Detail enhanced retinal vessel segmentation algorithm based on wavelet feature and U-Net
To address challenges such as feature loss and difficulty in extracting fine blood vessels in the task of retinal vessel segmentation,a detail enhanced algorithm for retinal vessel segmentation based on wavelet feature and U-Net is proposed.Firstly,to mitigate the loss of vascular features during the downsampling process,a wavelet feature compensation module is designed.This module employs wavelet analysis to independently model the high and low-frequency information in retinal images,thereby delivering a more enriched set of multi-frequency features.Secondly,considering the challenge of low contrast in retinal images and the difficulty in discerning fine blood vessels,a multi-dimension attention module is proposed to enhance the feature disparity between vessels and background noise.By integrating semantic information at different depths,this module conducts vessel feature detection and exploration along horizontal,vertical and channel dimensions.This approach achieves an organic fusion of spatial and channel information,thereby strengthening the algorithm's capability for detailed processing.Finally,to address discontinuities in vascular representation caused by insufficient algorithmic field of view,a connection enhancement module is designed.This module combines outputs from various layers of the decoder,utilizes dilated convolutions to expand the algorithm's receptive field,and captures multi-scale context information to enhance vascular connectivity.The algorithm's effectiveness is tested on the DRIVE and CHASE_DB1 datasets,yielding sensitivity values of 0.837 6 and 0.845 3,F1 scores of 0.834 7 and 0.837 2,and AUC values of 0.988 6 and 0.990 1,respectively.
万方数据
维普
