress has been made, especially with convolutional neural network-based methods, numerous challenges remain in scenarios such as blurred boundaries and blood vessels overlapping the boundaries. To address these issues, a novel optic cup and optic disc segmentation method based on multi-level wavelet subband-assisted learning and hybrid convolution-based dual self-attention are proposed. Firstly, a backbone network enhanced by multi-level wavelet subband auxiliary learning was designed. By introducing wavelet subband information of different directions and frequency bands at various levels of the backbone network, the network can focus more on key features relevant to the segmentation task and achieve more comprehensive feature extraction. Then, a hybrid convolution-based dual self-attention module is designed, which incorporates a ConvMixer module to extract diverse features, thereby enhancing the network's ability to adapt to different scales and shapes of the optic cups and discs. Subsequently, the features are processed by the dual self-attention module in both spatial and channel dimensions and are then reused for deeper feature extraction using different forms of convolution. Finally, feature map multiplication and skip connections are employed to suppress irrelevant features, amplify valuable ones, and refine the overall segmentation results of the optic cup and optic disc. Experimental results on the Drishti-Gs and RIM-ONE-r3 datasets show that the proposed method outperforms most current mainstream algorithms, and it achieves better segmentation results, especially along the edges of the optic cup and disc regions.