The fundus of the eye offers a unique opportunity to observe the microvasculature of the human body non-invasively. The analysis of the retinal vessel tree has advanced from measuring vessel diameters (1, 2) around the optic disk to analyzing the geometry of the retinal vascular structure in form of tortuosity measurements (3), branching angles (4), or fractal dimensions (5, 6). According to cross-sectional studies, retinal vascular geometric characteristics (RVGC) (7) can be used as quantifiable and objective early diagnostics in microvascular diseases, e.g., diabetes mellitus (8, 9), hypertension (10, 11), or cardiovascular disorders (12, 13). A few long-term studies also show promising results using RVGC as risk factors for the development of microvascular complications (14, 15). Such complications threaten the quality of life and are a burden to the healthcare system. Identifying patients with a particularly high risk could help allocating resources and support effectively, with the goal to prevent or slow down complications of e.g., diabetes.

Patients with diabetes are recommended to be regularly screened (16), and since the early 2000s fundus photographs started being digitally archived. Under the umbrella of the digital patient archive, caregivers and patients should be able to easily access data, including imaging material in the future.

To facilitate clinical usage of RVGC, reliable reference values need to be available. Establishing reference values is challenging due to a range of factors (17): First of all, measuring a three-dimensional object, as in the fundus of the eye, from a 2D image can only be an estimate. Caliber or angle measurements can therefore always only be a proxy. Thus unsurprisingly, RVGC differ between different vessel assessment software (18, 19). Secondly, even when the same analysis software is used, RVGC can vary for different cameras (due to resolution or field of view) (20) and, third, ocular magnification needs to be factored in to account for differences in axial length and refraction between eyes.

Historically, to measure lesions and compare their sizes between examinations, photographic magnification is corrected for by methods proposed by Littman (21, 22) and Bengtsson and Krakau (23) for the measurement of structures at the background of the eye or for the correction of optic disk measurements on fundus photographs. Those correction factors are device-specific and require additional measurements, such as axial length, and thus can be impractical in clinical practice. Correction factors should be applied with care (24, 25).

Vessel assessment software either processes all analyses in pixels [e.g., VAMPIRE (26, 27)], or employs an average image conversion factor (ICF) computed in a subset of images (e.g., SIVA) (28). This ICF is commonly calculated by the ratio of a reference optic disk (OD) size (e.g., 1,800 μm (29, 30) or 1,850 μm (31, 32), derived by Littmann’s method, based on the assumption of this being the average adult optic disk size) and the average optic disk size of a cohort. In studies where comparisons between different software packages have been attempted, the same conversion method should be used in both cases (19, 28).

To allow comparison of different software or analysis methods to establish RVGC reference values (17, 19), a large public data set is needed, containing fundus images and medical characteristics. This would help determine if associations with systemic variables might be software-dependent. None the less, large cross-sectional studies have shown that associations of RVGC measurements with medical characteristics persist despite a difference in software (18, 19). However, an average ICF does not enable long term follow-up of individual patients, when camera settings change or individual refraction differs between measurements. For example, in a retinopathy screening setting monitoring the development of diabetic retinopathy (DR) is long term (decades) and cameras are updated regularly during that time. However, for a patient in any screening setting it would be important to register RVGC changes between visits and make conclusion based on such developments. An individual ICF could be more reliable considering longitudinal follow-up, as it is based on the OD size of the analyzed eye. In contrast to constant ICF with calculations based on an average OD size, the individual ICF takes into consideration that OD size varies between individuals. We postulate that this could potentially enable comparison of images with variable imaging settings and subsequently changed resolution or field of view and this would also apply to changes in refraction. The use of different camera settings needing adaptation of the average ICF hinders the establishment of reference values from past long-term data (7, 33, 34). There is a need for the acquisition of reference values that are applicable for an individual risk assessment, however, all aforementioned problems and limitations complicate this, limiting the potential for patient care.

This study aims to use fundus image material from a long-term retinopathy follow-up study to identify problems created by changing imaging modalities or settings (e.g., image centering, resolution, viewing angle, illumination wavelength) and to offer solutions for longitudinal retinal vessel analysis. Special emphasis lies on the use of an individual ICF compared to a constant ICF and how the different analysis methods affect associations with obtained medical characteristics.

The analysis of fundus images in this study displays several challenges in connection with commonly used vessel assessment software. We present the effect of changed image modalities, variable image centering and use of ICF on measured RVGC. Problems were identified in the use of constant versus individual ICF application, changed centering of the image (ODC vs. MC) and its effect on clinical associations. The commonly used OD diameter used for ICF calculation influences the vessel diameter measurement given in μm.

