Design And Optimization Of Central Serous Chorioretinopathy Segmentation Based On Fully Convolutional Neural Network

Visual impairment is an important factor affecting people’s life and work.Recently,with the development of economy,the common usage of electronic products has led to an increasing incidence of eye diseases,such as central serous chorioretinopathy(CSC),which has caused serious decline of people’s vision.As the medical imaging technology develops rapidly,spectral-domain optical coherence tomography(SD-OCT)can provide the threedimensional fine structure inside the retina,providing the possibility to quantify the lesion area of the retina.Therefore,how to use image processing technology to automatically identify and quantify the lesion area has become a research hotspot.Based on the SD-OCT retinal imaging,this thesis mainly focuses on the development and optimization of CSC segmentation algorithm.The main works of this thesis are shown as followed:(1)We propose the double-branched and area-constraint fully convolutional neural network for the segmentation of CSC lesion region.In SD-OCT images,there are several problems,including the small lesion area,blur boundary between the lesion area and the normal area,and the existence of multiple lesion areas with different sizes.With regard to these problems,we first propose the double-branched structure to improve the model’s ability of learning multi-scale information.In addition,we also combine area constraint to help the model learn the relationship between the input image and the area of the lesion region,thus improving the accuracy of CSC segmentation in SD-OCT images.(2)We develop a neural network compression algorithm based on sparse self-representation and particle swarm optimization.Deep learning-based CSC segmentation model always needs large storage to save model and long time to predict segmentation.Therefore,we propose a neural network compression algorithm to deal with this problem.To reduce the performance loss after the model is compressed,the compression algorithm proposed in this thesis first projects all convolutional kernels into the base subspace to establish the similarity matrix,and then under the framework of spectral clustering,each convolutional kernel is clustered to different subspaces and is pruned according to the similarity matrix.At the same time,in order to trade off the compression ratio and the performance of the compressed model,we use particle swarm optimization algorithm to optimize the pruning process,achieving better overall compression performance.(3)We design a segmentation software for CSC.In this software,users can use the pretrained model for CSC segmentation,or use individual data to train and compress the model.With simple operation and abundant functions,this software integrates commonly used predefined items of operations and hyperparameters,which will bring great convenience for users.