Retinal Blood Vessel Segmentation Based on Multi-Scale and Multi-Attention Mechanisms

Aiming at the issues of subtle and complex retinal blood vessel structure, blurred boundaries, and high computational cost. To alleviate these problems, we propose UNet++ L3, which is retinal blood vessel segmentation based on multi-scale and multi-attention mechanisms. In the feature mapping of the model, the novel convolutional attention module obtains the channel spatial attention information under different fields of view by controlling the size of convolutional kernels, effectively reducing the occurrence of excessive artefacts during vessel segmentation. Besides, the global residual attention module effectively integrates global attention information with feature information, enhancing the network's ability to learn and extract features while improving its representation and generalization capabilities. Specifically, the global interpolation attention module effectively fuses the global attention information and feature information at different depths and resolutions through interpolation, improving the lost detailed information. Additionally, a new deep supervision formula is proposed to achieve the fusion of information across layers. Experiments were conducted on four common retinal datasets DRIVE, CHASE-DB1, STARE and DCA1. The F1 values were 82.34%, 80.29%, 83.65% and 77.80%. The accuracy rates were 96.81%, 97.35%, 97.36% and 97.54%. The sensitivity were 85.51%, 89.03%, 85.92% and 82.72%, AUC were 98.75%, 98.19%’ 99.08% and 99.12%, respectively. The experiments show that UNet++ L3 can effectively identify subtle, complex and fuzzy vascular structures.