Research On Intelligent Diagnosis Of Retinal Disease Based On OCT Image

Retinal layer segmentation in Optical Coherence Tomography(OCT)images is a key step in the diagnosis of retinal diseases.However,manual retinal layer segmentation is a subjective and consuming task.Therefore,automatic retinal layer segmentation can improve diagnostic efficiency and reduces the labeling error rate.The further construction of an intelligent diagnosis system for retinal diseases can reduce the workload of doctors,facilitate the early examination of patients and deal with the imbalance of medical resources.Intelligent diagnosis of retinal diseases based on OCT images remains a challenge due to the lack of reliable and interpretable analytical methods.In this paper,intelligent segmentation of the retina and intelligent diagnosis of retinal diseases are studied.The specific contents are as follows:(1)Theoretical and experimental research on intelligent retinal layer segmentation is carried out.A retinal layer segmentation dataset containing 1,200 OCT images was collected and constructed to divide the retina into 11 layers and fluid.Some existing advanced image segmentation algorithms were selected to conduct retinal layer segmentation experiments on this dataset and their results were analyzed.In order to solve the problem of inaccurate segmentation in the boundary region,a multi-task dual boundary aware network was proposed.which added the boundary regression task to enhance the learning of the boundary region.Based on the retinal structure.a dual boundary representation is proposed to encode retinal layers from two directions.The multi-task consistency constraint is used to promote the help of boundary regression tasks for retinal layer segmentation.The average Dice coefficient of 89.1%was achieved in this multi-task dual boundary aware network for retinal layer segmentation,which surpassed the existing image segmentation networks.(2)The intelligent diagnosis of retinal diseases is studied by combining OCT images with medical prior knowledge.A total of 4,547 OCT images from four different commercial devices from nine international medical centers were collected to finely divide the epiretinal membranes(ERM)into six disease stages.Some existing image classification networks were selected for the grading diagnosis of ERM in OCT images.and the help of integrating medical prior knowledge to the grading diagnosis of ERM was explored.In order to better perform an adaptive fusion of OCT images and medical prior knowledge,a bidirectional guidance module is proposed,which uses the same level of image features and medical prior features in the network to guide each other and promote feature extraction ability.The proposed two-stage network achieved 84.8%accuracy on the test data set in the ERM grading diagnostic task,2.9%higher than the diagnostic approach using OCT images only.(3)The retinal layer segmentation network and ERM diagnosis network are developed into two convenient cross-platform software.PyQt and other tools are used to combine the research results of retinal layer segmentation and retinal disease diagnosis to develop a retinal-aided segmentation system and a grading diagnosis system of epiretinal membrane.The retina-aided segmentation system performs layer segmentation and fluid segmentation for 2D OCT images and 3D OCT images,displays 2D segmentation results and 3D segmentation results,and supports the export of segmentation results into modified json annotation files.The epiretinal membrane grading diagnosis system makes diagnoses according to OCT images entered by users and patient information to obtain diagnosis results and diagnosis opinions.In order to enhance the interpretability and fault tolerance,the diagnostic basis of the system and the uncertainty of the diagnosis results are displayed at the same time.