Open AccessA Image Segmentation Method Based on Dual Convolutions
Author(s) -
Xiaoyu Dong,
Xichuan Hu
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 - 10.1088/1742-6596/2218/1/012014
Subject(s) - computer science , segmentation , artificial intelligence , sørensen–dice coefficient , fundus (uterus) , feature (linguistics) , pattern recognition (psychology) , diabetic retinopathy , feature extraction , computer vision , encoder , image segmentation , glaucoma , image (mathematics) , retinal , medicine , radiology , ophthalmology , diabetes mellitus , linguistics , philosophy , endocrinology , operating system
Fundus retinal vascular image has high medical diagnostic value. Its segmentation image can help assess diseases that cause retinopathy, such as cerebral small vessel disease, coronary artery disease, glaucoma, and diabetes. Classical encoder-decoder segmentation methods like Unet cannot extract subtle details as well in the retinal vessel due to its complex structure. In this paper, we propose an improved Unet algorithm based on dual convolutions and integrate MultiRes feature extraction module. The MultiRes module improves the feature learning ability of micro vessels, which can collect further subtle details. The dual convolutions module utilizes dual feature maps with varying scales and channels, exchanges feature information with each other, thus can capture retinal vasculatures with better performance. Experiment based on the DRIVE dataset indicates that the accuracy, sensitivity, specificity, AUC(area under curve), and Dice coefficient of the proposed model are 96.13%, 78.57%, 98.17%, 97.29% and 80.87%. Compared with the existing advanced algorithms, our proposed algorithm shows its effectiveness.