Lightweight Fundus Image Segmentation Network Combining Structured Convolution and Dual Attention Mechanism

Automatic segmentation of fundus blood vessel images plays an important role in computer-aided diagnosis of various ophthalmic diseases. In order to solve the problems of difficulty in segmenting fundus vascular images caused by vascular scale difference and image noise, limited feature field obtained by deep learning method using single-scale convolution operation, and high complexity of existing methods, a lightweight fundus image segmentation network combining structured convolution and dual attention mechanism is proposed. Through the design of coder-decoder network with encoder enhancement, downsampling times and feature depth reduction, the lightweight network with only 0.63M parameters is realized. In the coding stage, a structured convolution method is proposed, which effectively avoids overfitting of network training and improves the ability of network to capture differentiated vascular features. In the decoding stage, a dual attention mechanism based on spatial and channel is adopted to make the network pay more attention to the context and geometric spatial information of vascular features, and suppress the interference of lesion noise. Experiments on the DRIVE, CHASE_DB1 and STARE datasets show that the accuracy of the image segmentation is 96.92%, 97.57%, and 97.51%, the sensitivity is 83.68%, 84.99%, and 84.87%, and the area under curve of the receiver operating characteristic is 98.67%, 99.05%, and 99.02%, respectively. Cross-training on the DRIVE and STARE datasets validated the generalization ability of the network.