A Y-shaped spiking neural network for automatic retinal segmentation

The retinal layer contains important information for diagnosing ophthalmic diseases. Optical coherence tomography (OCT) technology allows doctors to directly observe fundus information in a non-invasive way, and the retinal layer segmentation of OCT images has always been an important task in diagnosis. Although existing deep learning-based segmentation methods have emerged in an endless stream, the segmentation methods based on convolutional neural networks (CNNs) are still limited in effect. We used deep residual spiking neural networks (SNNs) and Transformers to construct a Y-shaped network SOCT-Net for retinal OCT stratification. The network consists of a dual-path encoder and a single-path decoder, and the self-attention module is integrated into the network as the backbone of data processing. Compared with traditional convolutional neural networks, this network has a stronger retinal boundary fitting ability, especially unaffected by fundus effusion. We verified the effectiveness of this method on a public retinal OCT dataset and achieved a Dice score of 0.9086, which is better than the existing mainstream segmentation methods. The results show that the retinal automatic stratification method based on spiking neural networks has great potential in the field of OCT image segmentation.