Artificial Intelligence–Assisted Perfusion Density as Biomarker for Screening Diabetic Nephropathy

Abstract

Purpose: To identify a reliable biomarker for screening diabetic nephropathy (DN) using artificial intelligence (AI)–assisted ultra-widefield swept-source optical coherence tomography angiography (UWF SS-OCTA).

Methods: This study analyzed data from 169 patients (287 eyes) with type 2 diabetes mellitus (T2DM), resulting in 15,211 individual data points. These data points included basic demographic information, clinical data, and retinal and choroidal data obtained through UWF SS-OCTA for each eye. Statistical analysis, 10-fold cross-validation, and the random forest approach were employed for data processing.

Results: The degree of retinal microvascular damage in the diabetic retinopathy (DR) with the DN group was significantly greater than in the DR without DN group, as measured by SS-OCTA parameters. There were strong associations between perfusion density (PD) and DN diagnosis in both the T2DM population (r = −0.562 to −0.481, P < 0.001) and the DR population (r = −0.397 to −0.357, P < 0.001). The random forest model showed an average classification accuracy of 85.8442% for identifying DN patients based on perfusion density in the T2DM population and 82.5739% in the DR population.

Conclusions: Quantitative analysis of microvasculature reveals a correlation between DR and DN. UWF PD may serve as a significant and noninvasive biomarker for evaluating DN in patients through deep learning. AI-assisted SS-OCTA could be a rapid and reliable tool for screening DN.

Translational Relevance: We aim to study the pathological processes of DR and DN and determine the correspondence between their clinical and pathological manifestations to further clarify the potential of screening DN using AI-assisted UWF PD.

Introduction

Diabetic retinopathy (DR) and diabetic nephropathy (DN) are common microvascular complications of type 2 diabetes mellitus (T2DM) and can affect the macula and peripheral retina, leading to blindness in critically ill or end-stage renal disease patients.^1–3 Therefore, the early diagnosis of DR and DN is very important to prevent blindness in these patients. 

In recent years, rapid advancements in imaging technology have enabled dynamic, multilevel, three-dimensional, and high-resolution observation of the retina and choroid. Given the close relationship between the pathophysiological mechanisms and clinical changes of DN and DR,^2,4 researchers believe that these conditions can predict each other.^2,5,6 Previous studies have explored the correlation between DN and fundus images or blood flow in the macular region, finding that renal impairment is linked to the enlargement of the foveal avascular zone (FAZ) in diabetic patients.^7,8 However, microvascular lesions in DR are not confined to the macular fovea but are widespread throughout the retinal microvasculature. In this context, swept-source optical coherence tomography angiography (SS-OCTA) allows for the quick, highly detailed, and noninvasive quantification of a wide range of retinal features.⁹ Our preliminary SS-OCTA studies confirm that this technology provides valuable information for detecting potential blood flow damage at various stages of DR.¹⁰ 

Currently, the clinical diagnostic criteria for DN rely primarily on invasive serological examinations and renal biopsies.¹¹ Retinal screening for DN in T2DM patients is typically limited to fundus imaging and the FAZ evaluation, which do not provide comprehensive quantification of retinal and choroidal microvasculature characteristics. Additionally, due to the abnormal renal function in DN patients, invasive fluorescein fundus angiography (FFA) cannot be used to assess retinal and choroidal microvasculature damage. We hypothesized that certain parameters from the ultra-widefield (UWF) SS-OCTA could serve as rapid, noninvasive, and reliable screening biomarkers for DN. 

To test this hypothesis, we conducted a prospective study to observe retinal and choroidal microvasculature characteristics in patients with DR and DN using UWF SS-OCTA. We then classified the data using a random forest model to distinguish DN patients from the T2DM population (including a control group, a DR without DN group, and a DR with DN group) and the DR population (including a DR without DN group and a DR with DN group). The goal is to provide real-time monitoring for DR patients and assess the presence of DN using artificial intelligence (AI)–assisted UWF parameters. This approach aims to identify a quick, noninvasive, and reliable biomarker for DN screening. 

