Research On The Segmentation Of Subretinal Fluid Lesions Based On Deep Neural Network

The segmentation technology of spectral domain optical coherence tomography(SD-OCT)detects early subretinal fluid(SRF)by quantitatively analyzing the lesion area to assist physicians in clinical diagnosis and treatment.However,the interference of noise,as well as the differences in lesions between patients,makes the precise segmentation of SRF lesion areas extremely difficult.To improve the performance of SD-OCT image processing methods based on deep neural networks,the key issues in the segmentation of SRF lesions are studied in thesis.The main work of thesis is as follows:1.A progressive feature fusion attention dense network(PFFADN)is proposed in response to the problem that speckle noise contributes to degradation of raw SD-OCT image quality.First,densely connected dense blocks are arranged in the deep convolution network,and the shallow convolution feature map with the deep one extracted from each dense block is sequentially connected to form a residual block.Second,the key features are extracted and the irrelevant ones are suppressed by adding the attention mechanism to the network.Finally,the output feature maps from all dense blocks are fused and input to the reconstruction output layer.Compared to the state-of-the-art denoising algorithms on retinal SD-OCT images,PFFADN has a better denoising effect.2.A segmentation method based on deep attention convolutional neural network(DACNN)is proposed for the differences in lesions between patients.First,the improved residual structure is taken as the underlying convolution module of the segmentation model with U-Net as the basic framework.Second,a dual attention mechanism of spatial regions and feature channels is introduced in the process of fusing shallow and deep features of the network,enabling the model to adaptively select important information from the feature space for fusion.Finally,a deep attention module with hybrid kernel convolution is used to capture multi-scale contextual information in a fully convolutional neural network architecture with deep supervision and residual connection.Compared to the state-of-the-art segmentation algorithms on retinal SD-OCT images,DACNN has better segmentation performance.3.A deep network embedded with rough fuzzy discretization(RFDDN)is proposed for the uncertainty information in SD-OCT images.First,the information decision table of SD-OCT fundus image segmentation is established,and the membership degrees of pixels to each segmentation region category is calculated by fuzzy c-means clustering to achieve the fuzzification of pixel categories.Second,the individual fitness function based on rough fuzzy sets is designed and the genetic algorithm is employed to find the optimal breakpoints for feature discretization of SD-OCT fundus images to reduce the uncertainty caused by noise and redundant information.Finally,feature discretization based on rough fuzzy sets is taken as a pre-module of the deep neural network and the deep supervised attention mechanism is introduced to obtain the important multi-scale information.The experimental results show that RFDDN can effectively evaluate the uncertainty of data and eliminate the negative effects of redundant information,substantially improving the accuracy of SRF lesion region segmentation while taking into account interpretability and computational efficiency.