FormalPara Key Summary Points

There is a high and ever-increasing global prevalence of diabetic retinopathy (DR).

The practice of fluorescein angiography to confirm proliferative disease when the clinical examination is not conclusive is invasive and time-consuming.

We explored the ability of ultra-widefield (OCT-A) to differentiate severe from proliferative DR using both clinical and bespoke image analysis, automated measures.

We recruited 75 patients and analyzed 4 × 4 and 10 × 10 mm (widefield) and mosaic (ultra-widefield) OCT-As.

Results demonstrate clinical utility of particularly the ultra-widefield OCT-A to detect proliferative disease and significant differences in capillary density and ischemic areas, demonstrating potential for future image analysis.

Introduction

The World Health Organization regards diabetes as a major global public health problem, with an estimated 422 million [1] people already affected worldwide and an ever-increasing prevalence. One-third of those affected will develop ophthalmic complications and of those, another third will develop sight-threatening complications [2]. Proliferative diabetic retinopathy (PDR) is one of the most severe complications, defined by the growth of abnormal new vessels (NV) anterior to the internal limiting membrane (ILM) [3]. If left untreated, these can lead to vitreous hemorrhage, retinal detachment, or neovascular glaucoma. Treatment can be effective with pan retinal laser photocoagulation or anti-VEGF injections but both are invasive with potentially serious adverse effects [4, 5] and their application therefore needs to be carefully justified, with a clear discrimination of PDR from its precursor of severe non-proliferative diabetic retinopathy.

Currently in clinical practice, neovascularization is often identified through widefield color imaging and clinical examination. However, uncertain cases may require more sensitive means for detecting NV, with widefield fluorescein angiography (FA) [6]. This relies on an intravenous dye injection before imaging the fundus. It is highly effective for the detection of NV, however, because of its invasiveness, known side effects [7], and organizational requirements, it is not practical for it to be part of every diabetic clinic examination.

Widefield optical coherence tomography angiography (OCT-A) appears to show potential as a substitute to some FA investigations [8]. It is a much more rapid and non-invasive imaging tool based on blood flow detection using interferometric detection of low temporal coherence light backscattered by moving blood cells [9]. It allows the depiction of both retinal and choroidal blood vessel networks and with that the detection of ischemic areas or eventual abnormal vessel proliferation. Many studies have evaluated its potential in diabetic retinopathy (DR) assessment [10] demonstrating good detection rates for microaneurysms, intraretinal microvascular abnormalities, and NV.

OCT-A technology enables this detection of retinal NV by showing flow signal anterior to the ILM, which can be used to detect early and subtle NV [11] in a much more accessible way than FA. The main limiting factor for some years has been in the limited size of the field of the retinal image obtained. This limit has been gradually eroded, however, since the development of commercially available widefield OCT-A devices that allow a broader field of view of the retina. The practical assessment of the utility of these modern wide-field OCT-A imaging devices is clearly important to help define their precise clinical role in the care of patients with diabetic retinopathy.

In this prospective observational study, we assess a widely available CE-marked machine that performs widefield (10 × 10 mm) and ultra-widefield (17.5 × 17.5 mm mosaic) OCT-A imaging. We compare features of images acquired using these two modes in the two clinical groups of severe non-proliferative and proliferative diabetic retinopathy, recognizing the particular importance of distinguishing between these diagnostic groups in clinical practice.

We firstly document the clinically observed presence or absence of proliferative vessels to determine if the imaging characteristics permit identification of these disease defining characteristic new vessels. In further assessments, we use dedicated image analysis software to compare precise features of the superficial capillary plexus vasculature in each of these clinical entities for potential insights into the pathophysiology of the two clinical groups and potential markers for proliferative disease.

Methods

Ethical Approval

Ethical approval was obtained from Central Manchester University Hospitals NHS Trust Foundation (Manchester, UK), and NREC local ethics committee. The study was conducted in full conformance with all relevant legal requirements and the principles of the Declaration of Helsinki, Good Clinical Practice, and the UK Policy Framework for Health and Social Care Research (2017). Informed written consent was acquired from all participants.

