Research On OCT Image Segmentation And Generation Based On GA

The retinal Spectral domain optical coherence tomography(SD-OCT)image is obtained by three-dimensional sampling,with high resolution and no damage to human tissues during the acquisition process.It can provide a multi-view human eye’s fine structure of the retina,which can be applied to neurosensory retinal detachment(NRD)segmentation and other related research on retinopathy.OCTA(Optical Coherence Tomography Angiography)detects the movement of red blood cells by detecting changes in the OCT signal of the same cross-section.In the formed image,the signal intensity is correlated with the blood flow intensity at that location,providing more detailed blood flow information than OCT images.With the development of the adversarial generation network,the field of image segmentation has also carried out research on the use of new technologies for efficient and accurate positioning and segmentation of fundus lesions.Aiming at the current problems in OCT images and existing image segmentation models,this paper combines the characteristics of NRD lesions to carry out research on the use of adversarial generation network to locate and segment NRD,and rely on the ability of the adversarial generation network to transfer the style of the OCT image.The specific work is as follows:(1)A new Cycle GAN-based method for segmenting NRD lesions in SD-OCT images is proposed,and Res-UNet is introduced into Cycle GAN to enhance the fitting ability of the generator.A weighted attention mechanism is introduced to enhance the perception of features of the region of interest,and at the same time weaken the interference of invalid region features on the network,effectively improve the extraction of target features,and make the segmentation effect of the model more ideal.This paper compares the proposed method with four classic methods,verifies the feasibility and effectiveness of the model from multiple indicators,and draws a conclusion that the method proposed in this chapter has higher accuracy for the NRD segmentation results in SD-OCT images.(2)A hierarchical multi-scale fusion module embedded in pix2 pix network is proposed for the transform of OCT image to OCTA image and the feasibility and effectiveness of transform OCT images rich in retinal hierarchical structure to OCTA images rich in blood flow information are verified.The transform result shows the superiority of OCTA image compared to OCT image in terms of resolution and vascular blood flow imaging.Compared with traditional GAN,pix2 pix network,Cycle GAN and the optimized UNet network,multiple OCT to OCTA image transform experiments are compared.The experimental results of the deep,shallow and full-level projections are compared with analysis,qualitatively and quantitatively verifies the good performance of the proposed method in this transform experiment.(3)An OCT image segmentation and conversion system based on the generative adversarial network was designed and developed,and all the aforementioned experimental contents were integrated and applied.Encapsulates the NRD segmentation model and the OCTA image conversion model,carries out a modular design,provides an interactive application program interface,and visualizes and evaluates the experimental results.This system intuitively displays the research content and results of this thesis from the perspective of application,expands the content display of intermediate processes and comparison methods from the perspective of easy analysis,and designs multiple user-friendly modules from the perspective of ease of use.