The Application Of Semantic Segmentation Networks In Retinal Vessels Segmentation

The structure and form change of retinal blood vessels are highly correlated with many ocular diseases and angiocarpy diseases,which is an important evidence for diagnosing and screening these diseases.However the efficiency of manual diagnosis and experience needed is a severe drawback,so the study of automated retinal blood vessels segmentation algorithm is very significative.In this dissertation we propose a novel fully convolutional neural network that combine the new approach in semantic segmentation task for the seg-mentation of retinal blood vessels.Our proposed framework integrate dilated convolution,layer-by-layer decoder and multi-scale context fusion module,and test out model on public datasets.Our contributions can be summarized as:1.We discuss the preprocessing of retinal blood vessels segmentation dataset with a baseline model of U-Net framework.We analyze how the crop and sampling,image preprocessing and augmentation methods influence the final results,and choose the most effective combination of methods by comparative experiments.2.We summarize the current research status of semantic segmentation task,and test some representative approaches of semantic segmentation task on retinal blood vessels segmentation task to find out if these approaches can also be valid in this specific task.We proof that some usual approaches designed for semantic segmentation,such as dilated convolution and context modules,can also improve the performance of retinal vessels segmentation networks.3.We propose a fully convolutional neural network with dilated resid-ual convolution block,layer-by-layer decoder and context fusion module.The proposed method integrate dilated convolution block and context fusion mod-ule which is widely used in semantic segmentation frameworks,with skip-connection and layer-by-layer decoder that often applied by medical image seg-mentation networks.Our model can make full advantages of these approaches to generate prediction mask with both accurate class and exact boundary.Our proposed segmentation method can extract the structure and form of retinal blood vessels in color fundus image,and to the best of our knowledge,the proposed method shows state-of-the-art or second-best segmentation accu-racy in DRIVE dataset and CHASE_DB 1 dataset.Meanwhile,we proof that some approaches designed for semantic segmentation task can also improve the performance of retinal vessels segmentation task.