Study Design

Patient Recruitment and Imaging

This was a prospective observational cohort study. We included patients attending the diabetic ophthalmology clinic of Manchester Royal Eye Hospital (Manchester, UK) between November 2, 2022, and March 8, 2023, who were diagnosed with severe non-proliferative or proliferative diabetic retinopathy. Diagnosis was by specialist clinicians using initially clinical examination and widefield color imaging. For any cases that were not absolutely clear, clinicians had access to other imaging modalities including OCT A and fluorescein angiography to confirm their diagnosis. For this study, our exclusion criteria were a history of panretinal photocoagulation (PRP) in that eye, an anti-VEGF injection in the last 3 months or any condition affecting the patient’s compliance or understanding. We included one eye per patient. If both eyes were eligible, the eye with best imaging quality was included.

A review of electronic medical records was performed for all patients to record systemic and ocular parameters. For systemic parameters, we recorded age, gender, duration of DM, type of DM, cardiovascular condition, glycosylated hemoglobin if they had recent blood tests (less than 4 months). For ocular parameters, we recorded best-corrected distance visual acuity (BCVA, LogMAR scale), ocular comorbidity, and the level of media transparency, including lens status.

We used the CANON Xephilio OCT-A1 machine for all imaging. Our aim was to explore retinal findings in the highest possible resolution and largest possible retinal field achievable for the greatest number of patients. However, resolution may be reduced with increasing field and we therefore used multiple imaging modes to provide the greatest level of detail. We began by assessing single-shot 4 × 4 and 10 × 10 mm images and if we were able to acquire these images and the patient was agreeable and still comfortable to continue, we proceeded to acquiring the multiple images required for mosaic 17.5 × 17.5 mm imaging. We therefore ultimately planned to study three sets of images—4 × 4 mm (highest resolution to study foveal architecture), 10 × 10 mm-wide-field imaging, and finally 17.5 × 17.5 mm ultra-widefield mosaic (widest field of view to explore peripheral retinal changes).

For this study, we assessed superficial plexus images only. We extracted one mosaic for each eye, obtained by reconstruction on the Canon software, using four to five 10 × 10-mm images. All images were taken post dilation. Images were only accepted if quality was rated by the machine as level 3 or above and were exported without additional processing as bitmap images for analysis [12]. We used one eye per patient, which was the image with the highest quality, and if this was equal, the eye with the worst DR was entered into the study.

Presence of Definitive Proliferative Vessels

We defined OCT-A presence of proliferative disease by expert clinical examination as described by Russel et al. [13]; any abnormal vasculature located anterior to internal limiting membrane in a patient’s OCT-A images led to their classification as proliferative vasculature and the presence or absence of any such change was recorded for each patient.

Image Analysis

The images obtained were analyzed using custom-built MATLAB software (MATLAB version: 9.10.0.1602886 (R2021a) 2021; The MathWorks Inc., Natick, MA, USA) and the image processing Toolbox (Optimization Toolbox version: Version 11.3 (R2021a) 2021; The MathWorks Inc., Natick, MA, USA). The code automatically outputted the results of its analysis into an Excel (Microsoft) spreadsheet. The core functions of this software have previously been published and validated [14]. One version was used for 4 × 4 mm central/10 × 10 mm widefield images and another for 17.5 × 17.5 mm ultra-widefield mosaic images due to differences in image quality and characteristics.

Single-Shot Central and Widefield Images (4 × 4 and 10 × 10 mm)

We used the previously described dedicated software algorithm, which allowed input of 4 × 4 mm and 10 × 10 mm images and performed a combined analysis of these images. For each image, the user confirmed foveal center, area of optic disc, and any artefactual areas that were noted during imaging such as masking from apparent vitreous opacities. The software then excluded these areas and continued the analysis steps completely automatically, outputting data into an Excel (Microsoft) spreadsheet. Further details of algorithmic steps can be found in previous publications and in the supplementary materials (S1).

