Self-supervised Segmentation Of Optic Cup Optic Disk

Glaucoma disease as one of the most common serious eye disease often leads to people by irreversible permanent loss of visual acuity,the modern people to effectively ensure the health and normal eye vision rapidly promote normal quality of life means that through such fundus disease danger medical screening,early diagnosis and immediate inspection and treatment of the most active and reasonable.However,traditional manual screening diagnosis is complicated,time-consuming and costly,which is completely unsuitable for rapid screening of large-scale myopia.The automatic segmentation of optic disk and optic cup can solve the time-consuming and expensive problem of manual diagnosis and evaluation.Although the current academia found there are many new approaches to the study of the optic cup optic disk segmentation and preliminary achieved relatively good technical results,some model based on deep learning method research better performance than the traditional method,but also exist model and quantity change is bigger,the graphics card request is too high,and still there is no way to the multi-parameter comprehensive research consider scale information model.This paper takes this kind of problem as the related basic research and discusses the new breakthrough point.The main work is as follows: First,an image analysis and processing method of hierarchical automatic tracking and positioning of optic disk images based on deep convolutional neural network technology is proposed.Through computer processing,the classification and detection of optic disk images based on deep convolutional network technology are transformed to obtain significant image layer feedback image information.The problem of image highlight edge damage caused by image shooting quality and other problems is reduced,and the cross interference in the measurement results of highlight edge damage of optic disk appearance image or similar image is eliminated.The automatic positioning of optic disk image greatly improves the fast quantification and accuracy of detection data.Secondly,based on the practical problems such as large scale differences in resolution between different types of fundus photographic image data,the study introduces more multi-scale information in the coding stage,and finally achieves the goal of fast and accurate recognition of target features.Thirdly,boundary module and collaborative attention mechanism are introduced to quickly complete the design of semantic segmentation network Eye Net,so as to effectively achieve the goal of precise semantic segmentation network.