Study On Fundus Image Mosaic Technique Of Retinopathy Of Prematurity

Fundus images of preterm infants are an important basis for the diagnosis of retinopathy of prematurity and provide a wealth of information for the screening of neonatal eye disease,the diagnosis of retinal disease and the exploration of pathological mechanisms.Due to the limited field of view of the fundus camera and the limited amount of information available in a single fundus image,ophthalmologists often manually stitch together multiple fundus images of ROP in different orientations to obtain global fundus information to improve diagnostic accuracy,but this practice reduces the efficiency of film review and increases labour and time costs.Based on this,this paper focuses on the alignment and fusion of a set of multiple fundus images from different views to achieve fully automated stitching of retinopathy of prematurity fundus images in order to obtain global fundus information with a large visual field and improve the diagnostic efficiency of the physician.The main points and results of this paper are as follows:(1)To address the problems of blurred and distorted fundus images of retinopathy of prematurity acquired from the ophthalmology hospital,the original images were selected for greyscale processing using the green channel,brightness enhancement using gamma correction,contrast enhancement using Contrast Limited Adaptive Histogram Equalization(CLAHE),and geometric correction for subsequent stitching.(2)To address the problems of motion artifacts and difficulty in extracting eponymous feature points in most retinopathy of prematurity fundus images,this paper proposes a method for aligning retinopathy of prematurity fundus images by combining geometric constraints and robust feature description.The image is first diffused by non-linear filtering and a non-linear phase consistency moment map is generated using the phase consistency technique.A geometric mask constraint strategy was then designed to extract the feature points of the image using the KAZE algorithm.Combined with logarithmic polar coordinates,the feature vectors of the image are computed and repeated iteratively several times to generate the final descriptors.To measure the similarity between the feature vectors,the Euclidean distance is used as a metric and Fast Sample Consensus(FSC)algorithm is used to reject outliers and finally solve the transform model.A comparison with several conventional algorithms shows that the proposed method can achieve higher alignment accuracy.(3)To address the problems of uneven colour and the existence of stitching seams during the fusion of multiple fundus images of retinopathy of prematurity in different orientations,this paper uses an improved pyramid-based Laplace multi-band fusion algorithm to perform weighted fusion of multiple aligned fundus images in order to obtain high-quality fundus image stitching results.The algorithm is able to achieve a natural and smooth transition in the boundary and overlapping regions,thus effectively improving the quality of fundus image stitching.The method proposed in this thesis is robust in terms of feature extraction,feature description and fundus image stitching,and the stitched images that meet the visual effect are obtained through experiments,which is important for assisting doctors in clinical diagnosis to improve the diagnostic efficiency.