Methods

General Information

This study included a total of 169 patients with T2DM who were treated at the Shandong Eye Hospital from November 2020 to May 2022. The study protocol was approved by the Institutional Review Board of Shandong Eye Hospital (approval No. SDSYKYY202105). Written informed consent was obtained from all participating patients. All patients’ demographics and basic clinical characteristics were recorded. Patients without any systemic diseases or retinopathy were included in the healthy group. Patients with T2DM¹² and meeting the diagnostic criteria of DR¹³ were allotted to the DR without DN group. Patients with T2DM¹² and meeting the diagnostic criteria of DR¹³ but not of DN^14,15 were included in the DR with DN group. 

Machines for Eye Examination

All the patients were examined by an experienced physician using UWF SS-OCTA (SS-OCT, VG200D; SVision Imaging, Ltd, Luo Yang, China). The commercial SS-OCT equipment contained an SS laser with a central wavelength of approximately 1050 nm (full width of 990–1100 nm) and a scanning rate of 200,000 A-scans per second. The full width of the half-maximum axial resolution of the device in tissue was approximately 5 µm, and the estimated lateral resolution on the retinal surface was approximately 15 µm.¹⁶ Both optical coherence tomography (OCT) and OCTA data of 21 × 21 mm² area centered on the macular fovea were obtained by 1024 (horizontal) × 1024 (vertical) B-scans. By using the built-in eye-tracking mode of the device based on the integrated confocal scanning laser detector lens, eye movement artifacts during and between scans were minimized.¹⁷ 

Data Extraction and Deep Learning Analysis

The built-in AI identification and quantification software of the SS-OCTA recorded the area, perimeter and acircularity index, and fractal dimension 300 (FD300; blood flow density within a radius of 300 µm around FAZ)¹⁸ of FAZ, total retinal thickness (RT) (measured from the internal limiting membrane to the retinal pigment epithelium [RPE] in the central subfield),¹⁹ retinal vessel density (VD) and perfusion density (PD) of the deep vascular complex, choroidal perfusion (CP), choroidal vascularity volume (CVV; defined as the volume of the large and medium choroidal vessels), choroidal vascularity index (CVI; defined as the ratio of the volume of the large and medium choroidal vessels to the total choroidal volume), and choroidal thickness (CT) (measured from the outer edge of the hyperreflective RPE line to the inner edge of the sclera) in different circular radii around the fovea.^19–22 Data collection was carried out independently by an experienced investigator. 

Automatic measurement using SS-OCT built-in software offers an effective and objective method for quantitatively evaluating retinal and choroidal structures. Typically, the system's default hierarchical mode was used, with manual corrections made for any segmentation errors. However, scan artifacts present substantial challenges to accurate quantification of the retinal and choroidal layers.^23,24 To address this, the image quality score (IQS) of SS-OCTA images was assessed, and images with scores below 8 were excluded to eliminate large errors caused by projection, bulk motion, signal reduction, and other issues.²³ This evaluation was performed by two experienced ophthalmologists who were kept blinded to the participants’ data. In cases of disagreement, a more experienced ophthalmologist made the final decision. Samples with IQS scores below 8, which could not be corrected manually due to segmentation errors, motion artifacts, defocus, decentration, and masking, were excluded. 

A deep learning method was employed to standardize the data, selecting the most relevant optimal OCT parameters through 10-fold cross-validation (expressed by the importance of explanatory variables). Then, the data were randomly divided into 10 groups with approximately equal numbers to form a training set and a test set. The training set was divided into 10 folds for cross-validation, with each fold containing 90% training data and 10% verification data.²⁵ Finally, the training set data were then fitted to a random forest model, which was used to classify the data from the testing set.²⁶ Based on the optimal OCT parameters, we identified DN patients (DR with DN group) from the T2DM population (control group, DR without DN group, DR with DN group) and the DR population (DR without DN group, DR with DN group) and calculated the classification accuracy. All results were evaluated through a 10-fold cross-validation.²⁷ 

The obtained data were also statistically analyzed using SPSS v26.0 (SPSS, Inc., Chicago, IL, USA). The data were tested for normal distribution, followed by one-way analysis of variance or Kruskal–Wallis test (H test) as appropriate. Bonferroni correction was applied for multiple comparisons. Additionally, the Spearman correlation coefficient was used to evaluate the relationship between OCT parameters and grouping categories. 

