ODCS-NSNP: Optic disc and cup segmentation using deep networks enhanced by nonlinear spiking neural P systems

Abstract

Accurate segmentation of the optic cup (OC) and optic disc (OD) is crucial for early glaucoma detection. Despite various deep learning-based methods for OC and OD segmentation, challenges persist, including loss of prior information, blurring from down-sampling, and limitations of simple networks. To address these issues, this paper introduces ODCS-NSNP, a deep segmentation network for OC and OD based on nonlinear spiking neural P systems. ODCS-NSNP utilizes a densely connected depth-separable network unit (SDN-Unit) to enhance network width and depth. Unlike existing approaches, it integrates a nonlinear spiking neural convolution model to redesign convolutional units and resampling operators, facilitating the fusion of multi-scale features and preserving critical details and edge structures in retinal images. Experiments conducted on the RIM-ONE-r3, Drishti-GS and REFUGE datasets evaluate ODCS-NSNP’s effectiveness. The proposed method achieved high performance across three benchmark datasets, with average Dice, IoU, and Sensitivity scores of 0.9724/0.9815 (OD/OC) in RIM-ONE-r3, 0.9625/0.9325 (OD/OC) in Drishti-GS and 0.9687/0.990 (OD/OC) in REFUGE. The experimental results highlight the strong segmentation performance of our proposed network in comparison to the leading contemporary models. Our code is available at: https://github.com/liwangcsedu/ODCS-NSNP.