Vision-Threatening Diabetic Macular Ischemia Based on Inferred Progression Pathways in OCT Angiography

Abstract

Purpose

To elucidate the progression pathways of diabetic macular ischemia (DMI) using OCT angiography (OCTA) images and to assess changes in visual acuity (VA) associated with each pathway.

Design

A single-center, prospective case series study.

Participants

One hundred fifty-one eyes from 151 patients with a 3-year follow-up period.

Methods

We obtained 3 × 3 mm swept-source OCTA images and conducted analyses of en face images within a central 2.5 mm diameter circle. Nonperfusion squares (NPSs) were defined as 15 × 15-pixel squares without retinal vessels. Each eye at baseline and after 3 years was embedded into a 2-dimensional uniform manifold approximation and projection space and assigned to 1 of 5 severity grades—Initial, Mild, Superficial, Moderate, and Severe—using the k-nearest neighbors method. We assessed major transitions (involving ≥4 cases) during 3 years. Subsequent probabilistic analyses enabled the construction of a graphical model, wherein directed arrows represented inferred pathways of DMI progression. From this cohort, 103 eyes of 103 patients who did not receive any ocular treatments during the follow-up period were subsequently evaluated for VA changes.

Main Outcome Measures

Inference of DMI progression pathways.

Results

In most cases, NPS counts increased in both the superficial and deep layers. The major transitions between these severity groups at 3 years displayed a unique distribution, and probabilistic analyses suggested a directed graphical model comprising 7 inferred pathways of DMI progression: Initial to Mild, Initial to Superficial, Mild to Superficial, Mild to Moderate, Superficial to Moderate, Superficial to Severe, and Moderate to Severe. Eyes of the Mild and Superficial groups had greater increases in superficial NPS within the central sector than those of the Severe group. Additionally, deep NPS counts within the central sector decreased more in the eyes of the Initial group than in those of the Superficial and Moderate groups. Notably, the eyes of the Superficial and Moderate groups exhibited greater VA deterioration at 3 years compared with those in the Initial group.

Conclusions

A directed graphical model of DMI progression may serve as a useful tool for inferring progression pathways and predicting VA deterioration.

Financial Disclosure(s)

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

In the era of anti-VEGF treatment, diabetic macular ischemia (DMI) is the third vision-threatening diabetic retinopathy (DR), and effective therapeutic strategies should be established.^1,2 Capillary nonperfusion reduces the supply of oxygen and nutrients, subsequently leading to visual acuity (VA) reduction and poor retinal sensitivity.3, 4, 5, 6, 7, 8, 9, 10, 11 In addition, retinal ischemia is associated with 2 major pathological mechanisms: retinal edema and neurodegeneration.^1,12, 13, 14 Hypoxia-induced VEGF expression increases vascular hyperpermeability and plays a pivotal role in the pathogenesis of diabetic macular edema.^15,16 On the other hand, nonperfusion areas (NPAs) are accompanied by disorganization of retinal inner layers and disruption of the ellipsoid zone of photoreceptors on spectral-domain OCT images, which may represent layer-by-layer degenerative processes and resultant irreversible visual dysfunction.^14,17

Advances in fundus imaging are providing new insights into capillary nonperfusion. Retinal vessels in the macula reside in the ganglion cell layer and bifurcate into 2 capillary layers at the inner and outer boundaries of the inner nuclear layer.^18,19 A classic modality, fluorescein angiography, can delineate the foveal avascular zone (FAZ) but cannot visualize layer-by-layer capillaries in the parafovea.²⁰ In contrast, OCT angiography (OCTA) images with higher contrast have enabled the assessment of 3-dimensional vascular structures and the quantification of vascular parameters.^2,6,9,20, 21, 22 Automatic measurement of perfusion and nonperfusion metrics on OCTA images has improved the assessment of DMI severity compared with traditional FAZ parameters on fluorescein angiography images.¹ A recent publication has proposed a novel severity scale of DMI based on the distribution as well as the amounts of NPAs on OCTA images.²³ However, the gold standard for DMI diagnosis or grading remains to be determined.

