DMSU-Net++: A dual multiscale retinal vessel segmentation method based on improved U-Net++

Abstract

Retinal blood vessels are of different sizes and shapes, and even contain very fine capillaries with complex structural morphology, making accurate segmentation a difficult task. To address the above problems, we propose an improved retinal segmentation method DMSU-Net++ (Double Multiscale U-Net++) based on U-Net++. The method innovatively introduces a multiscale feature extraction module WTSAFM, which realises multiscale feature extraction via wavelet transform, and can capture image information of different frequencies more effectively while enhancing global context understanding. In addition, a dual multi-scale feature extraction module is constructed by cascading MFE modules to compensate for the spatial lack of information in WTSAFM, which further improves the accuracy of the model in dealing with different scale information. This method is experimented on the proposed method on two publicly available retinal vessel segmentation datasets, DRIVE and CHASE-DB1, and the experiments show that the F1-score of this method on the two datasets is 82.75 and 82.81, the Sensitivity is 83.74 and 85, and the AUC is 97.86 and 98.36, respectively. Compared with other methods, the method shows better segmentation performance with better accurate recognition.