Article Text
Abstract
Ultra-widefield fluorescein angiography (UWFA) is an emerging imaging modality used to characterise pathology in the retinal vasculature such as microaneurysms (MAs) and vascular leakage. Despite its potential value for diagnosis and disease surveillance, objective quantitative assessment of retinal pathology by UWFA is currently limited because it requires laborious manual segmentation by trained human graders. In this report, we describe a novel fully automated software platform, which segments MAs and leakage areas in native and dewarped UWFA images with retinal vascular disease. Comparison of the algorithm with human grader-generated gold standards demonstrated significant strong correlations for MA and leakage areas (intraclass correlation coefficient (ICC)=0.78–0.87 and ICC=0.70–0.86, respectively, p=2.1×10−7 to 3.5×10–10 and p=7.8×10−6 to 1.3×10–9, respectively). These results suggest the algorithm performs similarly to human graders in MA and leakage segmentation and may be of significant utility in clinical and research settings.
- Imaging
- Retina
- Macula
- Diagnostic tests/Investigation
Statistics from Altmetric.com
Introduction
The recent development of ultra-widefield fluorescein angiography (UWFA) imaging systems has produced images with a 200° field of view1 that can visualise up to 3.2 times more retinal area compared with the conventional 7-standard field imaging implemented by the Early Treatment of Diabetic Retinopathy Study.2 3 With this new technology, significant vascular pathology in retinal disorders can be observed in the retinal periphery, which was previously overlooked with standard field fluorescein angiography (FA). However, objective quantification and automated detection of pathology, including microaneurysms (MAs) and leakage regions, remains out of reach. Currently, quantification requires manual segmentation and remains too laborious to be practical. Consequently, clinical trials and in-office assessment of retinal vascular disease have been unable to use quantification of FA imaging as an evaluative clinical tool or as an endpoint in clinical trials.4–10
In this report, we describe the development and evaluation of a novel fully automated algorithm to detect and quantify MAs and leakage in UWFA images. This technology may provide new opportunities for detection, analysis and evaluation of underlying pathology, disease burden and disease characterisation that may have important clinical and research applications.
Methods
This is an institutional review board-approved image analysis study to retrospectively evaluate UWFA imaging using the Optos 200Tx (Optos, Scotland). A single early phase image and a single late-phase image were selected for MA and leakage analysis. Images included in the study belonged to patients with diabetic retinopathy treated at the Cleveland Clinic. Inclusion criteria for the study were as follows: successful acquisition of 200Tx images, diabetic retinopathy based on clinical examination and angiographic images in both early and late phases. Exclusion criteria for this study include prior laser photocoagulation, poor image quality, media opacity and severe artefacts obscuring view.
Image acquisition and dewarping
Native UWFA images from the Optos 200Tx have inherent warping and peripheral distortion due to the projection of a 3D spherical shape onto a 2D plane. Images were processed by dewarping the standard UWFA images through transformation of the image into a stereographic projection of the eye, as previously described.11 The 3D model was mapped to a 2D stereographic projection by projecting all relevant pixels to a plane through the equator of the eye.11 Both native and dewarped images were analysed with the segmentation platforms.
Automated MA detection platform
An automated segmentation and detection algorithm was developed for MA analysis based on a preliminary training set that used ongoing iterative feedback from expert readers. Following algorithm development, validation testing was performed. The MA algorithm developed used a two-step process for analysis. Candidate MAs were first identified. Initial detection of MAs involved application of a difference of Gaussians routine followed by spectral equalisation on an early phase image (figure 1). Following identification, the candidate objects were refined based on local intensity gradients. The intensity variance around each MA candidate was then evaluated. Objects that have a boundary intensity that is not significantly different from its immediate surrounding are eliminated as candidate MAs. Following refinement, the resulting segmented MAs are counted in a zonal fashion within concentric rings around the optic nerve centroid.
Automated leakage detection platform
Similar to the MA assessment algorithm, an automated leakage segmentation and detection algorithm was developed based on initial iterative feedback from expert readers on an initial preliminary training set. Following initial algorithm development, validation testing was performed. In order to evaluate the time-dependent change in hyperfluorescence that represents leakage, the algorithm uses both the early phase image and late-phase image to provide an assessment of change over time. Following retinal vascular segmentation, early and late image registration is first performed. Registration using Fourier correlation of retinal vascular patterns is used to account for rotation and translation of images during acquisition. Following registration, the late-phase image is ‘flattened’ to remove the intensity gradient from the optic disc to the image periphery (figure 2). The late-phase image is then spectrally enhanced to equalise the intensity of leakage regions and enable application of a fixed threshold for segmentation of candidate leakage areas. Vessels, which are accentuated in the late-phase image by this process, are removed by filtering corresponding vessels in the early phase images using a Gaussian convolution kernel. The segmented leakage regions are pseudo-coloured and superimposed on the original late-phase image.
