The Research Of Retinopathy Classification And Detection Task Based On Deep Learning

In recent years,the number of patients with retinopathy in China has increased significantly.The huge demand for diagnosis and treatment has led to the data explosion in medical institutions,which makes it of great significance to quickly classify medical images to achieve efficient query,retrieval and archiving.At the same time,a large number of film reading tasks have brought huge diagnosis and treatment pressure to professional doctors.Missed and wrong examinations sometimes occur in the diagnosis process,and the uneven distribution of medical resources leads to many patients’ failure to find and treat diseases in time,resulting in irreversible damage to their vision due to missing the best diagnosis and treatment time.Therefore,there is an urgent need for computer-aided diagnosis technology to assist doctors in more accurate and efficient clinical diagnosis.Due to its strong feature extraction ability and big data processing ability,deep learning technology has achieved excellent results in the field of image processing,and a large number of research work has integrated deep learning technology with related tasks in the medical field.Therefore,this paper uses deep learning technology to combine the two application scenarios of retinal image classification and macular disease auxiliary diagnosis.According to the different needs of the two tasks,a double branch multi-scale feature fusion network and an improved YOLOV5 network are proposed to realize the classification and detection of retinopathy.Firstly,the existing retinopathy classification methods based on deep learning have weak ability to extract small target features and poor classification effect on small target lesions.To solve this problem,a double branch multi-scale feature fusion network integrating gated attention mechanism and atrous spatial pyramid pooling module is proposed in this paper.In this method,deep features are used as gating signals in the backbone network to fuse with different levels of shallow features,eliminate redundant information in shallow features and highlight the lesion area.In the branch network,the atrous spatial pyramid pooling module is used to replace the down sampling operation,which avoids the loss of small target detail features caused by down sampling.The hole convolution is used to obtain different proportions of receptive fields and rich global context information while not reducing the size of the feature map,,and the final classification accuracy is 97.9%.Second,In view of the difficulty of multi-target lesion detection caused by the lack of retinal macular lesion data set and the similarity between different types of lesions,firstly,under the guidance of professional doctors,this paper labels the location of retinopathy and creates a multi-target detection data set of retinopathy.In addition,a retinopathy detection network based on improved YOLOV5 is proposed.Firstly,an improved attention mechanism module is added at the end of the backbone feature extraction network to weight the features from the two dimensions of channel and space,highlight the lesion area and suppress irrelevant features.This improvement not only improves the ability of network feature extraction,but also enhances the ability of network to locate the lesion.In addition,the enhanced feature extraction network of YOLOV5 is improved based on the idea of Bi FPN,the backbone features are fused on the basis of PANet,and the fused features are weighted,so that the network can learn the importance of features from different levels.The experimental results show that the detection accuracy of the improved model can reach 98.1%.