As the individual ICF’s differ from the constant ICF used in the first analysis, individual diameter measurements end up being either narrower or wider. This depends on the individual ICF being smaller or bigger than the constant ICF. Therefore, the diameter measurements of the individual ICF analysis are over-all wider on average than in the first analysis but only statistically significant for the CRVE (Figure 3 and Supplementary Table 1). The same could be observed in Figure 5 and Table 5, were the Bland-Altman analysis found poor agreement with more positive mean differences for the CRVE measurements when comparing measurements done with individual ICF vs. constant ICF. This means that the constant ICF measures statistical significantly smaller values for central retinal vein equivalents (mean difference of 11.6 and 7.5 μm), a trend for smaller values was also observed for central retinal artery equivalents. For arteriovenous ratios, fractal dimension and simple tortuosity, measurements with individual ICF were up to 0.03 times smaller, which was negligible. Numerical values can be obtained from Table 6 and plots can be found in Figure 5 and Supplementary Figures 3, 4. As can be seen in the box plots of Figures 3, 4 or Table 4, values that were obtained using an individual ICF are more likely realistic, showing a larger range of vessel diameters. To establish individual wellbeing of the vasculature, a constant ICF is not sufficient, because it bases its calculation on the average OD size of a cohort. Instead, the individual ICF factors in differences of OD of each eye to produce a risk assessment of a particular patient.

In the past (57), the average grid was converted to determine the size of the measurement zones. Now, with software, the ICF determines the mean zone size (28). An average ICF is usually calculated from the mean OD size of a sizable percentage (10% with SIVA) of the cohort or the mean optic disk diameter is computed from the whole set of images (19, 28). The calculation of the ICF is facilitated by an average OD diameter instead of raw pixel measurements of individual images adjusted for ocular magnification of the eye. The use of a fixed micron reference for OD for any cohort is a source of variation or error influencing caliber measurements. Optic disk size can vary significantly between subjects. Quigley et al. and others (58, 59) found that ethnicity and sex can affect OD size. For a Finnish or even Scandinavian cohort no reference measure for OD size was previously established. Generally, using a constant conversion factor (especially if the 1,800 μm is not a true reference diameter for the cohort) is particularly problematic for individuals with extreme optic disk diameter or if the cohort had considerable spread in OD diameter. This can be, arguably, more inconvenient than leaving the measurement in the original units of pixels.

Generally, a large ICF indicates a small OD diameter and therefore a smaller measurement zone. Hereby, vessels measured are nearer to the OD and therefore anatomically have the tendency to be wider. A small ICF indicates a wider OD diameter, which in turn widens the measurement zone. With this, vessels from farther away from the OD are included into the calculations, therefore, smaller ICFs have the tendency to be associated with narrower diameter measurements. Figure 2 illustrates this. In Table 4 this trend can be observed when looking at the columns of vessel diameters in μm.

Diameter measurements done with a constant ICF are limiting the variance of the RVGC results, artificially enlarging or shrinking values depending on the individual ICF, which can either be larger or smaller than the constant ICF in any individual eye. CRVE and CRAE computed with individual ICF are a better estimate for individual measurements, even if the variance within the cohort increases. Vessel assessment software users should be aware of this.

Generally, an individual ICF presents the individual range of caliber measurements, as can be observed in McGrory’s Supplementary Table 4 (28). In their study the difference between constant and individual diameter measurements was significant. In the current study not all results were statistically significantly, which could stem from a smaller difference between average individual and constant ICF. An individual ICF leads to different measurement values of the vessel diameters in μm. Its apparent non-use could be explained by the necessity to calculate a new ICF of every photograph. Future automation can remedy that. To avoid problems caused by conversion of pixel diameter to μm diameters with ICF, one could directly use the calculated central retinal equivalents measurements in pixel (17, 28). However, an approximation of width in μm helps clinical judgment, and by using pixels a comparison to existing studies would be hindered. A further option for utilizing central retinal equivalent for individual disease progression could be to establish a ratio for change between visits, which is only possible when there is no change in camera or if a correction factor can be obtained.

In a clinical setting RVGC are best measured by a trained technician. Acquiring RVGC with any software is likely to involve the knowledge of camera settings and other information to produce reliable results. In the future, there might be fully automated software with input from the fundus camera, which would produce for results simply by clicking a button. For now, RVGC analysis is not standardized enough to produce comparable results between studies (18, 20).

