Introduction

Diabetic retinopathy (DR) is a leading cause of blindness [1, 2]. By 2030, the World Health Organization estimates the prevalence of DM to reach 366 million adult patients and DR is estimated to affect 191 million [2]. This emphasizes the need for appropriate and early screening and, thus, the screening programs need to become as effective as possible [3]. Also, patient flow needs to be optimized in clinics with obtention of more information from less testing or without moving patients to different machines and/or rooms, which can be very time-consuming. This has become even more significant with the COVID-19 pandemic which has created a new healthcare reality.

MultiColor scanning laser imaging (MC) is a recent non-invasive modality which uses 3 laser wavelengths simultaneously to obtain high-resolution en face images of the retina. The blue, green, and infrared wavelengths have different penetration ranges and thus have the ability to show structures at different depths within the retina. The blue wavelength is more useful to study the vitreoretinal interface and inner retina; the green wavelength shows details of the retinal vasculature and intraretinal features such as blood and exudation and the infrared wavelength has a higher penetration and improves visualization at the level of the outer retina, choroid and retinal pigment epithelium [4, 5].

MC presents some advantages when compared to conventional fundus examination and color fundus photography (CFP), namely (1) a superior penetration through media opacities such as cataract, (2) a better performance in undilated pupils, (3) the confocal technology blockage of scattered light offers higher resolution and higher contrast, (4) the eye-tracking technology of image stability and averaging acquisition improves resolution and (5) it is incorporated in an optical coherence tomography (OCT)/OCT and fluorescein angiography (FA) platform, not requiring the need to move to a different machine [4]. As such, this new imaging modality can identify DR pathologic changes, such as retinal neovascularization in PDR [4, 6,7,8,9,10].

PDR is an important cause of vision loss in diabetic individuals and is characterized by the presence of neovascular complexes (NVC) on the optic disc (NVD) or elsewhere in the retina (NVE) [11, 12]. Although there have been significant advances in imaging techniques to evaluate PDR [13, 14], FA, described in 1961 [15], is considered the gold standard to classify disease activity, guide laser therapy and evaluate treatment response. Nonetheless, it is a time-consuming invasive technique that involves intravenous administration of a fluorescein dye to obtain a vascular map of the eye and, as such, cannot be performed at every visit [16,17,18]. In this situation, MC may be advantageous as it can reveal retinal neovascularization undetected by CFP or funduscopy by the use of the green laser wavelength [4, 6,7,8,9]. Also, when compared to FA, MC might be a good alternative in patients with poor quality FA or in which FA is contraindicated, due to its noninvasive character and higher penetration rate in opacities.

In this study, we evaluated the accuracy of MultiColor imaging in detecting PDR and associated DR features.

Methods

This retrospective case series included eyes from patients with PDR from the medical retina clinic at Hospital de Santa Maria, Lisbon, Portugal. This study was approved by the local ethics committee and was conducted according to the tenets of the Declaration of Helsinki. All subjects provided informed consent. Patients were identified using a database search and subsequent chart review. Exclusion criteria included other etiologies of proliferative retinopathy or reduced imaging quality due to significant media opacities.

Clinical and demographic data were collected along with same-day FA and MC images. It is our Institution’s protocol to obtain MC images when performing FA. All MC images were independently evaluated by 2 trained masked medical retina specialists—SVP and GDS—with ambiguities resolved by open adjudication to decide the final grading based on previous publications on MC imaging of DR lesions [4, 6,7,8,9,10, 19]. The FA grading was performed openly according to the ETDRS classification from fluorescein angiograms [20, 21]. Retinal neovascularization was classified as NVE and disc neovascularization or within 1 disc diameter from its margin as NVD [22]. Other imaging features classified were: presence of round artifact [23], microaneurysm (MA), intraretinal hemorrhage (IRH), vitreous hemorrhage (VH), preretinal hemorrhage (PRH), fibrosis, hard exudates (HE), diabetic macular edema (DME), epiretinal membrane (ERM), ischemia and laser spots (LS).

High-resolution digital MC images with a field of view of 55° were obtained using Spectralis HRA with MultiColor (Heidelberg Engineering, Heidelberg, Germany). Images were captured at the time of FA and retrospectively reviewed. An initial quality check was performed and only cases with at least 10 averaged scans were included. Images were analyzed using the Heidelberg Eye Explorer (version 1.10.4.0) and the software tools were modified for improved visualization of the neovascularization in MC, respectively the color balance was changed to green–blue enhances, the noise reduction was set to low, and the sharpen function to medium [5].