Mosaic Ultra-Widefield Images (17.5 × 17.5 mm)

We used a separate custom-designed algorithm for analysis of these images to account for the more irregular image and greater artefactual content of these images. Initial stages were similar to previous analyses, as there was an initial manual input to allow for cropping out of any evident artefacts and to select the center of the fovea. The subsequent analysis algorithms were adjusted, in particular using a Frangi filter 2D [15] to automatically segment vessels. This function computes the likeliness of an image region to contain vessels or other image ridges, according to the method described by Frangi [16]. It then segments the largest vessels from the capillaries and the ischemic areas based on their size and intensity. We then further segmented the images to isolate the most peripheral areas, excluding the posterior pole, defined by a circle of 400 pixels of diameter, centered on the fovea, as this was the area in these mosaic ultra-widefield images that we were most interested in obtaining information from. To ensure validity was maintained with these new analysis functions, the automated analyses were repeated manually by experienced ophthalmologists. The results concurred with those of automated analysis (supplementary materials (S2)).

The software algorithms we developed were able to segment different zones out from the center of the fovea and analyze them separately as well as perform multiple analyses on separately segmented capillaries and larger vessels (Fig. 1). However, for simplicity, in this paper we present the key features that were deemed a priori as most clinically appropriate based on evidence in related literature. Full analyses of these separate regions did not reveal any additional significant findings and can be found in the supplementary materials (S3).

Fig. 1
figure 1

Processed single-shot images a original 4 × 4 mm, b binarized 4 × 4 mm, c vessels and FAZ automatic segmentation on 4 × 4 mm, d original 10 × 10 mm, e binarized 10 × 10 mm, f vessels automatic segmentation on 10 × 10 mm after artefacts and optic disc removal

The analyzed characteristics of the images are presented below with brief descriptions:

4 × 4-mm image:

  • FAZ area; Foveal avascular zone area

  • FAZ circularity; Foveal avascular zone circularity

  • Mean vessel density; segmented vessel area/ total area of tissue

  • Mean vessel intensity; mean pixel intensity of areas defined as capillaries.

10 × 10-mm widefield image and 17.7 × 17.5 ultra-widefield image:

  • Mean vessel density;

  • Mean vessel intensity;

  • Total area of ischemia; combined area in pixels of all regions defined as ischemic

  • Number of ischemic areas; number of discrete regions identified as representing ischemic zones.

Statistical Analysis

We calculated and estimated the sample size for this study using the key metric mean capillary intensity from previous work on narrow field OCT-A. Using sample size calculation (details in supplementary materials S2) we obtained a result of 24. Our objective of recruitment was then a minimum of 25 patients in each group.

All statistical analyses were performed using JAMOVI software (The Jamovi Project (2022), Jamovi (Version 2.3)). The normal distribution of data was first assessed using the Shapiro–Wilk test. Depending on this result, either t tests or Mann–Whitney tests were used to compare the two DR groups. For qualitative patients’ characteristics, we used the Chi-square test. For secondary statistical analysis, we applied the Bonferroni correction. To take into account multiple statistical analyses, the p value cut-off was set to 0.0046 (Fig. 2).

Fig. 2
figure 2

Example of the algorithms for analysis of ultra-widefield images. On the left is the original image and on the right is the processed image with larger vessels showed in red, capillaries in green, and ischemic areas in blue with a cropped center

4 × 4 and 10 × 10-mm Widefield Images

We performed a one-sided test with the alternative hypothesis Ha that proliferative group would have lower values of capillary intensity than the severe non-proliferative group. We also performed other secondary analyses of the other metrics and patients’ characteristics.

17.7 × 17.5-mm Ultra-Widefield Images

We performed a two-sided analysis, with the null hypothesis that patients with proliferative retinopathy would have the same amount of ischemia as those with severe non-proliferative retinopathy.

Results

Baseline Patient Clinical Characteristics

4 × 4 and 10 × 10-mm Single-Shot Widefield Images (Tables 1 and 2)

Table 1 Description of epidemiological data on patients’ and scans’ characteristics and analyses—4 × 4 and 10 × 10-mm images (N number of patients, BCVA best-corrected visual acuity, SD standard deviation)
Table 2 Further description of epidemiological data on qualitative patients’ characteristics and analysis – 4 × 4 and 10 × 10-mm images

For these single-shot images, 75 out of 102 consented patients were able to be imaged successfully. All these images were subsequently able to be analyzed with the software.

