Analysis And Application Of OCT Image Retinopathy Based On Depth Detection Network

Choroidal neovascularization(CNV)is an important characterization of age-related macular degeneration(AMD).Effective CNV detection in spectral-domain optical coherence tomography(SD-OCT)images can vastly assist CNV clinical diagnosis.With the rapid development of deep learning,great progress has been made in the field of object detection.This paper aims at the problem of applying existing object detection models to medical images,and combines the characteristics of SD-OCT images and CNV lesions to carry out research and applications for the efficient and accurate detection of CNV using deep learning.The specific work is as follows:(1)A CNV detection model combining traditional features and nearby propagation algorithms is proposed.On the basis of YOLOv3,the traditional feature fusion original image is introduced to form a network input,which enhances the effective information of SD-OCT images while solving the problem of mismatch between natural images and medical image channels.According to the clustering accuracy variation curve of CNV lesions,the number of clusters in the anchors was adjusted.Proposed the nearby propagation algorithm,which reduces the detection loss rate and improves the consistency of the lesion.Use self-refine strategy to train the model and strengthen the ability to detect difficult samples by re-learning high-confidence test samples.In a comparison experiment with art of the stage object detection models,indicators such as average accuracy,recall rate,and missing rate are compared to demonstrate the effectiveness of our proposed method.(2)Based on the angle-independent regression method,a CNV detection model for predicting inclined bounding boxes is proposed.Based on the previous model,the direction of the bounding box is predicted by introducing independent regressions at the output layer.This solves the problems of using a horizontal bounding box to predict long and oblique lesions,predicting a large bounding box,resulting in a small proportion of lesions and inaccurate diagnosis.In order to apply to the rotated bounding box,we modified the model input,loss function,and non-maximum suppression algorithm,and proposed a new Io U calculation method.Compared with the object detection models of the prediction normal bounding box,the experimental results are qualitatively analyzed and quantitatively evaluated,and the advantages of the proposed algorithm are systematically discussed.(3)Based on the CNV detection model proposed in this paper,a CNV detection system was developed.The encapsulation technology is used to encapsulate the CNV detection model,and adjustable application program interfaces are provided.Through the modular design,the system functions are structurally divided,which simplifies the detection process and facilitates clinical and research use.Using a custom configuration file,the parameter adjustment in the interface is effectively passed to the background algorithm,and the parameter configuration is reused.Using a packaging tool,the system was released for easy system transmission.