Retinal Layer Segmentation And Disease Screening On OCT Images Based On Deep Learning Method

OCT is a high-resolution,non-invasive,noninvasive biological tissue imaging technology that has become an important tool for retinal 3D imaging and is widely used in clinical ophthalmology to diagnose eye diseases.Many important eye diseases as well as systemic diseases manifest themselves in the retina,so the development of retinal image processing technology is important for the diagnosis of retinal diseases.Retina is a complex hierarchical structure,the inner structure of the retina and retinal thickness information are an important indicator of the diagnosis of eye diseases.OCT technology can be used by ophthalmologists to analyze information on changes in the inner structure of the retina and retinal thickness.Thus,retina layer segmentation in OCT scans has become an important goal.Extensive approaches have been designed for the retinal layer segmentation in OCT images.Graph-based methods were commonly used in layer segmentation.However,most of these methods require a lot of human efforts for determining an appropriate model to compute good edge weights.In order to design an efficiently appropriate layer segmentation model,this paper provides a retinal layer segmentation method based on fully convolutional network and graph theory.In detail,firstly,we design a fully convolutional network with hidden layer supervision,which can automatically learn multi-scale and multi-level features to generate accurate boundary probabilities which means the probability value of each pixel belongs to actual boundary.Secondly,since the thickness of some retinal layers are thin,the boundaries between two layers are close and the boundaries are blurry.Thus,we enhance the proper probability maps in order to increase the contrast between boundaries and non-boundaries.At last,we use the values of boundary probabilities as the edge weights of a graph followed by a shortest path algorithm for the final segmentation.The proposed method is evaluated on a dataset with 130 OCT B-scans.Experimental results show the method get more accurate segmentation results.Automatic screening of retinal disease in optical coherence tomography(OCT)images is an important yet challenging task.Conventional studies generally discern a specific disease in macular OCT images or optic nerve head(ONH)centered OCT images,to assist ophthalmologists to diagnose more retinal diseases,this thesis propose a novel scheme to perform the respective classification over normal and abnormal retina in 3 types of OCT images:macular OCT images,ONH-centered OCT images and wide view OCT images acquired both at the macula and the ONH.This paper also propose a new 3D convolutional neural network for the classification task.Specially,an auxiliary supervision branch is introduced after a hidden layer to improve discrimination of learned features and the corresponding supervision function is constructed to be used as a regularization term.The final classification result is obtained by combining the output of the mainstream network and the auxiliary network.The proposed framework was tested on a large OCT image dataset with 873 volumetric OCT volumes and achieved an overall accuracy as high as 94.20%.