Despite its clinical relevance, the processes of DMI progression remain poorly understood. In most cases, capillary nonperfusion exacerbates irreversibly, highlighting the importance of both preventive and therapeutic strategies for DMI. Vascular morphologies on OCTA images demonstrate a continuum from healthy intercapillary spaces to NPAs in eyes with DR.²⁴ On the other hand, we often observe unique patterns of capillary nonperfusion in eyes with DMI, e.g., FAZ enlargement, capillary dropout in the temporal subfield, and nonperfusion predominantly in the deep layer.²⁵ This suggests that the clinical characterization of DMI progression sheds light on the neurovascular degeneration in DMI.

In this study, we investigated the progression pathways of DMI on OCTA images and the VA deterioration in each pathway after 3 years.

Methods

Participants

In this prospective study, 162 consecutive eyes of 162 patients with DR were initially enrolled at the Department of Ophthalmology in Kyoto University Hospital between July 2017 and December 2020. However, 11 patients were lost to follow-up over the 3-year period, resulting in a final cohort of 151 eyes. This cohort is completely independent of our previous study.²³ Among them, 103 eyes that received no ocular treatment during the follow-up period were included in the analysis of VA changes. This subgroup comprised both treatment-naïve and previously treated cases. This study was approved by the Kyoto University Graduate School and Faculty of Medicine Ethics Committee and adhered to the tenets of the Declaration of Helsinki. Written informed consent was obtained from all participants before inclusion in the study.

We evaluated eyes with DR for which swept-source OCTA (SS-OCTA) images of sufficient quality (signal strength index of ≥8) were acquired. The inclusion criteria were the presence of DR, the acquisition of 3 × 3 mm SS-OCTA images centered on the fovea at baseline and at the 3-year visit, and written informed consent. Exclusion criteria were severe media opacity, other ocular diseases that lead to visual impairment, an axial length of <22 mm or >26 mm, any other chorioretinal disease, including epiretinal membrane and vitreomacular traction, glaucoma, previous treatment within 6 months before imaging (e.g., anti-VEGF injection, ocular steroids, intraocular surgery, or photocoagulation), and vitrectomy with macular interventions such as membrane peeling during the follow-up period. For VA subgroup analyses, eyes that received any ocular treatment during the follow-up period were excluded. Additional exclusion criteria were severe segmentation error in the superficial slab or poor image quality (signal strength index ≤7). If both eyes met these criteria, the right eye was selected for this study.

Image Acquisition and Processing

After measurement of best-corrected decimal VA using a full protocol refraction and its conversion to the logarithm of the minimum angle of resolution (logMAR) VA, comprehensive ophthalmic examinations were performed, including assessment of DR severity based on the International Clinical Diabetic Retinopathy Severity Scales.²⁶ Central subfield thickness and axial length were measured using Spectralis OCT (Heidelberg Engineering) and partial coherence interferometry (IOL Master, Carl Zeiss Meditec, Inc), respectively. Eyes with a central subfield thickness of >305 μm or 320 μm for female or male patients, respectively, were diagnosed as center-involved diabetic macular edema.²⁷

We acquired SS-OCTA images within the nominal 3 × 3 mm square centered on the fovea using Plex Elite 9000 (Carl Zeiss Meditec, Inc). The images of the left eye were horizontally inverted to standardize the nasal subfield positioning. The en face OCTA images with 300 × 300 A-scans were digitally converted to a 1024 × 1024 pixel array for quantitative analyses. The en face images in the superficial layer (from the inner limiting membrane to the boundary of the inner plexiform layer and inner nuclear layer) were created based on the default setting of the installed software. The deep en face images (from the inner border of the inner nuclear layer to 70 μm above the retinal pigment epithelium) were prepared based on the custom setting of the same software, as described previously.²⁸

We semiautomatically determined the amounts and locations of NPAs on SS-OCTA images as described previously.²³ Briefly, the edges of retinal vessels were automatically detected, utilizing the Canny Edge Detector plugin of ImageJ software (National Institutes of Health, http://imagej.nih.gov/ij/).²⁹ The 2.5 mm circle was divided into 15 × 15 pixel squares, and the square with no signals of vascular edges was defined as nonperfusion squares (NPSs). To explore the distribution of capillary nonperfusion, we counted the NPSs in 5 sectors, i.e., the central (1-mm area) and 4 parafoveal (1–2.5 mm) subfields (nasal, superior, temporal, and inferior) of the ETDRS grid.

