Design And Implementation Of Algorithms For Vessel Segmentation Of Fundus Image And Joint Dlagnosis Of Cataract And DR Based On Deep Learning

Recently the number of patients with eye diseases has increased rapidly.However,the scarcity of professional doctors and the lack of resources in remote areas make people with eye diseases unable to get diagnosis in time,so missing the best treatment time.How to automate the analysis and diagnosis of eye diseases has become an urgent problem.Two important directions for studying ophthalmic diseases are vessel segmentation of fundus images and grading diagnosis of ophthalmic diseases.Therefore,we study and propose two automatic segmentation algorithms of fundus vessel structures and a multi-task learning-based automatic joint diagnosis algorithm for cataract and diabetic retinopathy in this paper.The vessel structures in fundus images can greatly reflect fundus diseases such as cataracts and diabetic retinopathy.Based on digital image processing technology and deep learning technology,we design two kinds of algorithms that can effectively perform automatic blood vessel segmentation.The segmentation algorithm based on digital image processing technology,which is called VFSeg in this paper,can obtain a clearer vessel structure through a series of processing such as image filtering,edge extraction,and image noise reduction without supervised label.U-shaped fundus vessel segmentation network based on depthwise separable convolution,which is called X-SegNet in this paper,uses encoder-decoder structure,and modules such as attention mechanism and multi-scale loss constraint.It can obtain high quality of segmentation results by learning labels.Compared with U-Net,the calculation volume and parameter volume of X-SegNet are 1/5 and 1/3 respectively.And the vessel segmentation indicators of X-SegNet on public datasets such as DRIVE,STARE,and CEASE DB1 are close to the best existing algorithms.A multi-task learning based algorithm for automatic joint diagnosis of cataract and diabetic retinopathy is proposed in this paper.By integrating cataract and diabetic retinopathy datasets,it can diagnose the two ophthalmic diseases by one network and then get their own graded results through different branches.The experimental results show that the accuracies of the diagnosis of the two fundus diseases are higher than those of the single disease diagnosis network by 2.5%and 5%,respectively.And the calculation volume of it is less than the sum of the two single disease diagnosis networks.It alleviates the problem of scarcity of single disease data in the medical field,and does not need to train multiple networks for diagnosing multiple diseases,providing a new idea for the diagnosis of multiple diseases.At the same time,the results of network visualization analysis prove the reliability of network classification.