Research On Fundus Retinal Image Segmentation Method

With the development of socio-economic science in China,significant progress has been made in disease prevention.The eye,as an important way to explore the world,is of great significance to humanity.The rapid development of deep learning has made assisted diagnosis a hot topic in the field of medical image processing.Color fundus retinal images are captured by a color fundus camera and can be used to judge whether a person has diseases such as diabetes and glaucoma based on the structure of the fundus.Because the various tissue structures in the retinal image are relatively complex,there are many capillaries and the optic disc features are not obvious.If doctors rely solely on manual segmentation of blood vessels and optic discs,it not only consumes time and effort but also has the risk of misdiagnosis and missed diagnosis.In the task of blood vessel segmentation,preprocessing was conducted to address the problem of unclear blood vessel features.The green channel of the image was extracted,and contrast-limited adaptive histogram equalization and gamma correction were performed to enhance the blood vessel features.Based on this,the data was augmented by image rotation and expanding the original image with the label map to meet the needs of network training.To address the issue of low accuracy in segmenting blood vessel endings and breaks,an improved U2-Net blood vessel segmentation network was constructed by combining the structural characteristics of the U2-Net network with the introduction of a deep attention module.Furthermore,a multi-scale feature aggregation module was introduced to enrich the blood vessel features extracted by the network and increase the edge detail information of the final feature map.In the task of optic disc segmentation,preprocessing was conducted to address the problem of unclear optic disc features.The red channel of the image was extracted,and homomorphic filtering was performed to enhance the optic disc features.Data augmentation was conducted by rotating the image.To address the problem of weak learning and feature extraction capabilities of the U-Net for optic disc features,an improved U-Net optic disc segmentation network was constructed by fusing the channel attention mechanism and the spatial attention mechanism to increase the weight of the optic disc in its surrounding pixels and increase the input quantity of effective features.Additionally,a dual-path feature extraction module was introduced to enhance the network’s ability to extract optic disc features.Experimental results showed that high accuracy segmentation of the retinal blood vessels and optic discs was achieved,with an accuracy rate of 96.98% for blood vessel segmentation,which is 8.8% higher than the basic model,and an accuracy rate of 98.77% for optic disc segmentation,which is 6.54% higher than the original basic model.