The eye is a structurally advanced optical organ with a complex vascular network. The maintenance of retinal function is critically dependent on the normal functioning of the blood-retinal barrier (Kaur et al., 2008; O’Leary and Campbell, 2023). Fundus diseases and related systematic diseases can cause functional changes in the fundus such as ischemia and hypoxia, which can damage the blood-retinal barrier, leading to retinal vascular permeability changes, hemorrhage, exudation, neovascularization, and other lesions (Pournaras et al., 2008; Ascunce et al., 2023; Monickaraj et al., 2023). Diabetes mellitus (DM) is a common endocrine metabolic disease with its microvascular complications in the eye that can result in diabetic retinopathy (DR), which is the leading cause of blindness in working-age people (Bourne et al., 2021; Steinmetz et al., 2021). Hypoxia is considered to be the core pathogenesis of DR besides of hyperglycemia in retinal blood vessels caused by DM (Hardarson, 2013). Hypoxia along with hyperglycemia stimulate vascular endothelium and eventually lead to vascular endothelial cell dysfunction, which in turn leads to the occurrence of DR. (Hardarson and Stefánsson, 2012; Khoobehi et al., 2013). Therefore, it is equally important to detect retinal oxygen saturation as well as blood glucose in DM patients, which can help to prompt the potential occurrence of DR and avoid further visual impairment.

Retinal vessel is the only vascular system in the body that can be observed in vivo (Hanssen et al., 2022), and non-invasive imaging technique is the first choice for retinal vascular oxygen saturation measurement. Nevertheless, there are only limited imaging instruments available to measure retinal oxygen saturation, in which dual-wavelength retinal oximetry is a promising technique. By measuring the light of different wavelengths reflected from the eye, it takes advantage of the light absorption variations between oxyhemoglobin and hemoglobin, and a calculation can be made that correlates directly to the retinal vascular oxygenation (Beach, 2014; Jeppesen and Bek, 2019; Garg et al., 2021). Previous studies using retinal oximetry have found that the oxygen saturation of retinal vessels in DR is significantly higher than that in normal subjects, and increases with the severity of DR (Hardarson and Stefánsson, 2012; Khoobehi et al., 2013; Jørgensen and Bek, 2014). Retinal oximetry has also been used to monitor retinal oxygen saturation in other eye diseases such as central retinal vein occlusion (CRVO) and glaucoma, showing varying degrees of increase in retinal oxygen saturation (Stefánsson et al., 2019). However, despite its potential, retinal oximetry is not widely used in ophthalmology clinics due to its recent commercialization.

Traditional color fundus cameras are the most commonly used equipment for fundus examination and disease screening (Karlsson et al., 2021). Their broad-spectrum cameras use complementary metal-oxide semiconductor (CMOS) or charge couple device (CCD) photoreceptors covered with red-green-blue (RGB) filters (Malvar et al., 2004). Thus, oxygen-sensitive and insensitive wavelengths reflected from retina vessels, which corresponds to different oxyhemoglobin and hemoglobin light absorptions, are generally received in the red and green channels, respectively, allowing for the estimation of blood oxygen function by extracting light absorption information from these channels.

In this study, we aim to develop a computational method for analyzing oxygen-saturation-related functional parameters with traditional color fundus photography based on precise vessel segmentation and light absorption analysis, providing an easy-to-apply image analysis algorithm for traditional color fundus cameras. With that, we also explore the characteristic alterations of retinal vascular oxygen function in type 2 DM patients to look for new potential image biomarkers for the monitoring of DM.

There is no difference in the age and sex between the 50 healthy volunteers in Normal group and the 50 patients in DM group (Table 1).

Table 1 alsodemonstrates the ODR for all vascular subgroups, ODR_v, as well as the t-statistic and p-values. In both healthy normal group and DM group, A ODR is higher than V ODR (mean ± SD, A vs. V, Normal: 0.713 ± 0.057 vs. 0.708 ± 0.063; DM: 0.755 ± 0.083 vs. 0.754 ± 0.082), main A ODR is higher than main V ODR (Normal: 0.728 ± 0.056 vs. 0.717 ± 0.058; DM: 0.773 ± 0.083 vs. 0.761 ± 0.082) and micro A is higher than micro V ODR (Normal: 0.634 ± 0.068 vs. 0.599 ± 0.082; DM: 0.672 ± 0.096 vs. 0.627 ± 0.117), which means that ODR is positively correlated with blood oxygen saturation. Compared to Normal group, ODR is significantly increased in all vascular subgroups except micro venular in DM group ( p < 0.05 , respectively). ODR_v is significantly decreased in DM group ( p < 0.001 ).