Statistical analysis was performed using Microsoft Office Excel 2020 for Mac version 16.39 and SPSS version 24.0 (IBM Corp, Armonk, NY). Data were analyzed with frequency and descriptive statistics. Categorical variables were presented as counts and percentages and continuous variables as mean ± standard deviation (SD). For each feature, we calculated the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), through cross-tabulation [24]. Receiver-operating characteristic (ROC) analysis was performed for several features with calculation of area under the curve (AUC) value with the respective 95% confidence interval (95%CI) and p-value for statistical significance compared to reference. AUC was classified as previously suggested [25, 26] and only results over 0.7 (acceptable) were included. Concordance between graders for the qualitative MC features assessed was determined with k index. The strength of agreement was considered fair from 0.21 to 0.40, moderate from 0.41 to 0.60, substantial from 0.61 to 0.80, and almost perfect from 0.81 to 1.00, as previously described [27]. P-value < 0.05 was considered statistically significant.

Results

Fifty-nine eyes from 38 individuals with a confirmed diagnosis of PDR were evaluated. The mean age ± SD was 56.6 ± 11.6, 20 (52.6%) were male, 28 (73.7%) Caucasian and 10 (26.3%) had African origin. Eight (21.1%) patients had type 1 DM and 30 (78.9%) had type 2 DM. Fifteen (39.5%) patients had bilateral disease. The affected eye was the right eye in 31 (52.5%) of cases, 2 (3.4%) eyes were treatment naïve, 42 (72.4%) received laser pan-retinal photocoagulation (PRP), 14 (24.1%) were submitted to a combination of PRP and anti-VEGF therapy and 5 (8.6%) eyes had additional surgery (pars plana vitrectomy).

Sensitivity, specificity, PPV, and NPV of MC in detecting NVC

The detection of eyes with PDR using MultiColor imaging compared to fluorescein angiography is presented in Table 1. Thirty-nine (66.1%) NVC were identified in both MC and FA. MC identified 41 (69.5%) NVC compared to 51 (86.4%) observed in FA (Figs. 1, 2, 3). The sensitivity for the detection of NVC (NVD & NVE) using MC was 95.1%, with a specificity of 40.0%, PPV of 92.9% and NPV of 50.0%, respectively (Table 2). When considering the NVD separately from the NVE, the sensitivity, specificity, PPV, and NPV were 88.9%, 76.9%, 80.0%, and 87.0%, respectively (Table 2). NVD was identified in 30 (50.8%) eyes using MC and in 29 (49.2%) eyes using FA, with 24 (40.7%) being present in both modalities (Figs. 1, 2). For the NVE, sensitivity, specificity, PPV and NPV were, respectively, 78.0%, 63.6%, 88.9%, and 43.8% (Table 2), meaning MC and FA identified 36 (61.0%) and 46 (78.0%) complexes, respectively (Figs. 1, 2, 3). Thirty-two (54.2%) NVE was positively recognized in both techniques (Figs. 1, 2, 3).

Table 1 Detection of eyes with PDR using MultiColor imaging compared to fluorescein angiography
Full size table
Fig. 1
figure 1

Example of 2 cases of high-risk PDR. MC (top left and bottom left) showing multiple areas of NVE corresponding to areas of late leakage on matching fluorescein angiography (top right and bottom right). Note the presence of macular hard exudates in both MC images, that are not apparent in the angiogram, as well as the rounded lesions from panretinal photocoagulation (top left) and the epiretinal membrane and fibrosis along the inferotemporal arcade (bottom left)

Full size image
Fig. 2
figure 2

Evaluation of NVD. MC showing red lesions corresponding to microaneurysms, retinal hemorrhages, and ischemia, with no apparent NVD or NVE (top left). Nevertheless, NVD can be identified in fluorescein angiography (top right), indicating PDR. On the other hand, the 2nd MC image (bottom left) clearly shows the abnormal vessels of NVD along with NVE along the superotemporal arcade, as confirmed by the angiogram (bottom, right). Note the laser scars (bottom left)

Full size image
Fig. 3
figure 3

Example of vitreous and preretinal hemorrhages. Preretinal hemorrhages can be easily identified in MC (top left and bottom left), as well as vitreous hemorrhages (bottom left). Note the macular striation with hard exudates from epiretinal membrane, as well as the fibrovascular NVE proliferating along the inferotemporal arcade (top left). The corresponding fluorescein angiogram (top right) shows the leakage from the NVE, as well as NVD that are not clearly apparent in the MC, but the macular changes are difficult to determine. MC image quality can be limited by the presence of a round artifact (bottom left), but in this case, it was still possible to identify the hemorrhages and abnormal vasculature. Matching angiogram (bottom right) confirms the presence of NVE and the absence of NVD

Full size image
Table 2 Sensitivity and specificity of MultiColor imaging in detecting NVC, NVD and NVE
Full size table

Sensitivity, specificity, PPV and NPV of MC in detecting additional features of PDR