Results

Patient Characteristics

The baseline demographic and clinical characteristics of the enrolled patients are summarized in Table 1. In this study, data from 169 patients (287 eyes) with T2DM were analyzed, comprising 15,211 individual data points. Ninety-eight control subjects (162 eyes) without retinal microvascular lesions were recruited, comprising 45 males and 53 females. Retinal microvascular lesions were identified in 71 subjects (125 eyes). There was a significant difference in gender distribution among the three groups (χ² = 6.978, P = 0.031). The average ages of patients among the control, DR without DN, and DR with DN groups were 54.31 ± 11.27 (range, 23–72), 56.76 ± 9.91 (range, 32–75), and 56.00 ± 9.67 (range, 39–79) years, respectively. There was no significant difference (F = 0.524, Bonferroni corrected P = 0.694). Patients in the control group did not have T2DM. The duration of T2DM in the DR without DN group (mean ± SD = 9.53 ± 6.71) was significantly shorter than that in the DR with DN group (mean ± SD = 13.31 ± 6.40) (t = −2.399, P = 0.02). 

Table 1.

 

Demographic and Clinical Characteristics of Enrolled Patients

 

Our results indicated that the best-corrected visual acuity (BCVA) measured using the international standard chart of vision and recorded as the logarithmic minimum angle of resolution (logMAR) was worse in the DR without DN group (mean ± SD = 0.30 ± 0.29) and DR with DN group (mean ± SD = 0.59 ± 0.42) compared to the control group (mean ± SD = 0.02 ± 0.04). The BCVA was significantly poorer in the DR with DN group (H = 91.827, Bonferroni corrected P < 0.001). There was no significant difference in intraocular pressure among the three groups (mean ± SD = 15.24 ± 1.95, 15.39 ± 2.55, and 15.56 ± 2.56; H = 0.759, P = 0.684). 

Retinal Vascular Complexes

In the UWF vascular images of the deep retinal vascular complexes (Fig. 1), the control group exhibited evenly distributed and densely packed blood vessels. In contrast, patients with DR showed nonperfusion areas, with significantly larger areas observed in those with DN. 

Figure 1.

 

The ultra-widefield fundus vascular images of retinal vascular complexes. The fundus vascular images of deep vascular complex in the control (A), DR without DN (B), and DR with DN (C) groups.

As depicted in Table 2 and Figure 2A, compared to the control group, the area (H = 13.397, Bonferroni corrected P = 0.001), perimeter (H = 11.344, Bonferroni corrected P = 0.003), and acircularity index (H = 16.236, Bonferroni corrected P < 0.001) of the FAZ were significantly increased in both the DR without DN and DR with DN groups. However, there was no significant difference in the FD300 of FAZ among the three groups (H = 5.757; Bonferroni corrected P = 0.056). 

Table 2.

 

The Characteristics of FAZ in the Control, DR Without DN, and DR With DN Groups Using UWF SS-OCTA.

 

Figure 2.

 

The OCTA parameters in the control, DR without DN, and DR with DN using UWF SS-OCTA. Comparison of the area, perimeter, acircularity index, and fractal dimension of FAZ (A), PD (B), VD (C), RT (D), CP (E), CVV (F), CVI (G), and CT (H) in different radii at the center of the macular fovea. ^*P ≤ 0.05, ^**P ≤ 0.01, ^***P ≤ 0.001.