Inference of Progression Pathways

We employed 4 steps of data analysis of high-dimensional data in order to explore the progression pathways of DMI: (1) embedding of each case onto 2-dimensional uniform manifold approximation and projection (UMAP) space, (2) determination of the cluster to which each case belonged, (3) calculation for major transitions, and (4) construction of a directed graphical model to show inferred pathways.³⁰ A recent publication has demonstrated a severity scale of DMI based on UMAP and k-means clustering in a cross-sectional study.²³ Briefly, each eye with high-dimensional data was mapped onto the 2-dimensional UMAP space by developing the manifold, as reported by McInnes et al.³⁰ Subsequent clustering divided all cases into 5 grades—Initial, Mild, Superficial, Moderate, and Severe—based on the amounts and spatial distribution of NPSs. As we previously reported, the Initial group had minimal nonperfusion in both superficial and deep layers, while the Mild group showed a slight increase, mainly in the deep layer. The Superficial group had larger NPAs in the superficial layer, especially in the central 1-mm area. The Moderate group exhibited NPAs extending into the temporal subfield in the deep layer, and the Severe group showed extensive nonperfusion in both layers.²³ Each eye at baseline and the 3-year visit was embedded into this space by the calculation of Euclidean distance to 3 nearest neighbors. We then determined the severity grade to which each eye belonged at baseline and the 3-year visit by identifying the predominant cluster among the 3 nearest points (k = 3) using the k-nearest neighbors method.

This study focused on observing cluster transitions at 3 years, defining “major transitions” as those from 1 grade to another involving ≥4 cases in this study. Among them, the pathways of DMI progression connecting adjacent clusters were explored to construct a directed graphical model because we hypothesized that capillary nonperfusion progresses gradually rather than continuously. We applied the following steps: (1) the determination of the starting and terminal clusters, (2) the establishment of the connections between clusters, (3) the application of conditional probability for pruning transitions, and (4) the iteration of steps 2 and 3, moving from terminal to starting clusters. First, we arranged the clusters in order according to the numbers of NPSs and determined the clusters with the least and most NPSs as the starting and terminal clusters, respectively. We then prepared the directed graph from the terminal cluster and the second one, based on the major transitions. When the third cluster was incorporated, 1-step or 2-step transitions were considered. We hypothesized 1-step or 2-step transitions between clusters (A→B or A→B→C) during 3 years and defined the probabilities of 2-step transitions (P_A→B→C) as P_A→B (probability of transition from cluster A to B) × P_B→C (probability from cluster B to C). When the sum of probabilities of possible 2-step transitions was less than the actual probabilities, we considered that the direct 1-step transition (A→C) must be present. In contrast, when the summed probabilities were greater, we rejected the 1-step transition (A→C) in this study. These processes were repeated until the starting cluster was reached. The directed arrows in the graphical model were referred to as “inferred pathways” of DMI progression in this study.

Statistical Analyses

All values were expressed as the median (interquartile range). Statistical significance was set at P < 0.05. We applied the Kruskal–Wallis test with a Bonferroni correction for continuous variables and the chi-square test for categorical variables. The statistical associations were evaluated using Spearman rank correlation coefficient. Multiple regression analysis was employed to assess the baseline factors associated with the changes in logMAR VA, using significant factors (P < 0.05) from univariate analyses (diabetic macular edema, central subfield thickness, and NPS counts in the deep layer). All statistical analyses were conducted using SPSS (version 24; IBM).

Results

Progression Pathway of DMI in a Longitudinal Study

We explored the NPS changes at 3 years in 151 eyes statistically (Table S1, available at www.ophthalmologyscience.org) and observed an overall increase in NPSs in both the superficial and deep layers (P < 0.001) (Fig 1). There are significant correlations between NPS counts at baseline and at 3 years in the superficial, deep, and both layers (Fig 1). Notably, some eyes with fewer baseline NPSs experienced significant NPS increases in the superficial layer. In the deep layer, most eyes with fewer NPSs at baseline had smaller increases in NPS counts (Fig 1).

Figure 1.

