SD-OCT Retinal Image Lesion Segmentation Based On Detail-preserving Semantic Segmentation Network And Its Applicatio

For SD-OCT images that are widely used in the field of retinopathy,a reliable and efficient OCT image segmentation algorithm can assist ophthalmologists to quickly assess the development of the disease and decide the treatment plan.Considering the advantages of deep learning semantic segmentation methods in robustness and adaptability,this paper conducts research on deep learning-based retinopathy segmentation methods.According to the characteristics of retinopathy,the focus of the work is on the preservation and reasonable application of detailed information.For the above reasons,some network modules and structure which emphasize on detail preservation and feature fusion are proposed.The specific research contents are as follows:(1)Aiming at the optimization of retinopathy contour segmentation model,a DataDependent Dual-Path network(D3PNet)was proposed.The network model designed the parallel structure of expansive path and guidance path from three perspectives of detail preservation,contour emphasizing and optimized feature fusion to adapt to the variation and diversity of retinopathy,retain more detailed descriptions of lesion areas and improve segmentation accuracy.A comparative experiment with mainstream semantic segmentation networks and medical image segmentation networks is designed.Compared with the most advanced network models,the method in this paper has a large improvement in the public CNV data set,and also make a breakthrough in the NRD data set.The ablation experiment further verified the effective role of different modules of the network.(2)For geographic atrophy(GA)growth prediction problem,a Long and Short follow-up information Adaptive lesion area prediction segmentation model(LSANet)is proposed.By introducing convolutional recurrent neural network and self-attention mechanism,LSANet learns the characteristics of current lesion follow-up and mines the law of lesion development between previous follow-ups.Therefore,LSANet can simultaneously segment retinal lesions and predict the changes of GA lesions on fundus projection images after a given follow-up time.LSANet also uses the proven upsampling path in D3 PNet to achieve detail information preservation.Comparison and ablation experiments are designed on the GA dataset to verify the advantages and effectiveness of the LSANet network model proposed in prediction.(3)A system for segmentation and prediction of retinopathy in SD-OCT images was designed and implemented.The system integrates D3 PNet and LSANet,and constructs five modules: function selection,image reading,lesion segmentation,lesion prediction,III result display and result storage.Users can select different networks,network parameters and application data through the interface to perform SD-OCT slice segmentation tasks and GA lesion prediction segmentation tasks for CNV and NRD lesions,and intuitively understand the segmentation and prediction results.