Choroid Neovascularization Growth Prediction Based On A Hybrid Model

The choroid is the vascular layer of the eye,lying between the retina and the sclera,and containing connective tissues.It also provides oxygen and nourishment to the outer layers of the retina.Choroidal neovascularization(CNV)is a proliferating blood vessel that grows between the innermost choroid and the retinal pigment epithelium.It comes from the choriocapillaris,which proliferates and spreads through cracks in the Bruch membrane.Choroidal neovascularization mostly occurs in the macular area,so it will do great harm to central vision.In the early stage of disease screening,if earlier stage of choroidal neovascularization is found through retinal imaging technology,the growth of the choroidal neovascularization can be predicted through modeling.Then timely clinical measures can be taken at the perfect time to help treat the disease and prevent further deterioration.During the period of disease treatment,predicting growth of choroidal neovascularization can guide therapeutic schedules and help doctors to implement personalized treatment for different patients.In the later stages of the disease recovery,prediction of reoccurrence of choroidal neovascularization can provide essential information and give reference and advice for subsequent treatment.In this thesis,we propose a noninvasive method based on ocular physiology to predict the growth of CNV from three-dimensional optical coherence tomography(OCT)images through a hybrid model.The method consists of three steps: preprocessing,CNV growth modeling and prediction.The preprocessing part mainly involves image enhancement,denoising,land-mark labeling,and image registration of the region of interest(ROI).In growth modeling part,we present a new hybrid model that combines the hyperelastic biomechanical model and the reaction-diffusion model with therapeutic factors through mass effect equation.Finally,prediction is achieved by optimization of the parameters.For parameter optimization,the genetic algorithm is applied to iteratively calculate the prediction parameters at each time point.Then,curve fitting is used to integrate parameters of all time points.The proposed method was tested on a dataset with 6 objects,each with 12 longitudinal 3D images.The experimental results showed that the average true positive volume fraction(TPVF),false positive fraction(FPVF)and Dice can reach 80.0±7.62%,23.4±8.36% and 78.9±7.54% respectively.