Nonperfusion square counts at baseline and after 3 years in all 151 eyes with diabetic retinopathy. A, The superficial layer. B, The deep layer. C, Both layers. NPS = nonperfusion square.

To investigate the pathways of DMI progression, we mapped each eye at baseline and the 3-year visit onto the 2-dimensional UMAP space established in a previous publication.²³ Eyes shifted from the right upper corner to the left lower corner over time, indicating NPA deterioration and subsequent DMI progression (Fig 2A, B). Eyes with DR underwent transitions between clusters at unique frequencies (Table S2, Fig S3, available at www.ophthalmologyscience.org). Each major transition involved specific amounts of NPS changes (Table 3, Table 4, and S5, available at www.ophthalmologyscience.org). This led us to hypothesize the main pathways of NPS progression and construct a directed graphical model (Fig 2C). Seven inferred pathways of DMI progression were identified: Initial to Mild, Initial to Superficial, Mild to Superficial, Mild to Moderate, Superficial to Moderate, Superficial to Severe, and Moderate to Severe.

Figure 2.

Embedding of all 151 eyes with diabetic retinopathy at baseline (A) and 3 years (B) onto the 2-dimensional UMAP space generated from the cross-sectional study in a prior publication.²³ Colored dots represent the dataset from the prior study, while hollow and filled black dots indicate the longitudinal cases analyzed in the current study. A, Eyes at baseline (hollow dots) are randomly located throughout the map. B, Eyes at 3 years (filled black dots) have shifted toward the left lower corner. C, Seven inferred pathways of diabetic macular ischemia progression are depicted. Arrow = inferred pathway; Black dot = centroid of each cluster; UMAP = uniform manifold approximation and projection.

Table 3.

Nonperfusion Square Changes in the Superficial Layer at Major Transitions between Clusters

Baseline
	Initial	Mild	Superficial	Moderate	Severe
3 yrs
Initial	52 (6–106)	-	-	-	-
Mild	105 (64–160)	64 (42–108)	-	-	-
Superficial	276 (174–451)	183 (76–284)	141 (100–175)	-	-
Moderate	173 (45–199)	151 (87–279)	206 (187–305)	130 (59–187)	-
Severe	-	200 (144–312)	84 (65–272)	220 (110–361)	184 (96–308)

Major transition = transition with ≥4 cases.

Table 4.

Nonperfusion Square Changes in the Deep Layer at Major Transitions between Clusters

Baseline
	Initial	Mild	Superficial	Moderate	Severe
3 yrs
Initial	60 (−5 to 176)	-	-	-	-
Mild	293 (268–345)	66 (56–97)	-	-	-
Superficial	442 (364–488)	270 (181–306)	149 (97–172)	-	-
Moderate	389 (365–518)	398 (311–496)	396 (364–418)	321 (235–482)	-
Severe	-	479 (417–783)	538 (386–662)	400 (315–568)	314 (243–412)

Major transition = transition with ≥4 cases.

Statistical comparisons demonstrated no significant differences in NPS changes across clusters, although specific trends were noted (Fig 4A, B). In the Severe group, the superficial NPS counts in the central sector remained stable, contrasting with their increases observed in the Mild and Superficial groups (Fig 4C). Eyes in the Initial group experienced more significant worsening in the deep NPS counts compared with those in the Superficial or Moderate groups (Fig 4D). In all parafoveal sectors, the NPS counts increased across all 5 clusters (Fig 4E, F).

Figure 4.

Nonperfusion square changes in each cluster of all 151 eyes with diabetic retinopathy. All sectors (A, B). The central sector (C, D). All parafoveal sectors (E, F). The superficial (A, C, E) and deep (B, D, F) layers. ∗P < 0.05; ^†P < 0.01. NPS = nonperfusion square.

Clinical Relevance of the DMI Progression Model

To evaluate the clinical relevance of our proposed model of DMI progression, we analyzed data from 103 untreated cases in a longitudinal cohort (Table S6, available at www.ophthalmologyscience.org). Visual acuity deteriorated in 46 eyes at 3 years (Fig S5, available at www.ophthalmologyscience.org). Multivariate analysis revealed no significant associations between baseline parameters and VA changes (Table S7, available at www.ophthalmologyscience.org). In contrast, worsening of logMAR VA was significantly associated with increases in parafoveal NPSs in the deep layer (ρ = −0.291, P = 0.003) (Table S8, available at www.ophthalmologyscience.org).

