Research On Image Classification Of Diabetic Retinopathy Based On Shallow Neural Network

Diabetic Retinopathy(DR)is the main cause of adult visual impairment.Early detection and treatment are very important to reduce the risk of blindness.The method based on deep Convolution Neural Network(CNN)can help to detect DR through the classification of retinal fundus images of patients.Such methods usually need a large labeled dataset to support model training,but in practical applications,the acquisition of high-quality and large-scale labeled datasets often faces big challenges.At the same time,the training of deep CNN model is not only time-consuming,but also easy to drop into overfitting.Therefore,it is meaningful to explore an CNN based approach to classifying retinal fundus images that is effective on small datasets.The main studies are as follows.(1)An approach to retinal fundus vessel segmentation based on wavelet transform is proposed.Since the analysis of retinal fundus blood vessels is very important for DR diagnosing,a fundus blood vessel segmentation approach based on wavelet transform decomposition and reconstruction is discussed,which takes the characteristics of retina fundus images,such as smooth edge,uneven gray scale and complex distribution of blood vessel structure,into consideration.The experimental results show that,compared with others,the proposed approach has better segmentation effect,and the segmented fundus blood vessels have higher integrity,clear structure,and less loss of small blood vessels.(2)An integrated approach to the classification and detection of DR based on multiscale shallow CNN is proposed.Experiments show that on the original dataset,the classification accuracy of the proposed approach is improved by 2%-9% compared with LCNN,VGG16 no FC,Compact Net and the current representative CNN integration approaches.On two small sample datasets,the classification accuracy is improved by 3.5%and 4.7% on average compared the integration approaches,Conv Net and CNN.In addition,when using only 10% sample of the original dataset,the classification accuracy of the prosed approach is 3% higher than that of VGG16 no FC on the original dataset.Meanwhile,though the accuracy of the proposed approach is close to the average accuracies of the three approaches,its time cost is largely lower.