SD-OCT Retinal Image Lesion Segmentation And Application Based On Multi-scale Semantic Segmentation Network

Based on deep neural network models,retinopathy segmentation has important academic research significance and broad application prospects in various fields,including pattern recognition,artificial intelligence,and medical image processing.Among them,in the neural network,the multi-scale feature extraction module is important in processing various sizes of segmentation objects.Aiming at the multi-scale object segmentation problem in lesion segmentation and other common problems in semantic segmentation,in this paper,some optimized multi-scale feature extraction modules and other performance optimization methods were proposed.The specific research works are as follows:(1)Aiming at the problems of the huge scale change span of NRD lesions and the similarity of gray distribution in foreground area and background area,this paper put forward a multilevel pyramid pooling model on the basis of residual network.Meanwhile,a difference loss function was mentioned to further optimize the segmentation effect.The proposed model was compared with some classic and state of the art algorithms in semantic segmentation,and other algorithms for NRD segmentation.Finally,the comparison experiments with other methods were carried out on three different datasets.The final experimental data and segmentation results proved that the multiple pyramid pooling residual convolution model proposed in this paper is significantly better than other algorithms in various evaluation indicators,and it is known that the segmentation effect is better in multi-scale in terms of avoiding holes and edge details segmentation etc.from the segmentation results.(2)According to multi-scale attention,a CNV segmentation model with classification assistant branches was proposed.In the CNV lesion segmentation,the imaging of CNV lesions in SD-OCT images is more complex,which not only has the problems of multi-scale,but also has different imaging differences and serious weak boundaries among different lesions that are caused by different symptoms of CNV lesions.These problems lead to the difficulty of CNV segmentation that is far more complicated than that of NRD.In order to solve the problems of CNV lesion segmentation,multi-scale attention module and classification assistant branch were added to the proposed network.In addition,for the weak boundary problem,a loss function with boundary constraint was added when training.Finally,a comparative experiment was conducted on a dataset of 128 cubes from 10 patients.The experimental results displayed that the proposed model has higher segmentation accuracy and stability than some other existing methods.(3)A retinopathy analysis system based on SD-OCT images was designed and implemented,which could select NRD lesions or CNV lesions to segment the retinal image and estimate the volume of the lesion area.It can also be applied to verify various evaluation indicators for the lesion image with the true label image.Most importantly,the system could provide convenient image browsing and demonstration functions,and thus 2D slice images and3 D images in different directions can be demonstrated.