Research On Optic Disc And Cup Segmentation Method Based On Deep Convolution Generative Adversarial Networks

Glaucoma is a chronic eye disease which leads to irreversible vision loss.The cup to disc ratio plays an important role in the screening and diagnosis of glaucoma.Thus,the accurate and automatic segmentation of optic disc（OD）and optic cup（OC）from fundus images is a fundamental task.Unlike other eye diseases such as cataracts and myopia,vision loss from glaucoma cannot be reversed.Early screening is thus essential for early treatment to preserve vision and maintain life quality.However,the manual assessment by trained clinicians is time consuming and costly,and not suitable for population screening.Most existing methods have insufficient feature information extraction,and hence they are susceptible to pathological regions and low quality images,and have a small receptive field when the depth of the network is fixed and cannot capture global context information of accurately different sizes of regions to be segmented,which result in insufficient extraction of feature information.Some methods rely on hand-made visual features,and do not consider the relationship between OD and OC,which result in poor model segmentation performance.Based on these problems,this paper proposes a multi-label deep learning model based on generative adversarial networks.The main research content of this paper can be summarized as follows:（1）Propose Generative Learn Networks（GL-Net）,it is a multi-label deep convolutional neural network model,that combines the generative adversarial networks,which realizes automatic segmentation of the OD and the OC simultaneously.GL-Net consists of two network structures:generator and discriminator.During training,the generator and discriminator perform minimax games and conditional constraints to optimize.In the loss function,we add the L₁ distance function and the cross entropy function to prevent the mode collapse when the model is trained,which makes the segmentation result more accurate.The downsampling factor of the generator is reduced to alleviate the loss of excessive feature information,and use the skip layer connection to combine the rough low-level segmentation information of the decoding layer with the fine appearance information of the encoding layer,which makes the segmentation feature map get more complete context information.The data augmentation is used to augment the training data and use the generalization ability of transfer learning in deep convolutional neural network,which reduces the over-fitting of the limited medical dataset by transforming the rich feature information learned from the relevant dataset.The experimental results show that GL-Net has better segmentation results than other models.（2）A new multi-label deep learning model Dilated Reconstruction Networks（DR-Net）is proposed,which is an improvement of the generator based on GL-Net.In the generator,the encoder part introduces cascaded dilated convolution,and the decoder part introduces scale reconstruction layer for upsampling.Due to the“gridding”issue of the dilated convolution,when the dilated rate is larger,the feature information collected is more sparse,and the captured feature information may be irrelevant,resulting in a lower accuracy of model segmentation.Therefore,a cascaded dilated convolution is proposed to alleviate this problem.The cascade convolution is cascaded by N convolutions with different dilated rates,which not only increases the receptive field of the model,but also all the feature information in the receptive field area can be fully covers without any holes and missing edges.It improves the model’s ability to understand local contextual information and maintains the correlation between feature information in the receptive field.In order to increase the speed of the model,the decoder uses the scale reconstruction layer to upsampling,which periodically rearranges the information in the feature map and compresses the number of channels in the feature map to expand the resolution to achieve upsampling.The scale reconstruction layer does not introduce extra parameters and computational cost,which can increase the speed of DR-Net,and it is learnable,so it can capture and recover detailed information lost in downsampling.The experimental results show that DR-Net further improves the accuracy of the OD and OC segmentation.