Research On Deep Learning Based Fundus Medical Image Segmentation Method

The accuracy of the segmentation results of fundus medical images affects the clinical analysis of doctors.The feature extraction method based on deep learning can effectively solve the problems of difficult traditional feature extraction and rough segmentation results.Therefore,the automatic segmentation model of fundus images based on deep learning is a hot topic in current medical image research.In this paper,we take the optic cup optic disc part and retinal vessel part of fundus images as the research content,carry out improved segmentation research based on U-Net network according to the characteristics of each part,and realize a new automatic segmentation network about optic cup optic disc images and retinal vessel images.The main research elements of this paper are as follows:（1）To address the problem that some existing segmentation network models for fundus images regarding optic cup optic disc images have the prediction of aberrant segmentation maps,based on IOSUDA（Input and Output Space Unsupervised Domain Adaptation）This paper proposes an unsupervised domain-adaptive optic cup optic disc image segmentation network based on effective shape constraints and triple attention:SCUDA（Shape-Constrained Unsupervised Domain Adaptation）.This network model utilizes a priori knowledge of the optic cup optic disc structure of fundus images and proposes a loss function based on a circular-like shape constraint,aiming at constraining optic cup optic disc image generation during segmentation.Secondly,for the original segmentation network U-Net,which is a classical network for segmenting medical images but suffers from inefficiency,a convolutional triple attention module is designed at the codec to improve the segmentation network,which is capable of capturing interactions across dimensions and providing a rich feature representation to improve segmentation accuracy.Extensive experiments have shown that the proposed SCUDA achieves good segmentation results compared to some other existing state-of-the-art network models on two publicly available datasets of optic cup optic disc images,Drishti-GS and RIM-ONE＿R3,where4)（80）values of 78.98%and 74.05%were obtained,respectively.（2）With regard to retinal vascular images in fundus images,which suffer from problems such as having complex vascular structures and uniformly positioned without distribution as well as unclear edge contours,this paper constructs a retinal vascular image segmentation network model based on criss-crossed hybrid-attention with multilayer global pyramid guidance:CG-CENet.The model fuses global and contextual information by using four pyramidal modules with different expansion rates,with the aim of providing information flow guidance for the feature decoding module to recover features.In addition,a criss-crossed hybrid-attention module is designed behind each decoder layer,which contains both spatial and channel dimensions and is able to capture rich contextual semantic information,thus allowing for better feature recovery representation for the decoder.Experiments were conducted on two publicly available retinal vascular image datasets,DRIVE and CHASE-DB1,with accuracies of 96.22%and 96.86%respectively,achieving the best results compared to other segmentation models.In addition,to validate the generalization performance of the model,generalization validation experiments were performed on the diabetic retinopathy dataset DIARETDB1（exudates and haemorrhages）.