In Table 3 and Figures 2B–D, we compared the PD and VD of the deep vascular complex and RT across various circular radii centered on the macular fovea. Except for PD_{0–1 mm} and VD_{0–1 mm}, significant differences in PD and VD of the deep vascular complex were found across different radii among the three groups (Bonferroni corrected P < 0.001). Both the average PD and VD were significantly decreased in the DR without DN and DR with DN groups compared to the control group. Moreover, the reduction in PD and VD in the DR with DN group was notably more pronounced than in the DR without DN group within certain radii. Moreover, significant differences in RT_{0–3 mm} (H = 11.272, Bonferroni corrected P = 0.004), RT_{0–6 mm} (H = 28.986, Bonferroni corrected P < 0.001), RT_{0–9 mm} (H = 22.306, Bonferroni corrected P < 0.001), and RT_{0–12 mm} (H = 10.861, Bonferroni corrected P = 0.004) were observed among the three groups. The mean RT values in both DR without DN and DR with DN groups were significantly higher compared to the control group. 

Table 3.

 

The OCTA Parameters of the Retina in the Control, DR Without DN, and DR With DN Groups Using UWF SS-OCTA

 

Table 4.

 

The OCTA Parameters of Choroid in the Control, DR Without DN, and DR With DN Groups Using UWF SS-OCTA

 

Our figure and data analysis revealed varying degrees of damage in the retinal vascular complex among patients in the DR without DN and DR with DN groups, particularly in PD and VD, with the DR with DN group exhibiting more severe damage. 

Choroidal Vascular Complexes

In Table 4 and Figures 2E–H, we compared the mean values of choroidal parameters (CP, CVV, CVI, and CT) across different circular radii centered on the macular fovea for all included subjects. 

Significant differences were observed among the three groups in CP_{0–3 mm} (H = 23.255, Bonferroni corrected P < 0.001), CP_{0–6 mm} (H = 23.645, Bonferroni corrected P < 0.001), CP_{0–9 mm} (H = 20.901, P < 0.001), CP_{0–18 mm} (H = 7.221, Bonferroni corrected P = 0.027), CVI_{0–15 mm} (F = 5.418, Bonferroni corrected P = 0.013), and CVI_{0–18 mm} (F = 5.039, Bonferroni corrected P = 0.01). In comparison to the control group, the CP near the macular fovea was significantly reduced in both the DR without DN and DR with DN groups. 

Correlation Between OCT Parameters and DN Diagnosis

Based on the statistical analysis and fundus image observations described above, we hypothesized that PD and VD of the deep vascular complex may serve as effective OCT parameters for screening DN. To test this hypothesis, Pearson's correlation coefficients were computed. Specifically, since the radii of 0 to 3 mm, 0 to 6 mm, 0 to 9 mm, and 0 to 12 mm are commonly used in UWF imaging, we conducted Pearson correlation analysis for OCT parameters within these ranges. 

As shown in Table 5, UWF PD demonstrated stronger correlations compared to other OCT parameters when screening for DN in both the T2DM and DR populations. 

Table 5.

 

Correlation Coefficients and Importance of Explanatory Variables by Random Forest Model for Screening DN Patients From the T2DM and DR Populations Using OCT Parameters

 

Screening Results of DN by Random Forest Model

We also employed the deep learning method to standardize our data, assessing the importance of explanatory variables through 10-fold cross-validation to identify the most relevant optimal OCT parameters. The importance of these variables, as determined by the random forest model for screening DN from both the T2DM and DR populations, is detailed in Table 5 (Importance). In the classification of DN from the T2DM population, key explanatory variables included PD (PD_{0–3 mm}: 0.093891, PD_{0–6 mm}: 0.086285, PD_{0–12 mm}: 0.251083) and VD (VD_{0–6 mm}: 0.081844, VD_{0–12 mm}: 0.089028). Similarly, in the classification of DN from the DR population, crucial variables were PD (PD_{0–3 mm}: 0.075730, PD_{0–6 mm}: 0.115155, PD_{0–9 mm}: 0.079301) and VD (VD_{0–6 mm}: 0.147807, VD_{0–9 mm}: 0.124143). These results suggest that UWF PD remains a robust OCT parameter for effectively screening patients with DN from both T2DM and DR populations. 

