Research On Multi-scale Retina Image Segmentation Based On Fully Convolutional Neural Network

Retinal images contain abundant information of tissue structure and lesion.The processing and analysis of retinal images can provide important clinical basis for early screening of various fundus diseases.However,due to the restrictions of imaging conditions and other objective factors,it is time-consuming and tedious to diagnose diseases only by ophthalmologists' manual screening of fundus images.It is effective to achieve automatic segmentation of retinal images by combining image processing and deep learning methods,which has great impact on large-scale early screening of diseases.In this paper,we mainly focus on retinal images segmentation based on deep learning and have achieved retinal vessel segmentation and optic disc segmentation.The contributions of this paper can be summarized into three aspects as follows:1)In the work of retinal vessel segmentation,we propose a fully convolution encoder-decoder network for retinal vessel segmentation without hand-crafted features or specific post-processing.Firstly,the network can combine the low-level features with the high-level semantic information.Then,we introduce residual learning and Dropout strategy into segmentation for further learning of edge and texture information.2)To address the problem of detecting small scale retinal vessel,we propose an atrous spatial pyramid pooling module for multi-scale segmentation by fusing hierarchical information from different receptive fields.On the one hand,the module can enlarge the receptive field and combine the context information to improve the ability of detecting small vessels.On the other hand,the atrous convolution greatly reduces the learning parameters,which contributes to the model training and optimization.Moreover,we apply the class-balanced loss function to solve the problem of imbalanced distribution of samples.Experimental results show that the proposed algorithm can preserve the details of images such as texture and edges.Our algorithm has superior performance compared with state-of-the-art methods.3)In the work of optic disc segmentation,we propose an improved method for optic disc segmentation based on attention mechanism and U-Net network.By introducing attention model into the encoder module,the precise disc segmentation can be achieved learning most relevant features of the target region with less prior knowledge.The proposed method is tested and analyzed on DRIONS and MESSIDOR datasets.Experiments show that the proposed method combined with the attention model can preserve the edge well and has better performance compared with the original U-Net segmentation network.