Certain cases within the Superficial group had poor logMAR VA at 3 years, but not at baseline (Fig 6A, B).²³ Notably, eyes in the Superficial and Moderate groups demonstrated more pronounced VA deterioration compared with those in the Initial group, suggesting that classification into these groups could serve as a predictor of future VA worsening at 3 years (Fig 6C). We further investigated VA changes during major transitions between clusters (Table S9, available at www.ophthalmologyscience.org). Transitions from the early stage (Initial and Mild groups) to the intermediate stage (Superficial and Moderate groups) were not accompanied by VA worsening. In contrast, VA deteriorated at 3 years in eyes that progressed from the Mild, Superficial, and Moderate groups to the Severe group or those that remained in the Superficial and Moderate groups (Table 10).

Figure 6.

Comparisons of logMAR VA across clusters in all 103 untreated eyes with diabetic retinopathy. LogMAR (A) at baseline and (B) at 3 years. C, Change in logMAR at 3 years. ∗P < 0.01. logMAR = logarithm of the minimum angle of resolution; VA = visual acuity.

Table 10.

Logarithm of the Minimum Angle of Resolution Visual Acuity Changes at Major Transitions between Clusters in Untreated 103 Eyes

Baseline
	Initial	Mild	Superficial	Moderate	Severe
3 yrs
Initial	0.023 (0.000–0.084)	-	-	-	-
Mild	0.000 (0.000–0.059)	-	-	-	-
Superficial	0.103 (0.073–0.109)	0.000 (−0.040 to 0.091)	−0.097 (−0.368 to 0.000)	-	-
Moderate	-	0.000 (−0.097 to 0.000)	0.000 (−0.125 to 0.000)	−0.079 (−0.125 to −0.046)	-
Severe	-	−0.097 (−0.137 to −0.023)	−0.125 (−0.207 to −0.073)	−0.051 (−0.176 to 0.000)	−0.026 (−0.145 to 0.059)

Major transition = transition with ≥4 cases.

Discussion

In this longitudinal study, we developed a directed graphical model to infer DMI progression pathways based on a recently reported severity scale.²³ To validate its clinical utility and investigate DMI progression over time, we established an independent cohort of 151 eyes with a 3-year follow-up. We selected this interval because preliminary data suggested that NPS changes were minimal during the first 1 or 2 years (data not shown). Further investigation regarding VA changes revealed that VA deteriorated at 3 years in the eyes of the Superficial and Moderate groups compared with those of the Initial group. In contrast, eyes with transition from the Initial or Mild group to the Superficial or Moderate group had no VA deterioration. These findings identify the Superficial, Moderate, and Severe groups as vision-threatening DMI, although future studies should elucidate the frequency of these transitions and the risk factors for progression over longer periods.

We applied a directed graphical model rather than statistical analyses to explore the clinical progression of DMI.31, 32, 33, 34 We then demonstrated its chronological outline: Initial and Mild as the first step without VA reduction, Superficial and Moderate as the second step, which predicts VA deterioration, and Severe as the third step with poor VA. Although there are no gold standards for the diagnosis and severity scales of the DMI continuum, this study suggests clinically relevant severity grades.²³ In other words, this graphical model for DMI progression might identify the Superficial, Moderate, and Severe groups as therapeutic targets and the Mild and Moderate groups for the preventive strategies. Despite its usefulness, we should incorporate statistical analyses in future studies to identify factors predictive of DMI progression or VA deterioration.

The comparisons of NPS counts between baseline and 3 years suggest the progression of DMI, which is consistent with rare transitions to less severe groups. We therefore constructed the directed graphical model to interpret DMI progression concisely, compared to the stepwise severity grades based on the FAZ morphologies on fluorescein angiography images.³¹ The graphical model elucidated 2 main pathways during the intermediate stages: Superficial and Moderate. The presence of these 2 distinct pathways may explain the different phenotypes of DMI.¹ The superficial vascular plexus primarily supplies the inner retina, and its ischemia has been linked to damages in the retinal nerve fiber layer and ganglion cell layer.³⁵ In contrast, the deep capillary plexuses, which supply the outer retina and photoreceptors, are more closely associated with ellipsoid zone disruption and photoreceptor degeneration.³⁶ Future studies should elucidate the functional and morphological biomarkers of neurodegeneration in each group and the processes of neurovascular degeneration leading to the Severe group.

