Research On Key Algorithms For Processing And Analysis Of Sequential Images From Fundus Fluorescein Angiography

Fundus fluorescein angiography(FFA)is one of the main diagnosing techniques for fundus diseases.Compared with other techniques,FFA is advantageous in recording the retinal blood flow and static vascular structures.However,numerous FFA sequential images are produced during imaging process.These gray scale FFA images contain rich spatial-temporal information of crisscross arteriovenous network,and the image quality is greatly affected by varying image background,strong fluorescein noise and eye-movement.Hence,it is challenging to develop accurate and fast algorithms for processing and analyzing FFA sequential images.Current preprocessing algorithms focus mainly on single static FFA image,without considering the characteristics of FFA sequential images.Hence,these algorithms show poor robustness and low accuracy in processing FFA sequential images.It results in difficulties in vessel width measurement and quantification of retinal blood flow from FFA images,which is disadvantageous to the accurate diagnosis and quantitative analysis of fundus diseases.With the aim to solve this problem and supported by Hunan Key Laboratory of Intelligent Robot Technology in Electronic Manufacturing,systematical studies on key algorithms for FFA sequential images are performed in the present work,which is of theoretical importance and has promising clinical applications.Preprocessing algorithms for FFA sequential images are firstly proposed in this paper,including a deep learning method for precise and robust vessel segmentation,and a regiongrowing-based method for accurate artery-venous classification.To facilitate accurate diagnosis of fundus diseases,quantitative measuring and analysis methods for FFA sequential images are then proposed,including a vessel width measuring method based on centerline correction and k-means clustering,modelling and analyzing methods for global retinal blood flow.The main contributions and conclusions of the present work are as follows:(1)Vessel segmentation of FFA images is a key preprocessing algorithm for accurate measurement of retinal vessels.However,the varying background and strong noise in FFA images make it difficult to segment vessels with high accuracy.To address this problem,a segmentation method based on multi-path cascaded U-net is proposed.This method can realize accurate vessel segmentation by fusing vessel features from the raw FFA image,an image with enhanced small vessels and an image with enhanced large vessels.The proposed method is tested on a public database and local database,and compared with 18 advanced segmentation methods(6 traditional methods and 12 deep learning methods).The results show that: the proposed method outperforms other methods on evaluation metrics of Se,F1-score,AUC,and can well retain the details of blood vessels.It also shows good robustness in processing FFA sequential images,and can be applied to accurate vessel segmentation of color fundus images.The proposed method lays a foundation for accurate measurement and analysis of retinal vessels.(2)Artery-venous classification of FFA images is a key preprocessing step for quantitative analysis of retinal blood flow.However,FFA images have no color information and contain many vessel crossings,which makes it difficult to classify arteries and veins with high accuracy.To address this problem,an artery-venous classification method based on sequential features and vascular structural features are proposed.The main procedures are:firstly,vascular graphs are established,with which vascular structural features are extracted;secondly,FFA images are aligned by a registration method,with which FFA sequential features are extracted;Then,with the obtained features,initial seeds of arteries and veins are generated.Finally,the obtained seeds are propagated outwards using a region growing strategy,until the entire artery-venous networks are classified.Experimental results on public and local databases show that: the proposed method can achieve a classification accuracy of 0.90.In comparison with 15 advanced classification methods,the proposed method is more advantageous in classifying small arteries and veins.Compared with semi-automatic methods(manual assistance is needed),the proposed method is fully automatic.The proposed method provides guarantee for automatic and accurate analysis of retinal blood flow within arteries and veins,and is more suitable for clinical application.(3)Abnormal morphology of vessels is considered as a precursor of diabetes and hypertension.Accurate measurement of vessel width in FFA images is thus the basis for accurate diagnosis.However,current methods may fail in measuring pathological vessels.To solve this problem,a method combing centerline correction and k-means clustering is proposed for width measurement of pathological vessels.The main procedures are: firstly,vessels are segmented using the multi-path cascaded U-net proposed in(1);secondly,vessel centerlines of pathological vessels are extracted and refined,with which the measuring axis perpendicular to the centerlines and the intensity profile of pixels along the measuring axis can be obtained.After that,vessel boundaries are detected by used of k-means clustering and boundary searching strategy.With the obtained vessel boundaries,the vessel width is finally determined.In order to validate the proposed method,normal vessel segments and pathological vessel segments are randomly selected from public and local databases and 900 measurements are performed.Tests shows that: measured values by using the proposed method are close to those by experts;compared with current methods,the proposed method obtains comparable results for normal blood vessels,but has obvious advantages in measuring pathological vessels.Moreover,the proposed method is also extended in the vessel width measurement for color fundus images,with satisfactory measuring accuracy.The proposed method is hopeful to be applied in quantitative measurement of pathological vessels,during diagnosis of retinal diseases caused by diabetes or hypertension.(4)Quantitative analysis of retinal blood flow is a key problem in the analysis of FFA sequential images.Current methods based on naked-eye observation can not meet the requirement for rapid and quantitative assessment in clinical usages.To solve this problem,a curve fitting method is proposed to describe quantitatively the global retinal blood flow,by use of FFA sequential images.With the obtained fitting model,blood flow parameters can be calculated quickly.The main procedures are: firstly,corrupted FFA images are identified and excluded by approximate nearest neighborhood principles;secondly,perfusion regions are segmented from the FFA images by multi-scale linear filters and then classified as arterial/venous perfusion regions using the classification methods proposed in(2);based on the pixel number in the perfusion region and the time interval of FFA sequential images,the perfusion curves for global retinal vessels,arterial vessels and venous vessels are obtained and fitted by damped least-squares regression.The obtained fitting models are then used to quantitatively describe the perfusing process of fluorescein in the vessels(or retinal blood flow process).Finally,based on the obtained curve fitting models,feature parameters describing the retinal blood flow of entire vessels,arteries,and veins are defined,including perfusion times,mean circulation times,and sectional perfusion rates etc.With these parameters,retinal blood flow process is characterized and can be quantitatively analyzed.Tests of the proposed method by 44 FFA videos(containing 1,3000 images)from public and local databases show that: the curve fitting models can correctly reflect retinal blood flow process,and the extracted feature parameters are close to those measured by manual method.Further analysis indicates that these parameters seem to be correlated with symptoms of fundus diseases such as retinal vein occlusion.Compared with current methods,the proposed method can provide feature parameters for both arterial and venous blood flow,and is thus more advantageous in accurate diagnosis of fundus diseases.Moreover,the proposed method is fully automatic and computing efficient in quantifying the blood flow process within the entire vessels,which is more suitable for clinical applications.