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

What is it about?

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

Featured Image

Photo by Harpreet Singh on Unsplash

Photo by Harpreet Singh on Unsplash

Why is it important?

In this paper, we propose a new method of retinal image segmentation based on U-Net. The method adds the feature extraction attention module SEM in the down-sampling process, which can better extract features and retain information, and select the appropriate attention stimulation operator according to the current network, which further improves the training speed and accuracy of the network. The SU-Net proposed in this paper solves the problem of low precision of retinal segmentation caused by the loss of retinal feature data, which is of great significance to the research of retinal segmentation. The experiment was carried out on the public dataset, and the three commonly used evaluation indexes were used for evaluation. The results show that the proposed method has obvious improvement compared with the traditional U-Net segmentation method. In future studies, we will extend to different types of data sets to verify the generality of SU-Net and whether it can be used for other image processing tasks, and test on large data sets to verify whether the model has a good performance.

Perspectives