Object Detection In Fundus Image Based On Deep Learning

The early symptoms of diabetic retinopathy(DR)mainly include microaneurysms,hemorrhage,and hard exudates.However,clinical diagnosis of DR mainly relies on manual screening,which is time-consuming.Therefore,automatic detection of early DR lesions is crucial,based on techniques of image processing.At present,the methods for DR detection are mainly divided into machine-learning-based methods and deep-learning-based methods.Due to the low representation of hand-crafted features,the performance of the machine-learning-based methods is limited,while deep-learning-based methods can automatically extract high-level semantic features to achieve the strong representation for the lesions,resulting in higher sensitivity.Nevertheless,multiple convolution and pooling layers will cause the loss of lesions information,due to the finite information of small objects in fundus image.The unbalanced data(i.e.,few positive samples and overwhelming number of negative samples)leads the over-fitting and reduces the performance of detection.Firstly,this paper proposes a deconvolutional neural network(DNN)to detect microaneurysms.This network uses the deconvolution layer,instead of pooling layer,to recover the features lost in the convolutional operation.The proposed method is tested on multiple public datasets,achieve significant detection results.Moreover,the proposed method achieves higher accuracy than the state-of-the-art methods.Furthermore,the proposed deconvolutional neural network is used to simultaneously classify microaneurysms,hemorrhage,and hard exudates from the public dataset DIARETDB1.The sensitivity of the three types of lesions are 82.6%,84.4%,87.4%,respectively.The total accuracy reaches 91.1%.In order to increase the difference between different objects and reduce the difference within the class,we apply the inception model on the basis of deconvolutional neural network.The purpose is to learn more high-level and diverse features by increasing the width of model.The proposed inception network use DIARETDB1 to train and test.The sensitivity of three types of lesions are up to 83.3%,83.2%,and 95.9%,respectively.Furthermore,the total accuracy reaches 93.4%.Experimental results demonstrate that the sensitivity of hard exudates is significantly improved,and the accuracy are higher than deconvolutional neural network without inception model.Finally,we use the deconvolutional neural network and the inception network for multi-class classification on the clinical data.The total accuracy of the two methods reach91.6% and 93.3%,respectively.It verifies that the proposed networks can effectively classify early lesions of DR in retinal fundus image.