Of the recruited 75 patients, 28 were in the proliferative group and 47 in the severe group. Seventeen of the proliferative patients had confirmatory FFA and 17 of the severe non-proliferatives. Tables 1 and 2 show the description and analysis of patients’ and scan characteristics. The only significant difference we found was for the BCVA, which was lower in the proliferative group (0.304 ± 0.2 vs. 0.152 ± 0.2). Mean age was close to 55 years in both groups. We had missing values in both groups for Hba1c, but mean value was 79.3 ± 18.4 in proliferative group and 77.2 ± 23.1 in the severe group. Most scans had an image quality of 6–7/10 in both groups with no significant difference. We recruited more male than females overall (50 vs. 25) but they were equally distributed in both groups.

17.5 × 17.5-mm Mosaic Ultra-Widefield Images (Tables 3 and 4)

Table 3 Epidemiological description of patients’ and scans’ characteristics and analysis – ultra-widefield images (N Number of patients, SD standard deviation)
Table 4 Further description of qualitative patients’ characteristics and analysis – ultra-widefield images

Fifty-one out of seventy-five consented patients were successfully imaged with the multiple shots required to produce mosaic images. Among those, 40 had good enough quality to undergo semi-automated analysis.

Of the 40 patients included in the semi-automated analysis, 16 had PDR and 24 had severe DR. Thirteen of the proliferative patients had confirmatory FFA and six of the severe non-proliferatives. Tables 3 and 4 show the description and analysis of patients’ characteristics and also scans’ quality. We found significant differences on best-corrected visual acuity with lower vision in the proliferative group (0.347 ± 0.2 vs. 0.156 ± 0.1). The PDR group also had a lower male proportion compared to the severe DR group (44 vs. 83%). The other analyses did not show any significant differences.

Outcome Measures

Presence of Definitive Proliferative Vessels

Patients with PDR were defined as such through clinical expert diagnosis using clinical exam and all required modes of imaging including widefield fluorescein angiography that was often performed. Our results show that 25% of all patients with proliferative disease (7/28) had clear signs of neovascularization on the single-shot 10 × 10-mm widefield OCT-A image. No patients in the non-proliferative groups had such signs (0/47).

For the ultra-widefield 17.5 × 17.5-mm mosaic imaging, 100% of patients (16/16) identified as having proliferative disease also had clear evidence of proliferative vasculature on the wider field mosaic examination images. Again, no patients in the non-proliferative groups had such signs (0/35).

Image Analysis Comparison Between Proliferative and Severe Groups

Single-shot, 4 × 4 and 10 × 10-mm widefield images (Table 5)

Table 5 Description and statistical analysis of the image analysis metrics obtained from single-shot 4 × 4 and 10 × 10-mm images (N number of patients, FAZ foveal avascular zone, SD standard deviation)

The FAZ area measured in pixels was lower at 1847.3 ± 341.9 for the proliferative group versus 1931.9 ± 479.1 for the severe group but it was not statistically significant. There were no significant differences in the vessel area, density or foveal avascular zone circularity.

There were no significant differences for our main assessment in intensity of capillaries - 131.4 ± 5.4 in the proliferative group versus 130.7 ± 5.0 in the severe group. We did not find any significant differences regarding the other key vessels metric of density.

There were no differences also in the number of ischemic areas (2.679 ± 1 vs. 2.596 ± 1; p value 0.428) or average size of the ischemic areas (810.4 ± 392 vs. 1460.0 ± 332.5; p value 0.512) between proliferative and severe non-proliferative cases.

17.7 × 17.5-mm Ultra-Widefield Mosaic Images (Table 6)

Table 6 Description and analysis of the metrics obtained from ultra-widefield images (N number of patients, SD standard deviation)

For the mosaic images, we found no significant difference in vessels density: 23.5% ± 2.8 in the proliferative group versus 21.7 ± 3.0 in the severe group, with a p value of 0.07.

