Semantic Segmentation Of Macular Edema Based On U~2-Net Fundus OCT Image

With the growth of age and different lifestyles,the number of people suffering from eye diseases increases gradually.The symptoms of eye diseases directly affect people’s daily life.Optical coherence tomography(OCT)based on fundus plays a very important role in the diagnosis of retinal diseases.In order to accurately and effectively help ophthalmologists complete diagnosis and analysis,overcome the technical problems of small focus area and low contrast,improve the accuracy of macular edema segmentation and speed up the speed of model image processing,the segmentation research of OCT image of fundus macular edema in this paper mainly includes the following two parts:In view of the difficulties of small lesion area and low contrast,which seriously affect the segmentation accuracy of macular edema lesions,a two-level nested U~2-Net semantic segmentation model with U-shaped structure is proposed to segment macular edema effectively.The multi-scale features in the stage are captured by the residual U-block(RSU),and the local features and multi-scale features are fused,obtain stronger feature extraction ability of macular edema region.The improved model is applied to the public data set and compared with other methods.The experimental results show that the optimized model can accurately segment the macular edema region of the image,and improve the accuracy and accuracy of edema region segmentation to a certain extent.In order to reduce computation and reduce spatial redundancy while maintaining the segmentation accuracy,this paper proposes a macular edema segmentation of U~2-Net fundus OCT image based on attention mechanism.The attention mechanism is introduced to strengthen the feature extraction of important lesions.Octave convolution(Oct Conv)is used to replace the ordinary convolution in the residual U-block,make full use of the efficient information exchange characteristics between high frequency and low frequency of octave convolution and reducing spatial redundancy.In addition,octave convolution uses the convolution of response to process low-frequency information,effectively expand the receptive field,obtain deep pathological features and capture the multi-scale information of the image.In the public data set,this method is compared with several proposed segmentation methods in recent years.The model proposed in this paper not only obtains the dice score of0.82,which is higher than that of Re Lay Net,Res U-Net++and other methods,but also the average segmentation time of a single image is 65.1ms lower than that of U~2-Net model,which is 26.6ms less.This method can accurately and quickly segment the fundus macular edema area,and has certain auxiliary significance in clinical diagnosis.