Researches On Analysis Method Of Retinal Disease In OCT Images Based On Deep Learning

Optical coherence tomography(OCT)is a novel medical imaging technique,which can provide in-vivo imaging of human retina with high resolution.Due to its great advantages of non-contact,non-invasion,and fast imaging speed,OCT has been widely used in clinical ophthalmology,and become one of the indispensable tools for the diagnosis of retinal diseases.At present,the diagnosis for retinal diseases mainly depends on ophthalmologists’ manual analysis of retinal OCT images.However,such a process of manual analysis is extremely boring,time-consuming,laborious,and prone to yield subjective diagnosis results.Therefore,it is greatly significant and promising to develop computer-aided diagnosis(CAD)based technique for retinal disease analysis in OCT images.In recent years,deep learning(DL)technique has achieved great development and breakthrough in many fields,such as image classification and recognition,object detection and tracking,image segmentation,video processing and so on.Under this background,this paper mainly focuses on the following three aspects of retinal disease analysis by using deep learning models: retinal layer segmentation,retinal lesion detection,and retinal disease classification.(1)It is challenging for traditional retinal layer segmentation algorithms to effectively represent and identify retinal layer distortions caused by various kinds of retinal lesion structures,and thus their segmentation performances on OCT images with complex lesion structures still leave much to be desired.To tackle this problem,this paper proposes a convolutional neural network(CNN)based method for retinal layers segmentation in OCT images.The method first utilizes CNN to learn multiple probability maps of retinal layer boundaries from OCT images,and then uses a graph search model to calculate the accurate optimal layer boundary from its corresponding probability map.The method can effectively learn rich abstract semantic features and thus fully represent the retinal lesion structures.Experimental results show that the proposed method can effectively improve the segmentation accuracy of retinal layers in OCT images with complex retinal lesion structures.Experimental results on 2915 AMD OCT images exhibit that the average segmentation errors of the proposed method for INL,OPL and RPE layers are 0.22,0.66 and 1.04 pixels,respectively.(2)The traditional retinal lesion detection algorithms are either based on fixed mathematical model or utilizing hand-crafted features/rules to detect lesion structures.However,due to the co-existence of multiple lesion structures within the same OCT image,high variabilities in type,number,shape,size,and location of lesion structures,as well as the negative influence of retinal blood vessels,those traditional algorithms are very difficult to achieve satisfactory detection results.To deal with this problem,this paper designs a fully convolutional network(FCN)to detect retinal lesion structures from OCT images.The network is mainly composed of convolutional module,local contrast-enhancing module,as well as deconvolutional module,which can accurately identify retinal lesion structures while fully preserving their spatial locations.Experimental results on 2500 drusen,CNV,DME and normal OCT images demonstrate that the proposed method can achieve high detection effect of Dice metrics of 0.92,0.87,0.89 and 0.95,respectively.(3)The type,number,size,and shape of retinal lesion structures are one of the important criterions for ophthalmologists to make diagnostic decisions on retinal diseases.Therefore,retinal lesion structures can be used to guide the classification of retinal diseases.However,existing classification methods do not make full use of the information of retinal lesion structures.To handle this problem,this paper proposes a lesion guided network(LGN)for retinal disease classification from OCT images.The method employs a series of novel lesion guided module to introduce the information of retinal lesion structures to CNN classification network so that the network can extract local discriminative features of lesion structures for accurate and efficient classification while still retaining useful information in global OCT images.Experimental results on 70404 OCT images show that the proposed method can obtain high overall accuracy of 90.1% in classifying drusen,CNV,DME and normal retina.