Deep Learning-based Segmentation Of Choroidal Neovascularization

Choroidal neovascularization(CNV)is a retinal vascular disease.It seriously endangers the visual health of patients and is one of the main causes of blindness.The pathological features of CNV are complex and diverse.It is helpful for clinicians to use automatic algorithm to segment CNV accurately.This paper studies CNV segmentation based on Spectral Domain Optical Coherence Tomography(SD-OCT)and(Optical coherence tomography angiography)OCTA images.The main research contents are as follows(1)A 3D CNN based 3D segmentation method for choroidal neovascularization was proposed.3D U-Net was used as the basic structure,and 3D convolution was used to extract features from SD-OCT images.In order to achieve the balance of space occupation and model efficiency,the convolution kernel with the size of 7 7 3 and the step size of 2was applied in the first layer of the network.As a result,the receptive field of the network was increased and the size of the feature map was rapidly reduced.In order to better extract the features of CNV lesions,a cross convolution module was proposed,which employs two convolution kernels with size of 7 3 3 and 3 7 3.In order to further enlarge the receptive field of the network and improve the ability of multi-scale feature extraction,a dilated connection layer was proposed,in which dilated convolutions with different dilation rates were added to the skip connection structure of the network.The results of experiments showed that the method we proposed is able to segment CNV region effectively on SD-OCT images.(2)A choroidal neovascularization segmentation strategy based on multimodal feature fusion was proposed,which uses SD-OCT and OCTA modal data for fusion and segmentation.The strategy consists of two stages: in the first stage,a U-Net model was trained with OCTA images,and the final model parameters were selected according to the convergence of the network training process.In the second stage,a feature fusion network of two-branch encoder was constructed based on U-Net,and a feature fusion module was designed to select and fuse the feature maps from two branches.The parameters of encoder obtained in the first stage were loaded into the OCTA branch of the network and were frozen.The feature maps of each module of the two branches were fused through the fusion module,and then used as the input of the next module of the SD-OCT branch.Finally,the network took pairs of SD-OCT and OCTA data as input,and used the gold standard as the label for training and optimization.4-fold Cross Validation and patient independent experiments verified the effectiveness of our scheme.