Open AccessDM-Net:a Depth-separable convolution and Multi-Scale Network for retinal blood vessel segmentation
Author(s) -
Wei Li,
Wei Cong,
Yuqing Cheng,
Yang Liu
Publication year - 2022
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/2213/1/012040
Subject(s) - convolution (computer science) , segmentation , computer science , residual , separable space , feature (linguistics) , artificial intelligence , net (polyhedron) , feature extraction , pattern recognition (psychology) , scale (ratio) , set (abstract data type) , channel (broadcasting) , convergence (economics) , field (mathematics) , artificial neural network , algorithm , mathematics , telecommunications , philosophy , physics , geometry , quantum mechanics , pure mathematics , economics , programming language , economic growth , mathematical analysis , linguistics
Retina segmentation plays an important role in the medical field,In recent years, some proposed networks have some problems, such as single receptive field, huge parameters, and difficulty in training, which affect the segmentation results. In this paper, a U-Net-based DM-Net with deep separable convolution and multi-scale is proposed, a residual multiscale module is designed to reduce the parameters and improve the feature extraction ability. In order to cope with the feature information fusion at different levels and the sudden decrease in the number of feature channels in the decoder, the channel attention mechanism is applied. Experiments on the public data set CHASE_DB 1 show that DM-Net has achieved good results compared with other networks, especially in ACC (0.9748) and SP (0.9882). At the same time, it has few parameters and fast convergence speed.