Anterior Segment Optical Coherence Tomography Image Analysis Based On Deep Learning

Primary angle-closure glaucoma(PACG)is the major ophthalmic disease of blindness in Asian populations.Anterior chamber angle(ACA)classification is key task for diagnosis of angle-closure glaucoma in Anterior Segment Optical Coherence Tomography(AS-OCT).The existing ACA classification methods focus on binary classification(i.e.,open angle and angleclosure)in 2D slice.However,in clinical diagnosis,it requires a more precise ACA manner(i.e.,open,narrow,and synechia angles)to benefit clinicians who seek better to understand the progression of the spectrum of angle-closure glaucoma types.But,due to the dynamic adhesions in the anterior segment,the classifier based on static AS-OCT image may not be able to accurately predict the results.Therefore,this paper proposes a series of deep learning algorithms in AS-OCT image,which can not only classify open angle and angle-closure glaucoma from a single AS-OCT image,but also recognize the time dynamic information of the features from the AS-OCT image sequence for ACA classification.The main contents of this paper are as follows:(1)In this paper,the ResU-Net method based on the attention mechanism is used to segment iris.Based on the ResU-Net structure,the multi-scale residual module is introduced to improve the ability of convolution layer to extract multi-scale features;the attention mechanism module is introduced to optimize the weight of the target features in the encoder and decoder,which can enhance the features of the target area while suppressing the background and noise area;the multi-scale image input and multi-level function output modules are introduced to supervise the feature learning of each layer.(2)This paper proposes a classification model based on multi-scale feature aggregation.Based on the clinical prior knowledge,the ACA region in AS-OCT image is extracted and classified.In the method proposed,a multi-scale feature discrimination aggregation module is designed,and a multi-scale aggregation network is constructed to classify the extracted ACA area.(3)This paper proposes a classification model based on multi sequence feature fusion.In this paper,2D image is processed into 3D data(image sequence).The convolutional neural network encoder is used to extract the depth features of each image in the sequence,and then input the convolutional recurrent neural network(Conv LSTM)to extract the sequence image features effectively.Therefore,this paper proposes time weighted cross entropy loss(TC-loss),and when making a decision on each time,we take the average value of all the sequence features in front of this time.And it can connect the features of all the sequence feature,and input them to the classifier for ACA classification.