The data of the additional features observed is included in Tables 3 and 4. A round artifact was identified in 55 (93.2%) MC images and in none (0%) of the FA images evaluated (Fig. 3), with a specificity of 6.8% and an NPV of 100%. MA was observed in 49 (83.1%) eyes using MC and in 58 (98.3%) with FA, for a sensitivity of 98.0% and PPV of 100.0% (Fig. 2). IRH was more frequently observed in MC—40 (67.8%) versus 36 (61.0%) using FA (Fig. 2), with a sensitivity of 100.0%, specificity of 76.9%, PPV of 91.7%, and NPV of 100%. The grading of VH was quite similar between the modalities, 6 (10.2%) eyes in MC and 8 (13.6%) in FA (Fig. 3), with a specificity, specificity, PPV, and NPV of 62.5%, 100.0%, 100.0%, and 94.1%, respectively. The presence of PRH was balanced in both MC and FA – 5 (8.5%) for each – (Fig. 3), meaning a sensitivity, specificity, PPV and NPV of 80.0%, 98.1%, 80.0%, and 98.1%, correspondingly. The presence of HE had a sensitivity and NPV of 100% and specificity and PPV of 16.0% and 2.1%, respectively. HE was observed in 45 (76.3%) of MC image and in just one (1.7%) FA image (Fig. 1). Fibrosis was observed in 40 (67.8%) eyes in MC and in 22 (37.3%) when using FA (Fig. 1), for a specificity of 95.5%, specificity of 50.0%, PPV of 58.0%, and NPV of 93.8%. The identification of DME had a sensitivity of 45.5%, specificity of 100%, PPV of 100%, and NPV of 76.2%, being identified in 5 (8.5%) of MC and in 21 (36.1%) FAs. An ERM was found in both images in 2 (3.4%) cases, being more apparent in MC images (9; 15.3%) (Figs. 1, 2), for a sensitivity, specificity, PPV and NPV of 100.0%, 83.3%, 28.6% and 100%, respectively. Ischemia was registered in both modalities in 26 (44.1%) eyes and absent in 13 (22.0%) eyes, with disagreement between modalities in 20 (33.9%) eyes. MC presented sensitivity, specificity, PPV, and NPV values for ischemia of 74.3%, 68.4%, 81.3% and 59.1%, respectively. The observation of LS was similar between the 2 imaging modalities – 38 cases (64.4%) in MC and 37 cases (62.7%) in FA (Figs. 1, 2), reflecting a sensitivity and NPV of 100.0%, a specificity of 89.5% and PPV of 94.7%.

Table 3 Detection of eyes with additional DR features using MultiColor imaging compared to fluorescein angiography
Full size table
Table 4 Sensitivity and specificity of MultiColor imaging in detecting additional features of DR
Full size table

ROC curve of MC in detecting NVC, NVD, NVE, and other features

As shown and highlighted in bold in Table 5 the AUC analysis of MC was excellent for the detection of NVD (0.83, 95% confidence interval (CI), 0.71–0.95, p < 0.001), excellent for detection of VH (0.81, 95% CI 0.60–1.00, p = 0.005), IRH (0.89, 95% CI 0.74–1.00, p < 0.001) and PRH (0.89, 95% CI 0.68–1.00, p = 0.004) and outstanding for detection of LS (0.95, 95% CI 0.87–1.00, p < 0.001). This analysis also showed that MC was acceptable for the detection of NVE (0.71, 95% CI 0.53–0.89, p = 0.035), fibrosis (0.73, 95% CI 0.59–0.86, p = 0.005), and ischemia (0.71, 95% CI 0.57–0.86, p = 0.010). The AUC analysis for ERM detection was outstanding but without being statistically significant (0.92, 95% CI 0.81–1.00, p = 0.052). The AUC analysis of HE was inferior to acceptable (AUC < 0.7).

Table 5 Receiver-operating characteristics of MultiColor imaging in detecting NVC, NVD, NVE and additional features
Full size table

Intergrader agreement

The intergrader agreement for the MC features accessed was almost perfect with a weighted kappa of 0.86 (standard error: 0.02, 95% confidence interval: 0.82–0.88) (highlighted in bold in Table 6).

Table 6 Intergrader agreement for the MultiColor features accessed
Full size table

Discussion

The hallmark of PDR is the presence of neovascularization, which is induced by global retina ischemia [11, 28]. CFP and FA are still considered the gold standard for PDR although there have been recent advances in other imaging modalities [4, 6, 29,30,31,32]. The NVC is friable and can bleed, thus hemorrhagic complications such as the presence of VH or PRH are a sign of neovascularization and their presence determines the presence of high-risk PDR [22, 28, 33]. In this study, we compared the accuracy of MC to conventional FA to evaluate the presence of PDR and associated features. A screening test should be able to accurately identify diseased and non-diseased individuals and ideally should be highly sensitive (high probability of detecting disease) and specific (high probability of excluding the disease). However, sensitivity and specificity are inversely proportional, which means that an increase in sensitivity will result in a decrease in specificity and vice versa [24].