Quantitative assessment of algorithm performance
MA and leakage were manually segmented and quantified in early phase and late-phase FA images, respectively, by two independent expert graders using ImageJ v1.49. Prior to manual segmentation, graders went through a training course assembled by an expert reader to standardise definitions of MA and leakage. MAs were defined as small circular objects significantly hyperfluorescent compared with surrounding choroidal background in early–mid-phase UWFA images. A minimum fluorescence intensity similar to adjacent vessels was also required. Leakage was defined as a region of increasing hyperfluorescence in size and intensity in the late phase angiogram compared with arteriovenous-transit phase, noted as hyperfluorescent regions in the transformed processed images. Registered MAs and leakage areas were quantified by MA counts and pixel area, respectively, using ImageJ.
Algorithm performance was evaluated against manual segmentation gold standards through correlation analysis. The gold standards for MA and leakage were derived using a two-way analysis of variance approach. Specifically, an estimated individual effect for each human grader was derived by a fitted two-way ANOVA model. Each estimated individual effect was removed from total MA counts or leakage areas before values were averaged among human graders to generate the gold standard. Intraclass correlation coefficients (ICCs) and Pearson’s r (r) between algorithm-defined and gold standard for MA counts and leakage areas were calculated to assess for correlation.
Results
In total, 56 dewarped UWFA images and 56 standard UWFA images were analysed with algorithm segmentation of MA and leakage in standard and dewarped UWFA images. Representative comparisons between automated versus manual MA and leakage segmentation are displayed in figures 3 and 4, respectively. Comparison of raw MA counts and leakage areas between the gold standard set and algorithm confirmed significant direct correlation in both dewarped and native images (figure 5). Specifically in MA detection, the ICC was 0.78 in the dewarped images and 0.87 in the native images (p value=2.1×10–7 and 3.5×10–10, respectively) and the r values were 0.788 and 0.885 for dewarped and native images, respectively (figure 5A, p value = 6.47x10–7 and 4.27×10–10, respectively). In leakage detection, the ICC was 0.70 in the dewarped images and 0.86 in the native images (p value=7.8×10–6 and 1.3×10–9) and the r values were 0.940 and 0.939 for dewarped and native images, respectively (figure 5B, p value=1.24e-13 and 4.53e-13, respectively). Bland-Altman plots demonstrate strong association between the manual and automated methods of quantification in both MA and leakage parameters (figure 5C and D).
Discussion
This study represents a new innovation in fully automated quantitative assessment of MA and leakage in UWFA images. Our results displayed significant correlation between automated segmentation and expert reader-generated gold standards. Until now, FA remains a critical tool for diagnosing and assessing progression of retinal disease in clinical and research settings. Segmentation in MA and leakage is challenging given image artefacts and various media opacities (eg, eyelashes, lenticular opacities). Additional previous reports of automated leakage quantification have assessed retinopathy secondary to malaria and diabetic retinopathy.12 13 One major difference between the technique described in our report and other previously described analysis platforms is the capability for panretinal assessment through the use of the UWFA. In one previous report, the field evaluated was from 50° images obtained by Topcon 50-EX System.12 In the diabetic retinopathy report, image analysis focused on the posterior pole.13. The assessment of global retinal leakage by UWFA using this technology can highlight potentially clinically significant pathology in the retinal periphery and facilitate monitoring of overall disease activity in retinal vascular diseases, such as diabetic retinopathy and inflammatory eye disease.
Challenges encountered with segmentation include image quality issues, media opacities and artefacts (eg, lashes). Other image features that require unique consideration for segmentation include chorioretinal scarring, such as from previous laser photocoagulation. Additional development plans for automated assessment will include zonal-based assessment, higher-order pathology discrimination (eg, perivascular vs generalised leakage) and identification of artefacts and regions of interest. Additional prospective research is needed to better understand the role of automated detection and quantification in outcomes and therapeutic decision making. Our group is currently evaluating this algorithm in two prospective clinical trials to provide robust data for assessment of quantitative angiography in retinal vascular disease (NCT01982435 and NCT02503540). Areas of active research include development of automated quantification of retinal ischaemia, integrative pattern analysis and utilisation of these integrative tools as imaging biomarkers.
References
Footnotes
Contributors All authors contributed to the manuscript and to this project through the analysis of data, gathering of data and revision of the manuscript. JPE and KW both wrote the manuscript and contributed equally to the project and are designated as co-first authors.
Funding NIH/NEIK23-EY022947-01A1 (JPE); Ohio Department of Development TECH-13-059 (JPE, SKS).
Competing interests None declared.
Patient consent This was a retrospective study and did not require consent.
Ethics approval Cleveland Clinic institutional review board.
Provenance and peer review Not commissioned; externally peer reviewed.