Visual acuity reduction is generally associated with a decrease in the perfusion metrics or an increase in the nonperfusion metrics in DR, in accordance with previous publications.4, 5, 6, 7, 8, 9^,37 In this study, VA deteriorated during the transition from the Mild, Superficial, and Moderate groups to the Severe group. We speculate that malnutrition and hypoxia within the NPAs irreversibly contribute to visual disturbance during the Superficial and Moderate stages of transitions to the Severe stage. These stages may share common pathological mechanisms, leading to concurrent degenerative processes in both neurons and vessels.³⁸ Neurotoxic factors might be released from or passed through vascular endothelial cells, continuing until the complete loss of vascular cells terminates the neurodegenerative processes.³⁹ Vice versa, neurodegeneration in the inner retinal layers might lead to loss of angiogenic factors and subsequent capillary obstruction.^40,41 Multivariate analysis showed that baseline NPS counts across all untreated eyes were not significantly associated with VA changes. In contrast, the graphical model based on the DMI severity scale revealed more pronounced VA deterioration during the intermediate stages, emphasizing the importance of a stage-specific approach in understanding DMI progression and its impact on visual function.

Visual acuity was not deteriorated at the transitions from the Initial or Mild groups to the Superficial and Moderate groups, despite aggravating capillary nonperfusion. It is consistent with no differences in VA between the Initial, Mild, Superficial, and Moderate groups in a cross-sectional study.²³ We hypothesized 2 possible explanations: the threshold of nutritional deficiency for retinal function and different mechanisms underpinning capillary nonperfusion with and without visual dysfunction. Retinas are redundantly nourished by retinal arteries and the choroidal vessels, so mild reduction of retinal capillaries might not influence the function of retinal neurons.⁸ As another perspective, we should consider that transient and permanent capillary occlusion reduces the flow signals on OCTA images. Blood cells, e.g., leukocytes, erythrocytes, and platelets, could induce transient plugging of capillary lumen, whereas neurovascular degeneration might aggravate cell death in vascular cells and permanent occlusion with extracellular matrix.^42,43

There are several limitations to this study. The single-center design with the inclusion and exclusion criteria may result in selection bias. In particular, several artifacts at the acquisition and creation of OCTA images may influence the quantitative analyses.⁴⁴ The use of a specific SS-OCTA device and image processing methods may limit the generalizability of our findings. We selected NPSs as a nonperfusion metric, although we have to validate the consistency with other perfusion and nonperfusion metrics.¹ Multicenter studies incorporating independent cohorts, diverse imaging technologies, and analytical algorithms will be necessary to validate and expand upon our results. We selected the UMAP for dimensionality reduction and better visualization of the data space, although future studies should confirm the reproducibility by other methods to reduce dimensionality.^30,45 In our prior study, the Superficial and Moderate groups were classified as distinct based on objective clustering.²³ However, given their shared status as intermediate stages in DMI progression, future studies may refine or even merge these groups based on additional data. Our directed graphical model considered 1-step and 2-step transitions but not 3-step transitions between clusters because the inferred probabilities of transitions were not consistent with the actual probabilities. When 3-step transitions were included, the discrepancy between actual and estimated probabilities increased (data not shown), suggesting that 3-step transitions are less likely to occur within a 3-year period. This could be further clarified through more frequent acquisitions of OCTA images in future studies. While our focus on major transitions provided valuable insights into the most clinically relevant changes, further investigation into the characteristics of eyes with minor transitions is warranted.

In conclusion, this longitudinal study validates the use of a novel classification system for DMI on OCTA images and introduces a directed graphical model for understanding its progression. These findings contribute to the ongoing effort to establish reliable diagnostic and staging criteria for DMI, with the ultimate goal of improving patient outcomes through timely and targeted interventions.

Manuscript no. XOPS-D-24-00483.