We report that MC alone revealed an excellent performance in diagnosing NVCs due to its high sensitivity and PPV (very low rate of false positives), which could be helpful in diagnosing PDR, as the green reflectance highlights both the vascular and fibrotic component of neovascularization [4, 6,7,8,9]. However, MC was not so useful for PDR screening purposes due to its low specificity and NPV value, resulting in low accuracy for excluding negative cases (high rate of false negatives). We also found that the specificity and NPV were higher for NVD compared to NVE. This means that MC was able to correctly identify 76.9% of patients as having no NVD, while it properly returned a negative result for 63.6% of NVE. We believe this result may be due to better contrast in the MC image at the disc area, as the vessels are reddish and the optic disc is yellowish, and also because the area of interest is much smaller, as the definition of NVD is neovascularization at or within 1 disc diameter of the disc margin [8, 22]. Indeed, the ROC analysis showed MC had excellent and statistically significant discrimination for NVD compared to FA and an only acceptable and statistically significant discrimination for NVE. This could be explained by some difficulty in distinguishing early NVE from IRMA in MC, which led to a more precise classification when using FA due to the identification of fluorescein leakage in NVE [20, 21]. Nevertheless, albeit the presence of NVE determines a more severe disease, IRMA is a known risk factor for PDR [31, 34, 35] and, in the natural disease course, about 50% of patients with severe non-proliferative DR develop PDR within a year and 15% develop high-risk PDR [36]. Besides NVD, MC was also excellent in discriminating VH and PRH, with a higher sensitivity (80.0% vs 62.5%) for PRH and a higher specificity for VH (100% vs 98.1%). The presence of VH and PRH commonly indicates the presence of active NVC and MC was found to be statistically significant superior to FA.

Regarding other features of DR, MC was highly sensitive in detecting MA, IRH, fibrosis, and HE. These results are likely due to the contrast and quality of the images, allowing for better discrimination [4, 6,7,8,9,10]. Also, FA is black and white, and therefore, whitish fibrotic tissue and yellow HE are more easily unapparent in FA. MC has been reported to offer the possibility of identifying ischemic areas because of the green wavelength [4, 6, 7]. We identified the areas of nonperfusion as areas with a red enhancement when compared to the surrounding homogeneous orange background color. The ischemia appeared to be reddish and faded and could be better highlighted as hyporeflectant if seen with the blue-green wavelength filters. We found a PPV of 81.3% for ischemia and NPV of 59.1%, with acceptable discrimination in the AUC analysis. Nevertheless, MC was outstanding in the identification of laser scars from PDR treatment, with high sensitivity, specificity, PPV, and NPV.

We have demonstrated that MC has advantages over FA and it can be acquired along OCT and/or FA as it is integrated into the Heidelberg Spectralis platform (Heidelberg Engineering, Heidelberg, Germany), obviating the need to move the patient to another imaging device [4, 6], which is even more valuable as we go through the current COVID-19 pandemic. Also, our work reinforces the conclusions of other authors [4, 6, 9], that one single device allows saving time in high-volume clinics. Albeit the sensitivity was higher than the specificity when compared to FA, Roy et al. [9] validated it against CFP and so MC may be useful for diabetic screening as it can be incorporated in the Spectralis-HRA, Spectralis-OCT, or Spectralis HRA + OCT. This could be advantageous in countries that regularly perform diabetic retinal screening and are going to upgrade it with OCT. The combination of Spectralis-OCT with MC instead of separate OCT and CFP could ultimately lead to less referrals of “ungradable” cases as MC performs better in media opacities and does not require dilation [4, 6, 9]. Nonetheless, we should consider its limitations. In fact, the technique is not widely available and requires a learning curve for appropriate clinical correlation as MC incorporates 3 wavelengths, demanding a longer duration of fixation for acquisition, and the findings appear different from CFP [4, 5]. We used two trained retina specialists, familiarized with this imaging modality and thus the excellent agreement. Furthermore, the study is retrospective in nature and includes a small sample; further multicenter clinical studies are needed to complement our observations. We have chosen to compare MC to FA as the Spectralis imaging platform does not incorporate standard CFP and we stopped performing it in our routine practice. Finally, the negative impact of imaging artifacts, such as ghost maculopathy, should also be considered [4, 23]. The presence of the artifact precluded a proper assessment of macular features such as DME and ERM, which led to many cases of unsure/ungradable.

In conclusion, MC detected some PDR features such as NVD, VH, PRH, IRH, and laser spots, more accurately than FA. These findings make MC a useful test for the diagnosis and follow-up evaluation of PDR, complementing the noninvasive imaging of this disease.