Retinal Image Interoperability In Chronic Disease Information System And Its Vessel Network Quantitative Analysis

Background:Chronic diseases have been public health issues for their high incidence, high recurrence, high disability, high mortality, and yearly increasing costs. Currently, there are two important problems that chronic disease information system facing, one is lack of valid image data, and the other is data sharing. Retinal structures deform with the process of chronic disease development. It’s of great significance for their quantification and automatic analysis due to difficulties in accurate illustration and annotation of shapes, texture, curvature, spatial relationships of retinal vessels. Therefore, this study aims to investigate the interoperability of retinal image data in choronic disease information system and methods for quantifing retinal vessel network changes, which could provide automatic analysis for retinal vessel changes in chronic diseases.Methods:Firstly, we constructed a browser/server (B/S) retinal picture archiving and communication system (PACS) based on the digital image communication of medicine (DICOM) 3.0 standards. We designed the DICOM-structured report(SR) in a list and hierarchy way, and visited the patients’retinal image and report data through web acess to DICOM persistant object (WADO) on the basis of HTTP.Secondly, we made image enhancement using undecimated discreted wavelet transform(UWT) filter and clustered retinal image pixels by their grayvalue and textural features with fuzzy C mean clustering method (FCM). Then, we proposed a automatic seed tracing framework to classify segmented vessels into artery and vein to provide basis for further morphology features detection and analysis. In this study, we modeled vessel centerline as stochastic process, detected the feature points in the centerline according to neighbor information in image, and obtained pixel coordinates in each segments between the feature points. After that, we performed second order analysis to model the tortuosity of the vessel centerline and tested our algorithms in constructed curves and real retinal images. For width modeling, we firstly obtained the sectional profiles of tested vessel segments, and selected mixed gaussian curves to fit such profiles by considering the reflex phenomenon on retinal vessels. Then, we calculated the half height full width using differential calculus to obtain the final average width. However, such models have limits in the vessel bifurcations. Instead, we proposed a method based morphology method. In this method, we isolated vessel segments and calculated area, major axis length of the vessels to get mean width values of such segments. And we compared the consistency of above-mentioned width meaments.Finally, we proposed a method to model retinal vessel topology hierachies, and performed different space analysis with our before-mentioned quantification algorithms on images from patients with diabetes and hypertension. We made statistics of retinal vessel network on four squares on the whole image and performed fractal dimension calculation on diabetes. For hypertension, we selected different regions of interest (ROI) in the optic disk (OD) centered image (1.5 DD,3.5DD) to measure width. For diabetes, we made tortuosity analysis and comparison at the different topological hierarchy levels.Results:1. The constructed B/S retinal image PACS could store and communicate DICOM images effectively. Users from hetergenous systems could click on links of the patients’lists through HTTP and WADO service to request and download patients’ related images and reports.2. In this study, the sensitivity of our segmentation method was 0.82, and area under curve (AUC) was 0.9375. The correctness of artery and vein seeds indentification was 80.13%, mean interations during FCM process were 12 and mean time costs were about 2 s. Our proposed tracing algorithm could correctly mark continuous branching vessels and cross vessels. Compared with traditional arc length ratio, our proposed tortuosity models could better reflect real vessel tortuosity after the validation in simulated and real vessels and could be stable in different scale and rotation conditions. Our proposed mixed gaussian curves fitting algorithm could reflect vessel profiles better and our morphology-based method could automatically obtain each vessel segments’ width, and both measurements had high consistency as shown in Pearson analysis(r=0.91).3. Our proposed topology modeling method successfully identified the topological hierarchy levels of retinal vessels. The paired T test showed that the dimension of fraction (Df) was 1.74±0.01 in the control group, while increased in the diabetes group (1.78±0.01, P<0.05). The width measurements in vessels of different ROIs indicated that the arteries’width as well as artery vein diameter ratio (AVR) decreased in the hypertension group compared with that of the control group (both P<0.05). The tortuosity analysis based on different topological hierarchies in macular-centered images from patients with diabetes showed that the tortuosity increased with the growing hierarchy levels and was larger than the control group (P<0.05).Conclusions:1. Our proposed retinal image integration and sharing method could solve the image interoperability issues between hetergenous chronic disease information systems, and our designed DICOM-SR model is of importance for computer treatment and data mining with text and image quantification information in such systems.2. Our proposed retinal vessel network quantification methods could quantify vessels’ tortuosity and width, which provide convinence for medical staff in their description of retinal vessels as well as feature extraction, and is the basis for image-based computer aided diagnosis.3. Our proposed topological hierarchy-based retinal vessel network feature quantification idea is a feasible framework for image information extraction and retinal vessel morpoloical changes investigation in diagnosis and monitering of chronic diseases.