Semantic Segmentation For Wet Age-related Macular Degeneration Based On Deep Learning

Age-related macular degeneration（AMD）is a degenerative retinal disease that is the main leading cause of severe visual impairment in older adults.There are 2 types of AMD: dry and wet.Between 80 and 90 percent of the blindness and severe vision loss occurs in the wet form of the disease.In clinical diagnosis,ophthalmologists need to manually identify lesion regions and then make diagnosis decisions of these lesion regions.Such manual analysis is time-consuming and demanding for expert graders,which is also prone to yield subjective results.Consequently,building an automatic and reliable computer-assisted fundus image analysis technique with computer vision is required for the efficient diagnosis of wet AMD.In this work,we propose AMD-Net,a semantic segmentation network of wet AMD based on deep learning.We use this new architecture to jointly detect and segment edema and hemorrhage lesions.AMD-Net uses the simplified U-Net as the base network architecture.To extract and fuse multi-scale features,the AMD-Net proposes an encoder feature fuse unit.And with the attention module,the encoder feature fuse unit can give important features higher weight.In the skip connection part,the AMD-Net introduces a skip connection block to reduce the semantic gap between encoder features and decoder features.Decoder attention block is applied to fuse multi-scale decoder features,and using the global context fuse block instead attention module can enhance the contribution of high-level features in tiny regions segmentation.The proposed method is compared with several advanced segmentation networks using the self-constructed fundus dataset.The experimental results indicate that the proposed method achieves the best results and demonstrates its effectiveness.We have also investigated the effects of encoder feature fuse unit,skip connection block,decoder attention block and global context fuse block on the performance of the proposed method in the ablation experiment.