Research And Implementation Of Semantic Retinal Image Semantic Segmentation Based On Semi-supervised Transfer Learning

The research of fundus image analysis and processing technology has very important clinical guiding significance for the diagnosis of various ophthalmic diseases and other diseases of the human body.Therefore,it is of great significance to perform a classification of the fundus image on the diabetic retinopathy image and to accurately segment the blood vessels of one of the important components of the fundus image anatomy based on the lesion level.In this thesis,SqueezeNet v1.1 was used to classify the fundus image,and the improved FCN and semi-supervised migration and U-dimensional method were used to achieve segmentation of blood vessels.The specific work is as follows:(1)A classification method for fundus images of diabetic retinopathy based on ultra-lightweight SqueezeNet network was proposed.Firstly,a series of pretreatments are performed on the fundus image,and the SLO(Scanning Laser Ophthalmoscope)image is adjusted and converted into a pseudo color image,followed by data amplification to solve the class imbalance problem.Secondly,according to the four types of diabetic retina Based on the characteristics of the lesion image,an improved ultra-lightweight SqueezeNet v1.1 deep convolutional neural network was to achieve accurate classification of pathological images.(2)Aiming at the problem of insufficient medical labeled image data,a full-convolution neural network blood vessel segmentation method based on transfer learning is proposed.First,the DRIVE data set after data amplification is input;secondly,using the parameter transfer method,the first five sets of parameters(a total of 13 layers of convolutional layers)of VGG19 trained on ImageNet are transferred to the full convolutional neural network to complete the pre-training.It has been proved by experiments that the jump structure with parameter transfer is adopted between FCN32s,FCN16s and FCN8s,and the accuracy is obviously improved compared with the non-parameter transfer,about 6.5%.(3)Aiming at the problem of small blood vessel segmentation and data scarcity,an iterative semi-supervised transfer learning semantic segmentation method is proposed.Firstly,the method of self-data amplification and cropping is proposed.The small blood vessels are used as the main body of segmentation.In the overlapping splicing method,the relationship between pixels is used to improve the fault tolerance rate,and the segmentation result is corrected.Secondly,the data is scarce and the lesions are affected.Transfer learning and feature learning propose an iterative semi-supervised migration learning method,which performs iterative transfer through simplified dimension reduction Unet network learning features,and continuously promotes the fusion of source domain and target domain,and achieves reasonable amplification of dataset.The experimental results show that the proposed algorithm can segment small blood vessels better,and the accuracy rate reaches 97%,which exceeds the accuracy rate of human experts.