Research On Fovea Location Based On Deep Learning

The macula is an important region in the central part of the retina responsible for central vision function.Macular lesions can lead to issues such as vision loss,color vision impairment,and loss of depth perception.Therefore,accurate diagnosis and treatment of macular lesions are crucial for ocular health.The central fovea plays a crucial role in assessing macular function and diagnosing macular lesions,while also influencing the positioning and measurement of other fundus structures.This article addresses the issues of insufficient positioning accuracy and difficulty in locating lesion samples in current central fovea positioning algorithms.Two central fovea positioning models are designed,and the following main contributions are achieved:(1)A multi-module localization model is established based on traditional threshold segmentation and deep learning-based semantic segmentation methods.The model includes a binarization processing module(BPM),which calculates the threshold by utilizing the area,histogram information,and optic disc radius information of local images,and optimizes the segmentation results through morphological techniques.Simultaneously,a segmentation result determination module(SREM)is designed to evaluate the positioning results of the BPM module,filter out samples with failed positioning,and pass them to the Unet segmentation module for processing,achieving a mutually corrective structure between the BPM and Unet modules.Experimental results on the Messidor dataset demonstrate that the proposed model can effectively locate the central fovea of the macula with high efficiency and accuracy.(2)Based on image generation and image restoration techniques,an end-to-end model is established.The image restoration technique is utilized to address the issue of insufficient localization accuracy caused by macular feature loss and lesion interference.In the model,a global scene information inpainting module(GSIIM)and a local scene information inpainting module(LSIIM)are designed based on keyvalue memory networks.The GSIIM module integrates global information from template images into the image to be inpainted,improving the overall consistency of the inpainting results.The LSIIM module incorporates features from non-masked regions into masked regions,enhancing the detail information of the inpainted image.Experimental results on the Messidor dataset and IDRi D dataset both achieve the current best accuracy,indicating that the macular fovea positioning model based on image inpainting can effectively handle highly lesioned samples.