SU-Net: A retinal segmentation model based on improved U-Net network

• DOI: 10.1145/3584376.3584545
• Publication Date: 2023-04-19T22:54:51Z
• Authors: Mengzhu Yang; Yongfang Wang; Guoqiang Li; Lintao Zhang; Dong Zhu; Chengchao Wang
• Keywords: Sørensen–Dice coefficient; Segmentation-based object categorization; Feature (linguistics)

Image segmentation plays a very important role in medical diagnosis. It can extract information such as the area of interest, human tissue, and lesion size. Diseases nervous system, leukemia, diabetes cause eye problems. To observe changes distribution, structure, morphological characteristics blood vessels retinal images by image segmentation, it also help diagnose degree lesions above diseases to certain extent. Although commonly used artificial is gold standard, has disadvantages being time-consuming, power-consuming, unable reproduce, so research on accurate efficient automatic method focus research. Because problems, partial feature data loss, low accuracy, pathological errors that may occur traditional U-Net model during we proposed an improved based – SU-Net. In this method, attention module added coding process, which fully capture context improve accuracy extraction. The effectiveness was verified testing publicly available retina set. average IoU, Dice coefficient, global were taken evaluation indexes. Compared with model, experiments show Dice, increased 0.7, 0.9, 0.2, reached 82.4%, 82.2%, 95.5% respectively.