Research On The Generation And Leakage Segmentation Of High Resolution Fundus Fluorescein Angiography Images

Fluorescein angiography is a diagnostic method used to evaluate ocular diseases by injecting fluorescein dye and observing its distribution and dynamic changes in the retinal circulation.Despite being considered the "gold standard" for the diagnosis of retinal diseases,FA is known to have certain adverse effects and is contraindicated in patients with conditions such as hypertension and heart disease.Therefore,the generation of highresolution fluorescein angiography images through image generation algorithms from retinal structure images or leakage and segmentation masks has significant practical implications in the prevention,auxiliary diagnosis,and guidance of retinal-related diseases.Currently,high-resolution generation is one of the most active research directions in the field of image generation.This thesis focuses on two widely used applications of deep learning in FFA images:image generation and leakage segmentation.Specifically,the following questions were addressed:(1)Generation of high-resolution fluorescein angiography images from retinal structure images: Given the potential risks associated with FA and the limitations of existing methods in generating high-resolution images(typically 256x256 or lower),we proposed a retinal fluorescein angiography image generation algorithm based on patching and temporal sequences.Compared to existing methods,our approach can generate images with a resolution of 768x768,with a structural similarity index of 0.7126.Our algorithm performs better in generating vessel structures and leakage details and has promising clinical diagnostic applications.(2)Generation of high-resolution fluorescein angiography images from leakage and vessel masks: To address the challenge of the lack of training images in medical deep learning,we proposed a generation algorithm based on progressive learning.Our approach gradually learns to generate higher resolution images while introducing designed vessel and leakage perception loss functions to assist in generating retinal vessel details and leakage information.Experimental results show that our generated images have a high degree of similarity with real images,with a discriminative distance of 31.3.(3)Leakage segmentation in fluorescein angiography: To address the timeconsuming and inefficient manual segmentation of leakage in clinical practice,we proposed a generative image-driven fluorescein angiography leakage segmentation algorithm.The generated images were validated using both supervised and unsupervised learning,and experimental results show that the similarity coefficient of the leakage segmentation results with manual segmentation is 0.881,demonstrating that data augmentation using generated images can effectively improve leakage segmentation performance.In summary,this thesis investigated two applications of deep learning in fluorescein angiography images,providing a foundation for subsequent practical applications.