Subsequently, the UWF PD data underwent further processing using the 10-fold cross-validation method and random forest model to screen for DN within the T2DM and DR populations. The classification accuracy of each group in the 10-fold cross-validation method, along with the average classification accuracy of the random forest method, is presented in Table 6. The average classification accuracy of the random forest model for identifying patients with DN using PD was 85.8442% from the T2DM population and 82.5739% from the DR population. 

Table 6.

 

Classification Accuracy of DN Patients From the T2DM and DR Populations Using Random Forest Method

 

Discussion

The complications of T2DM lead to a series of global health problems that persist over time. DR and DN are common microangiopathic complications in patients with long-term T2DM.²⁸ Microalbuminuria is an early marker of endothelial injury. Its presence significantly increases the risk of DR, indicating a shared pathophysiological mechanism²⁹ and a potential “common development path” between DR and DN.^30,31 

Due to challenges in accessing medical care and the uneven distribution of medical resources, telemedicine has gained unprecedented attention. Our team urgently needed to identify a simple and quick biomarker to screen for diabetic microvascular complications. The severity of DR correlates with the severity of DN, making it a potential marker for predicting the progression of chronic kidney disease. Many studies have evaluated the relationship between macular microvascular changes and the severity of DR in nephropathic patients using OCTA.^7,32–35 These studies also show that renal impairment, as a systemic risk factor, is associated with an enlarged FAZ area in diabetic patients.⁷ As DN progresses, RVD gradually decreases³⁶ and the retinal microvasculature in both the superficial vascular plexus and deep vascular plexus becomes sparse.³⁷ Alé-Chilet et al.³⁸ reported that VD and FAZ areas can detect the degree of DN in patients with type 1 diabetes mellitus in a noninvasive and objective quantitative way. There is no potential advantage of PD in judging the type of DN. 

In our research, compared with the control group, the average PD and VD of the DR without DN and DR with DN groups decreased significantly. Moreover, the decrease in PD and VD was significantly greater in the DR with DN group than in the DR without DN group. There was a correlation between UWF PD and DN diagnosis in both the T2DM and DR populations. This prompted us to consider whether UWF PD could be a reliable indicator for accurately screening DN. We used a deep learning method to analyze the UWF SS-OCT data, and the importance of explanatory variables by a random forest model between UWF PD and DN diagnosis from T2DM and DR populations was still high. The average classification accuracy of the random forest model for patients with DN screened out by PD was 85.8442% from the T2DM population and 82.5739% from the DR population. This suggests that UWF PD could be a quick and reliable biomarker for screening patients with DN from T2DM and DR populations using deep learning. 

The diagnosis of DN typically depends on laboratory examination or renal biopsy, which is not optimal in terms of operational safety and cost-effectiveness. Patients with DR rarely realize the need for renal function examination.³⁹ Although FFA is the gold standard for detailed measurement of retinal microvasculature, this invasive procedure may increase the metabolic burden on patients with DN. With advancements in technology and multimodal imaging, OCTA achieves a higher blood vessel contrast than fundus photography and is not affected by dye leakage, which can obscure blood vessels in dye-based angiography.²¹ Furthermore, UWF SS-OCTA offers a larger examination range and faster scanning speed,⁴⁰ allowing real-time and accurate assessment of peripheral retinal blood flow in DN patients. AI based on deep learning has garnered tremendous global interest recently.⁴¹ By providing noninvasive, high-resolution, deep-resolution images of retinal and choroidal vessels, SS-OCTA significantly enhances our ability to use these images for disease screening and evaluating therapeutic efficacy and response.²¹ Some scholars have used deep learning models to detect and predict the incidence of chronic kidney disease (CKD) from two-dimensional retinal fundus images.⁴² Combining these superior biomarker detection capabilities with noninvasive procedures makes SS-OCTA, aided by deep learning, a promising screening technology for clinical practice.²¹ 

The BCVA (logMAR) of the DR without DN group and DR with DN group was significantly lower than that of the control group, with the reduction in the DR with DN group being greater. As T2DM progresses, retinal microvascular injury in DN patients worsens, leading to significant vision decline. In China, middle-aged and elderly patients often do not fully understand the complications of T2DM. It is important to note that vision decline can reflect DR progression in T2DM patients and warn of the possibility of DN, highlighting the need for disease awareness and screening. 