There were, however, significant differences in the capillary density in patients with proliferative disease compared to severe non-proliferative disease: 9.1% ± 1.9 in the proliferative group versus 11.0% ± 1.9 for severe group.

Also, we found a significant difference in the metrics of ischemic areas; average area of ischemic zones [253,930.1 ± 108,636 for the proliferative group versus 149,104.2 ± 55,101.8 for the severe group, see violin plot (Fig. 3)] and density of ischemia (8.9 ± 3.9 for proliferative group versus 5.3 ± 2.0 for severe group) were both significantly higher in patients with proliferative disease.

Fig. 3
figure 3

Average area of ischemic zones (in pixels)—analysis of ultra-widefield images

One patient from the PDR group was removed from the final analysis because it had extreme measurements, which led to a final number of 39 mosaics, and among them, 15 PDR and 24 severe.

Discussion

In this study, we used modern widefield and ultra-widefield mosaic OCT-A imaging to investigate the differences between patients who had been clinically diagnosed with severe DR, compared to PDR. We assessed a range of OCT-A image fields of view and firstly recorded if the characteristics of each image type were large enough to permit identification of defining characteristic new vessels. We then determined image analysis characteristics for each of the image sizes to assess if there were significant differences between the two clinical patient groups.

It should be noted that not all patients were able to undergo the mosaic imaging for different reasons. Firstly, because it involved taking four extra pictures for the research purposes, some patients refused the examination. Secondly, some patients, with poor vision or difficulties understanding instructions, were not able to stare at the moving target. Thirdly, some were able to undergo the imaging, but because of media opacity or eye movements, the poor quality of the images did not allow a semi-automated analysis leading to their exclusion.

We found firstly that the defining neovascular tissues were seen in 25% of OCT-A images with 10 × 10-mm size field (7/28 cases), with no false-positive areas identified. However, this low detection rate increased to 100% of cases (16/16) when we used ultra-widefield mosaic images to assess patients, again with no false-positive cases. The reasons for this disparity appear to be purely due to the wider field of imaging reaching the peripheral location’s new vessels.

This finding is consistent with an emerging body of literature [17,18,19] using similar modern widefield modalities and has important clinical implications. Although the gold standard for detecting proliferative diabetic retinopathy remains slit-lamp examination or widefield color imaging, the more invasive and time-consuming fluorescein angiography is used when doubt persists as a more sensitive and specific method to screen for this condition and was used for many of our patients. In our practice, an OCT-A needed about 15 min to be performed and computed into a mosaic image whereas FA needed at least 30 min including preassessment. Also, whereas FA needs a nurse to inject the dye as well as an imaging technician, OCT-A only required an imaging technician and so also does not require a separate clinic visit.

With limited numbers in this study and not all patients receiving fluorescein angiography, we cannot say with confidence that absence of neovascularization on even ultra-widefield OCT A will always exclude the possibility of neovascularization.

Our results do suggest, however, that in clinical practice widefield OCT-A imaging will often be able to confirm neovascularization, negating the need for fluorescein angiography and could therefore be a useful addition to the diabetic clinic. In particular, the wider the field the greater the likelihood of detection of any neovascularization with modern mosaic imaging appearing particularly effective demonstrating detection of NV in 100% of our cases assessed with no false positives. As the imaging modality is relatively quick and is non-invasive, clinical use of even the 10 × 10-mm single-shot imaging would likely still reveal the NV in a significant percentage of proliferative patients, obviating the need for them to have invasive angiography.

In terms of image processing and analysis metrics, our study did not show any significant differences between OCT-A metrics of vessel intensity or density, nor of ischemic area indices between severe and proliferative diabetic retinopathy on the 4 × 4 and 10 × 10-mm widefield images.

