Multi-model Recognition Of Diabetic Retinopathy Based On Target Detection

With the recent rapid development of artificial intelligence technology,the development of high-quality model machines can well assist or even replace manual work.In the field of CV(computer vision)technology,the use of deep learning technology has brought many revolutionary breakthroughs for human beings,including face recognition,autopilot,virtual reality and so on.This paper mainly describes how to use deep learning technology to make auxiliary diagnosis on medical images,which is different from the basic image convolution processing,and proposes multi-model migration medical image recognition based on target detection.Thanks to the application of deep learning technology,many studies have been carried out on the image classification of a single model.However,it still has great limitations,such as insufficient utilization of picture information or excessive acquisition of redundant information of pictures,and large noise data,so the robustness of the training model is not enough.In a picture,useful information may be concentrated in some areas,and doctors can diagnose the problem by using only local information when diagnosing the disease.For this reason,we use target detection technology,first locate useful information on the original image to obtain the characteristics of important information,and then use the migrated multi-model integration to complete image recognition.The main research work is as follows:(1)the detection network DR-Faster-RCNN,is proposed,which is improved on the basis of Faster-RCNN to extract important feature information from the original graph.This paper also compares the single-stage SSD and YOLO methods,and analyzes the different advantages of each method in speed and accuracy.As a detection network,DRFaster-RCNN network is divided into two parts in the training and testing stages.In the training phase,two-stage training is used,while the testing only uses the RPN layer to extract important feature information.Because DR-Faster-RCNN is an improved network structure,we also study how to retrain the whole model and combine the classification network.(2)aiming at the classification design of important feature information,we construct the classification network structure of multi-model fusion judgment,using three different ResNeXt-101-32x8 d networks,DenseNet101 networks and EfficientNets B3 networks.ResNeXt-101-32x8 d network has the advantage of deep structure,DenseNet101 network has the advantage of multi-channel feature fusion,and EfficientNets B3 network verifies the balance between network depth,width and resolution to optimize the network.We also need to study how to integrate the model to achieve the best results.(3)In order to improve the performance of each model,we study a series of optimization algorithms to optimize the model.This paper studies the SWA method for the performance of this model on the data set of diabetic retinopathy.SWA is a model integration method,which uses two same models to find the optimal value on the loss function plane to improve the accuracy of the model.Secondly,the influence of ROIPooling on feature input is studied,and the way of weighted feature input is compared.It is found that the feature information can be better obtained after ROIPooling pooling.In addition,the improvement of Focal Loss loss function is studied,which is improved on the basis of CrossEntropy,adding modulation coefficient.Through the adjustment of modulation coefficient,the model can focus on the training of difficult and easy training samples.Finally,the improved method and the proposed multi-model integration method based on target detection are used to achieve 0.941 accuracy on the diabetic retinopathy data set,and the best results are achieved.