OCT Image Segmentation Of Choroidal Neovascularization Based On Deep Learning

Age-related macular degeneration(AMD)is a major cause of vision loss and blindness in older adults worldwide.Choroidal neovascularization(CNV)is a main sign of AMD.The segmentation of CNV lesion area and the corresponding retinal layers is helpful for doctors to judge the disease course.Optical coherence tomography(OCT)is clinically widely used in CNV diagnosis due to its non-invasive,high resolution and fast imaging.Computer-aided OCT segmentation method contribute to segmenting accurately and quickly,which not only reduce the manual segmentation burden of doctors,but also reduce the difference in labels made by different observersA fully automatic Deep Learning method is proposed in this paper for segmentation of OCT images which objects suffering from CNV.The proposed algorithm framework,based on the "encoder-decoder" convolutional neural network,consists of several new strategies for retinal layers' segmentation.To deal with the contrast reduction of adjacent retinal layers caused by CNV,a Channel to Feature map pixel block is firstly proposed to convert particular channels into pixels in one bigger feature map,to realize feature map amplification.A convolution layer is followed and optimized by an auxiliary edge loss to strengthen the edges in feature maps.Faced with large morphological changes of retinal layers caused by CNV,the attention mechanism is then introduced to extract more context information.Global-Detail Attention Network and Lesion Attention Network are proposed.These two networks make the model pays attention to the both global information of retinal structure in feature map and detail edge information,also the lesion area.The strategies proposed in this paper can effectively segment the CNV areas out as well as related retinal layersThis paper use 60 3D OCT images collected by Shantou University·the Chinese University of Hong Kong and Shantou International Ophthalmology Center using Zeiss OCT and Suzhou Big Vision Medical Technology using Topcon OCT.Each of 3D OCT images has 128 2D slices.Then,7680 2D images are used in total.The experimental results were compared with the popular Deep Learning networks in terms of Dice coefficients,IoU and Accuracy.Quantitative experimental results showed that the proposed method can achieve good performance for OCT image segmentation of objects suffering from CNV.