Design And Implementation Of Interactive Fundus Labeling System Based On Deep Learning

Glaucoma is a fundus disease that causes vision loss due to optic nerve damage.After catarac,it is the second most common blind eye disease in the world,and the leading cause of blindness.Because vision damage caused by glaucoma is irreversible,early screening and diagnosis are critical to maintaining vision.Clinical glaucoma screening can be obtained by fundus photography to obtain a retinal fundus image,and glaucoma risk is assessed based on the ratio of the diameter of the optic cup(OC)to the optic disc(OD)in the vertical direction(CDR)in the fundus image.Much work has been done to obtain CDR values by automatically segmenting the optic disc.Accurate cup disc splitting is the key to accurately calculating CDR values.The improvement of the segmentation accuracy of the optic disc requires a large amount of data with segmentation and labeling.At present,the data is manually labeled by the doctor,which will consume a lot of manpower and time.In order to reduce the doctor’s workload and improve the credibility of the annotation data,this paper proposes an interactive fundus iamge based on deep learning.The doctor only needs to make a simple annotation on the fundus image to get precise label on the cup and the optic disc.This article first introduces the interactive fundus image annotation scheme,including detailed requirements analysis of each module of the annotation system,the overall design of the system,and the convolutional neural network model used by the system.Then introduce the coding implementation of each functional module of the system and the implementation of the training code of the convolutional neural network model.Since the annotation is performed on the PC side,the system uses the OkHttp network communication technology to solve the data interaction and sharing problem between the PC and the algorithm,and the system also realizes the manual labeling when the algorithm result is not good.Finally,the convolutional neural network model is evaluated through experiments,and the interactive fundus label system is tested and analyzed,and some shortcomings of the current system are summarized,and the future work is expected.