As shown in the logistics regression results in Table 2, higher ODRs are significantly associated with the occurrence of DM, especially in main A (OR 1.953, 95% CI 1.252-3.046) and main V (OR 1.908, 95% CI 1.227-2.968), and also in micro A (OR 1.604, 1.047-2.455). While Lower ODR_v (OR 0.381, 95% CI 0.210-0.693) is significantly associated with the occurrence of DM. The C-statistic for the discrimination of DM from the healthy normal group combining all the acquired oxygen-saturation-related functional parameters is 0.777 (95% CI 0.687-0.867, p < 0.001 ) (Figure 2).

In this study, we present a computational approach for extracting oxygen-saturation-related functional parameters through precise vascular segmentation and light reflection channel analysis using single traditional color fundus photography. By applying this method to the comprehensive analysis of the fundus images acquired from normal individuals and patients with type 2 diabetes, we revealed that the retinal vascular ODRs were increased while the ODR variability was reduced significantly in DM patients, providing new potential image biomarkers for the management of DM.

The successful extraction of retinal vascular ODRs with color fundus photography offers a new protocol for retinal vascular function evaluation, providing additional quantitative functional information based on the most commonly used fundus imaging tools in clinical settings. Compared to the current existed retinal oximetry, which only reveals oxygen saturation level in main retinal vessels and requires subjects to withstand long light exposure time (Hardarson, 2013), our method achieves information in microvascular level without any additional hardware implementation, which ensures its convenience and accessibility with greatly reduced cost. These advantages suggest extensive potential application value in the diagnosis and screening of retinal diseases.

In our study, we have shown that the increase of proposed oxygen-saturation-related ODRs were statistically significant at the stage of DM, which could be related to the stronger affinity of glycosylated hemoglobin to oxygen in DM (Graham et al., 1980; van Kampen and Zijlstra, 1983). Indeed, former studies using retinal oximetry have found obvious increase in retinal oxygen saturation level during DR, especially in proliferative DR, but no significant change was shown in DM stage before (Khoobehi et al., 2013). This indicates that our newly proposed ODRs might be more sensitive parameters than that of commercial retinal oximetry in finding alterations in blood oxygen function. Moreover, our exploration of ODR in the microvascular groups revealing significant increase in micro arterioles of DM patients further proves the high sensitivity of our method.

We have proposed and analyzed the ODR variability for the first time on the basis of extracted ODR of the full retinal vascular network and found it to be significantly decreased in DM patients compared to normal subjects. While glycosylated hemoglobin has higher affinity to oxygen (Graham et al., 1980), it damps the oxygen-release capacity (Ditzel, 1972) of retinal blood, results in reduced oxygen saturation variations. Studies have also found that the retinal blood flow velocity increases in DM patients (Landa et al., 2011), meaning less oxygen release time of blood in retinal vessels, further reduced the ODR variability.

Our method suffers limited retinal vessel resolution and imaging depth due to the physical nature of fundus photography. Only the superficial retinal vessels can be analyzed without 3-dimensional (3D) illustration. The same limitations also exist in current commercial retinal oximetry. Recently, visible-light optical coherence tomography based techniques for retinal oxygen saturation measurement have been gradually proposed (Pi et al., 2020; Poddar and Basu, 2020), achieving high resolution 3D imaging of retinal vascular oxygen function information (Pi et al., 2020), although most of these studies are still in laboratorial stage.

Our attempt to distinguish DM from normal subjects with all the extracted functional ODR parameters showed a relative low C statistic result (AUC = 0.777), which is not sufficient for DM diagnosis or screening. Since our previous study found that the morphological parameters of retinal vessels in DM patients also changed significantly (Li et al., 2021), further combining functional and morphological parameters of retinal blood vessels and multi-center studies or follow-up studies may help us better understand the changes in retinal blood vessel characteristics during the occurrence and development of DM and even DR, so as to diagnose DM or DR More accurately. Furthermore, applying our method to other eye diseases related to retinal vascular oxygen function alterations such as glaucoma (Vandewalle et al., 2014), central retinal vein occlusion (Hardarson and Stefánsson, 2010), age-related macular degeneration (Geirsdottir et al., 2014), etc. in the future would further demonstrate the reliability and universality of our method.