Regarding choroidal indices, a meta-analysis by Kase et al.⁴³ indicated that monitoring choroidal vessels in diabetic eyes could help detect DR onset through longitudinal observation. However, there are still controversies about changes in CT and blood vessels in DR patients.^44,45 Studies have shown that CT is thinner and the choroidal vessel area is reduced in diabetic patients, while others have reported thicker CT and increased choroidal vessel area in DR patients.⁴⁶ One study found no difference in choroidal changes between the DR and control groups but did find differences between the DN and control groups.⁴⁷ The result was controversial. CVI and CVV are new quantitative parameters of choroidal vascular health.^48,49 Our research found that the CP (CP_{0–3 mm}, CP_{0–6 mm}, CP_{0–9 mm}, CP_{0–18 mm}) and CVI (CVI_{0–15 mm}, CVI_{0–18 mm}) decreased in patients from the DR without DN and DR with DN groups, but there was no significant difference in CT and CVV. We speculate that the decrease in CP may be related to the narrowing of choroidal arterioles, choriocapillaris atrophy, and capillary dropout. We aim to improve classification methods to explore the characteristics of choroidal microvasculature in DR and DN patients. 

There are differing opinions on whether the duration of diabetes is a reliable predictor of DN.⁵⁰ Hung et al.⁵⁰ conducted a meta-analysis providing this view, while Sharma et al.⁵¹ showed that diabetes duration was the strongest predictor of DN, with durations over 12 years being the best predictor. In our study, the duration of T2DM in the DR with DN group was 13.31 ± 6.4 years, compared to 9.53 ± 6.71 years in the DR without DN group. This supports the idea that when the duration of T2DM exceeds 12 years, DN screening should be prioritized. 

For ophthalmologists, early detection and diagnosis of DR and DN in T2DM patients should be emphasized. For patients with long duration of T2DM and DR, timely DN diagnosis is crucial to prevent DN progression, not just focusing on DR treatment. For endocrinologists, even without the guidance of ophthalmologists and nephrologists, UWF SS-OCTA provides a quick and reliable method to preliminarily assess DN in T2DM patients through AI. Monitoring UWF PD alleviates the pressure of uneven medical resource distribution and serves as a telemedicine-based screening scheme for T2DM complications. 

There were some limitations to this study. First, the actual duration of T2DM may have been longer than recorded due to delayed diagnosis for various reasons. Despite our efforts to include DR at the same stage, strict differences between DR and DN existed according to different development stages. Additionally, we only included PD and VD for the deep vascular complex when collecting data. In the future, we will study the influence of PD on both the deep vascular complex and the superficial vascular complex in DN patients. 

Conclusions

Our results suggest that when evaluating patients with DN in terms of retinal microvasculature, AI-guided UWF SS-OCTA detection of PD may serve as a rapid, reliable, and noninvasive early warning tool. Whether changes in choroidal microvasculature in patients with DR can be used as a predictive index remains to be verified. To positively impact the quality of life of T2DM patients, it is crucial to detect diabetic microvascular complications (DR and DN) as early as possible for implementing combined treatment strategies. 

Acknowledgments

Supported by the Bethune Langmu Young Scholars Research Fund Project (BJ-LM2021007J) and New Ophthalmology Technology Incubation Fund Project. 

The data sets used and/or analyzed during the current study are available from the corresponding author on reasonable request. 

Disclosure: X. Xie, None; W. Wang, None; H. Wang, None; Z. Zhang, None; X. Yuan, None; Y. Shi, None; Y. Liu, None; Q. Zhou, None; T. Liu, None