Segmentation Of Retinal Layers With Macular Hole And Cystoid Edema In OCT Images

Retina is an important part of the fundus with thin but delicate layer structure.The thickness and intensity changes of each layer are important clinical features for many retinal diseases.Non-invasive optical coherence tomography technology(OCT)provides a guarantee for the automatic segmentation and analysis of retinal layers and lesion areas.Macular hole(MH)is a common retinal disease,which is usually accompanied by macular cystoid edema(CME).The precise segmentation of MH and CME can obtain visual acuity indicators such as optical density and defect length.Due to the variety of MH and CME,the structure of the retinal layer is changed greatly,which greatly increases the challenge of automatic segmentation of the retinal layer.In this thesis,a retinal layer segmentation framework based on convolutional neural network(CNN)and 2D dynamic constraint graph search(2D-DCGS)is proposed for the automatic segmentation of retinal layer and MH and CME lesion region in retinal OCT image,in which the MH and CME segmentation results based on CNN are used to guide the retinal layer edge detection based on 2D-DCGS.The main steps include:(1)Because the intensity of MH is similar to the background,it is difficult to locate the up boundary of MH.An auxiliary network with a residual module is proposed to locate the left and right vertices of the up boundary of MH.(2)A multi-object segmentation network based on mixed downsampling module(MDM)and self-entropy loss function(SE Loss)is proposed to segment MH and CME simultaneously.(3)The anisotropic diffusion filtering(ADF)algorithm is used to suppress the speckle noise.(4)The MH and CME segmentation results are used to guide the multi-scale 2D-DCGS to obtain the retinal layer boundaries.Twenty retinal OCT images(12 lines radial scan mode)with idiopathic MH(with coexistence of CME)provided by Shantou International Ophthalmology Center were used for the performance evaluation of the proposed method.The average Dice coefficient for the segmentation of MH and CME reached 94.59%and 84.06%respectively,while the Dice correlation coefficient and the mean absolute distance(MAD)of the retinal layer segmentation were 89.87%and 4.58 μm,respectively.Compared with other methods,both the performances of segmentation of MH and CME and segmentation of retinal layer have been greatly improved,which indicates that the proposed framework is feasible and effective.