However, in our analysis of mosaic images, we showed significant differences between the two groups on ischemic area assessment metrics of number of areas and average area. This significant difference was not found in the widefield field scans of 10 × 10 mm, confirming that the key areas of ischemia found with progression to proliferative disease are often the more peripheral zones. Although there are high numbers of ischemic areas detected by software in both groups, only the ultra-widefield mosaic periphery images showed significant difference.

Our findings are not easily comparable to other published works as publications in this field often differ in terms not only of patient populations, but also imaging depths (superficial or deep capillary plexus), imaging field sizes, machines used and image analysis algorithms and metrics. Some of the authors used regular OCT-A of 6 × 6 and others used widefield OCT-A of up to 12 × 12 mm. Those metrics were analyzed sometimes after dividing posterior pole in different zones, up to 25 for Cunha-Vaz [20]. Many vessel metrics have been described in the literature [20,21,22,23,24,25,26,27,28], and we have used similar measures in our studies. Garg et al. [22] analyzed moderate and severe DR but did not find significant differences compared to a PDR group. Other groups confirmed assessments of superficial capillary plexus to be at least as relevant as deep capillary plexus [28, 29]. Ryu et al. [29] developed a deep learning algorithm (Convolutional neural network, CNN) able to classify DR on OCT-A, and achieved some accuracy for detecting severe (0.871) and proliferative (0.913) disease on 6 × 6 mm images. However, those results should be interpreted with caution as such accuracy values can be artificially inflated depending on the proportions of the disease in the database used for evaluation. Dong et al. [30] also published similar work, also using a multibranch CNN on widefield OCT-A images. They obtained an accuracy of 90.56% on DR staging, putting moderate and severe DR in the same group.

Limitations of our study include low numbers of patients, particularly in the ultra-widefield group. A high proportion of patients recruited were not able to be imaged with the OCT-A – this was due to a combination of issues that may have included machine requirements, poor dilation, and view due to incipient cataract, as these were not automatically excluded from the study as many patients with these conditions could still achieve adequate imaging. These statistics are likely to improve with newer imaging machines and dedicated imaging staff. Although the mosaic data showed significant differences for many measured outcomes, those that did not reach significant should be taken with caution as we were not able to conduct sample size calculations for this group as we had no prior data using this imaging.

As this study was observational, we relied on everyday clinical practice for the diagnosis of severe versus PDR. We did not have FA confirmation for all diagnoses and on occasion depended upon only clinical examination and other imaging modalities. This may have been prone in theory to potential misdiagnosis, but is representative of real-life ophthalmologic gold-standard clinical management of those patients.

One aim of our study was to assess if there were any early warning signs of features of high-risk individuals who might not yet have proliferative disease but be at risk of it. Our results from the mosaic ultra-widefield imaging demonstrate that measures such as wider field ischemic area have the potential to be used in this way. Figure 3 demonstrates that although large areas of ischemia are present in severe non-proliferative and proliferative, after a certain point, there is a far higher likelihood of the patient having proliferative disease. However, for this information to be used to determine which severe non proliferative patients are likely to progress to proliferative, we would need a prospective study assessing high-risk characteristics and monitoring which patients progress over time from severe non-proliferative to proliferative.

Our results are promising, as they add to the existing literature suggesting that there is currently available technology that shows promise to become a standard imaging exam for identifying a high proportion of proliferative DR. The main advantage of ultra-widefield OCT-A over fluorescein angiography is its non-invasiveness and repeatability without contraindication or adverse effects allowing routine assessments. With more prospective research studies, image analysis algorithms could potentially identify those patients who were at highest risk of becoming proliferative. It should be noted from our study that we still had to exclude some patients due to inability to take adequate quality images, but this is likely to improve with time. In those cases, FA will still need to be performed.

Conclusions

Our study assesses image analysis and clinical assessment of widefield and ultra-widefield OCT-A metrics for the evaluation of severe and proliferative diabetic retinopathy. We demonstrate that the mosaic ultra-widefield (17.5 × 17.5 mm) OCT-A in particular has a high sensitivity for detecting proliferative disease and our studies suggest image analysis of aspects such as ischemia levels may be useful in identifying higher-risk groups as a warning sign for future conversion to neovascularization.