In this study the comparison of MC and ODC images gave statistically significant results for the central retinal equivalents (Figure 3 or Supplementary Table 1) with poor agreement from Bland-Altman analyses (Figure 6 or Table 6), this is comparable to Mookiah et al. (51). Even though all RVGC-software recommend the use of ODC images, some studies only have MC images available (26, 60). Therefore, it was paramount to find out how the centering is related to the characteristics measured. In the Bland-Altman analysis for ODC vs. MC images we found that RVGC were generally larger measured from ODC images, compared to MC images (Table 6) (camera and resolution settings were kept the same). Therefore, future studies using MC images need to keep in mind that RVGC might be smaller in comparison to ODC images, or vice versa. In the case of our 60° images, only the vessel caliber results are affected significantly (Figures 3, 6 or Supplementary Table 1 and Table 6). The lack of agreement for the caliber measurements could also be caused by zone C being cut off to a varying extent in MC images, as can be seen in Figure 2. In the past, diameter measurements were shown to be more easily affected by changed image modalities and different analysis software (18). Also, their correlation coefficients with clinical characteristics can change (28). In this study AVR were identical between MC and ODC images (Table 6 and Supplementary Figure 5), with Bland-Altman analysis indicating negligible differences. This indicates that a factor between ODC and MC cancels itself out when it is derived from CRAE and CRVE. Therefore, despite variable image centering, measurements can be directly compared between ODC and the 60° MC images employed in this study, where Zone C is complete in both points of view. Generally, the AVR is a more robust characteristic when it comes to measurement changes due to resolution, angle, or calculation technique (49, 52), but important information about exact changes in arteriolar and venous vessels is lost (61).

Measurements of fractal dimension and simple tortuosity also showed good agreement in the Bland-Altman analysis, proving that they are similar despite different image centering. This again indicates that a factor between ODC and MC cancels itself out when Df and ST are computed and therefore such derived measurements can be directly compared between MC and ODC. Numerical values can be obtained from Table 6 and plots can be found in Supplementary Figures 5, 6.

Individual ICF allows for a more flexible inclusion of different image modalities, this is also possible for single patients. As mentioned before, the same vessels need to be selected to always make analysis methods comparable. Generally, if MC images are used to analyze RVGC it is important to select images showing a larger degree of the retina, including at least zone B and most of zone C. Preferably, vessels included in the analysis are not on the edge of the image to avoid possible distortion. Therefore, studies using MC images need to take the above into account, when deriving RVGC. This warning applies to macula centered images with 45° or 30° field of view, for example, where Zone C can be substantially incomplete. In these cases vessels detected would be reduced and the results of the calculations could be affected.

Generally, diameter measurements show association with clinical characteristics (15), for example, Broe et al. measured the vessel caliber in an adolescent T1D cohort and found correlation between diabetes duration, HbA1c and CRVE, and CRAE with systolic and diastolic blood pressure (33). The correlation coefficients of RVGC with clinical parameters of this study differed between the three analysis methods (Table 7). For example, the duration of diabetes correlated with six RVGCs measured with an individual ICF and only with two RVGC measured with a constant ICF and only with two from MC images. Part of this, the STt and Sta were the only characteristics to be statistically significant correlated through all analysis methods with duration of diabetes. The positive correlation of STt and STa indicated that a longer duration results in more curved arteries (62, 63). After Bonferroni correction only the correlation between diabetes duration and STt calculated with individual ICF for ODC and MC images persisted. Associations evaluated with multivariable linear regression also differed between the analysis methods (Table 8), however, testing their effect on associations with systemic variables is outside the scope of this manuscript. Contrary to prior studies involving adults with diabetes (64), the caliber measurements (calculated with constant ICFs in this study) showed no statistically significant correlation with clinical characteristics. This is not surprising, as this study consists of children who were of good health except for a still relatively short duration of diabetes and only a few with even mild retinopathy.

Scanned fundus images can be analyzed with modern tools. This is possible when using an individual ICF for analyzing ODC images or 60° MC images. RVGC, except for central retinal equivalents can be used to assess the vasculature of eyes. The results of the three analyses illustrate the challenges of establishing normal and abnormal values of RVGC in a clinical setting. On top of the general limitations of employing a constant reference OD for pixel-to-micron conversion, the analyses with constant and individual ICF result in differing values for retinal caliber measurements in μm, e.g., a refractive error outside the norm will affect central equivalent measurements disproportional with a constant ICF compared to an individual ICF. As shown by the 60° images of the current work, where Zone C is complete in both points of view, the analysis of MC and ODC images resulted in good agreement for arteriovenous ratio, fractal dimension and simple tortuosity measurements, signaling that MC and ODC images could be used interchangeably when Zone C is complete. However, CRAE and CRVE are statistically significant, presenting further evidence of the susceptibility of diameter measurement to changed imaging settings. In summary, we recommend using ODC images for Zone C analysis, but if not available, MC images can be utilized under the recommendations above.

This study enhances the awareness that central retinal equivalent values calculated with different vessel assessment software or between different studies are not by default interchangeable. For evaluating the relationship of changed analysis methods on systemic associations with RVGC, a larger publicly available image and medical data base is needed. This could facilitate enhanced collaboration in the field to produce comparable individual RVGC risk values between studies.

The images analyzed for this study can be found in the health atlas repository (https://health-atlas.de/projects/43), where access to the images can be requested after registration with the health atlas.

The studies involving human participants were reviewed and approved by the Ethics Committee of the Oulu University Hospital. The patients/participants provided their written informed consent to participate in this study.

CS, AF, and NH planned the study. CS scanned the images and used SIVA for analyzing the RVGC. CS and FGR did statistical analyses. All authors wrote and revised the manuscript.