Cross-domain Retinal Fluid Classification Based On Deep Learning

As one of the main manifestations of retinal diseases,subretinal fluid has become one of the leading causes of blindness in developed countries.An accurate assessment of the type of fluid in the fundus helps the clinician plan treatment and save the patient from vision loss.Spectral Domain optical coherence tomography(SD-OCT)has been widely used in the computer-assisted diagnosis of retinal diseases due to its non-invasive,high resolution,and real-time characteristics.In recent years,the deep learning model has achieved good results in the classification of SD-OCT images obtained from a single device.However,when the model trained on SD-OCT images obtained from one device is applied to the classification of SD-OCT images obtained from another device,it often results in significant degradation of model performance.This is because their data distribution is markedly different.Although labeling images and then fine-tuning the model can solve this problem to some extent,in the medical field,image labeling often requires more labor and material resources.Domain adaptive technology provides a new way to solve these problems.A classifier with strong generalization performance can be trained by learning the domain invariant features of SD-OCT images from two devices.The knowledge learned on labeled data sets can be applied to unlabeled datasets and obtain good classification performance.This paper studied the domain adaptation problem of cross-domain subretinal fluid classification,and the main research contents were as follows:(1)A multi-stage domain adaptive cross-domain classification method for subretinal fluid was proposed to learn generalized valid domain invariant features from labeled source domain SD-OCT images and unlabeled target domain SD-OCT images and then be used for cross-domain classification of subretinal fluid.The task-independent feature alignment(TiFA)module firstly maps SD-OCT images from different domains to the same feature space to effectively preserve the original image information to perform preliminary feature alignment.Subsequently,a downstream task-related feature alignment(TsFA)module extracted class-related features from the TiFA output,then used for fluid classification.The experimental results show that this method can effectively judge the type of subretinal fluid on SD-OCT images.(2)A cross-domain subretinal fluid classification method based on coarse label guidance and angular contrastive learning is proposed to solve the problem that the differences between healthy images in SD-OCT images from different devices are slight and the healthy images occupy a large proportion in the data set.In this method,a coarse label guidance module is first used to dichotomize the health and fluid of SD-OCT images from source and target domains.Then,guided by the classification results,the feature angular contrastive learning module is used to learn the feature differences between the two categories of health and fluid to obtain more effective domain invariant features.At the same time,the TsFA module was used further to extract the detailed category-related information of domain invariant features and then used for classifying subretinal fluid.The experimental results show that this method is more superior and advanced in the task of cross-domain subretinal fluid classification.