Researches On Intelligent Analysis Method For OCT Retinal Images Self-Supervised Learning

Eye is one of the most important organs in human body.In recent years,the amount of patients with ophthalmic diseases has been increasing and the types of diseases are complex and changeable.Optical coherence tomography(OCT)can provide high-resolution imaging by scanning the human retina,which is an indispensable tool for the diagnosis of ophthalmic diseases.At present,the diagnosis of retinal diseases mainly relies on the subjective judgment of ophthalmologists on OCT retinal images.However,manual diagnosis is time-consuming,laborious and requires high professional level of doctors,resulting in misdiagnosis and missed diagnosis from time to time.Therefore,the research on automatic analysis technology and diagnosis system based on OCT retinal images has extremely high clinical significance.In recent years,with the rapid development of deep learning,many researchers have applied it to OCT retinopathy analysis,and achieved good performance.However,this success heavily relies on large-scale uniformly distributed tagged datasets,and it’s very difficult to obtain a large amount of annotation data from different devices in the field of medical images.In this context,this paper focuses on the research of OCT retinal image lesions classification,diseases segmentation,and designs an intelligent diagnosis system for OCT images in ophthalmology.Specifically,1.OCT image labeling is difficult and the amount of relevant labeling datasets is limited,which leads to the unsatisfactory classification results of lesions.To tackle this problem,a self-supervised patient specific learning algorithm for OCT retinal image lesion classification is proposed in this paper.In this algorithm,patient-related selfsupervised tasks are designed according to the characteristics of OCT retinal images.By self-constructing labels,the model learns the characteristics of OCT retinal images,such as film thickness,texture and structure.The experimental results show that this method can mine the significant information from the unlabeled data,assist the actual classification task,and reduce the requirements of well OCT retinal image classification accuracy on the amount of labeled data.2.Retinal image acquisition devices are diverse,the distribution of training data(source domain)and clinical data(target domain)is frequently different,resulting in unsatisfactory actual segmentation results.To solve this problem,this paper proposes a self-supervised feature alignment algorithm for OCT retinal image lesion segmentation.This algorithm is designed based on the U-net network,according to the inter-frame continuity of OCT image scanning results and the separability of data sources,the time sequence prediction task and feature domain source prediction task are designed to learn the common temporal and spatial features of data in different domains.The experimental results show that this method can render the data from different domains to achieve consistent distribution at feature level,thereby improving the segmentation accuracy of target domain.3.At present,most researches on retinal OCT image lesion analysis are still in the theoretical stage.In this paper,an intelligent analysis system for OCT retinal images is designed based on Py Qt5 and Python.The two retinal image analysis methods proposed in this paper are integrated into the system.In addition,due to the limitation of the dataset in this paper,the sorts of lesions that can be analyzed by this system are restricted.We have additionally integrated excellent lesion analysis codes in relevant fields,so that this system finally realizes the functions of basic file management and OCT image denoising,lesion classification and lesion segmentation,which can be used as a tool to assist experts in diagnosis.