The Behavior of the External Limiting Membrane and Ellipsoid Zone in Diabetic Macular Edema: A Population AI-Assisted Optical Coherence Tomography Study

Abstract

Purpose

To investigate the behavior of the external limiting membrane (ELM) and ellipsoid zone (EZ) in diabetic macular edema (DME) using artificial intelligence (AI)-based optical coherence tomography (OCT) quantification.

Methods

This was a cross-sectional study. OCT scans of eyes affected by DME were analyzed using a validated AI platform to quantify intraretinal fluid (IRF), subretinal fluid (SRF), inflammatory hyper-reflective retinal foci (I-HRF), and the percentage of interruption of the ELM and EZ. Eyes were classified according to ELM/EZ interruption and grouped into four patterns of integrity: no interruption, EZ interruption >ELM, ELM > EZ, and ELM = EZ. The distribution of the ELM/EZ disruption and associations with other biomarkers, and clinical data, including visual acuity, were analyzed.

Results

We analyzed 2355 eyes from 1688 patients affected by DME. ELM interruption was observed in 21.9% of eyes and EZ interruption in 37.6%. A greater prevalence of EZ disruption was also identified in both previously treated and untreated eyes. Eyes with an interrupted ELM/EZ showed significantly greater IRF and SRF volumes and a more centrally located IRF distribution. I-HRF counts were significantly higher in eyes with predominant EZ damage. ELM disruption showed the strongest correlation with visual acuity and SRF burden.

Conclusions

Quantitative AI-based OCT analysis demonstrated that the ELM and EZ have distinct yet complementary relationships with fluid biomarkers and visual acuity. EZ alterations are more prevalent and associated with I-HRF increases, whereas ELM disruption appears to reflect deeper structural damage and stronger functional impairment. Integrating outer retinal layer integrity with other features of DME may provide a more comprehensive characterization of DME and support a more personalized prognostic assessment and therapeutic strategies.

Diabetic macular edema (DME) is the leading cause of moderate and severe visual impairment in the working-age population globally and affects approximately 11% to 15% of patients with diabetic retinopathy (DR).¹ DME has benefited in recent decades from a diagnostic revolution thanks to optical coherence tomography (OCT). Although the central subfield thickness (CST) has long been the main benchmark in clinical trials, it is now evident that the correlation between retinal thickness and best-corrected visual acuity (BCVA) is limited and a higher correlation has been identified with the structures of the outer retina several years ago.^2,3 In this context, two structures take a fundamental importance: the external limiting membrane (ELM) and the ellipsoid zone (EZ). The ELM represents the junction between Müller cells and photoreceptors and appears on OCT as the first hyper-reflective band external to the outer nuclear layer. The integrity of the tight junctions within the ELM, guaranteed by the presence of occludin protein, acts as a barrier against the infiltration of fluids and macromolecules, and it is distinct from the RPE, which forms the outer blood–retinal barrier.^4,5

The EZ, the second hyper-reflective band on OCT, corresponds with the portion of the outer segments of photoreceptors dense in mitochondria, and it is widely regarded as an OCT biomarker reflecting photoreceptor integrity.^1,6 To correctly quantify and understand the dynamics of these structures is essential: damage to the external retina is not only associated with edema, but it is also an important biomarker to define patient's possible functional recovery and best treatment.^7,8

Despite the clinical importance of the ELM and the EZ, their evaluation has historically been limited by the subjective nature of human analysis. The manual segmentation of these layers is a time-consuming process, subject to high inter- and intraobserver variability, which precluded their routine use in clinical practice and in large epidemiological studies. The integration of artificial intelligence (AI) systems and deep learning algorithms is redefining this scenario. The ability to identify and quantify the ELM and the EZ using AI allows us to better understand the behavior of these structures in treated or untreated eyes with DME, and their reciprocal changes influencing the retinal morphology.⁹

Methods

This multicenter, retrospective, observational study included consecutive eyes with DME collected in an anonymized way. The study adhered to the Declaration of Helsinki; local approvals and consent procedures followed national regulations, and only fully anonymized data were analyzed.

Study Population and Clinical Data

Eligible eyes had type 1 or type 2 diabetes, center‑involving DME, and a complete baseline spectral‑domain OCT dataset suitable for AI‑based quantitative analysis. For each eye, the database included age, laterality (right eye/left eye), CST, BCVA (Early Treatment Diabetic Retinopathy Study [ETDRS] letters), diabetes type and duration, previous treatment for DME (yes/no), time since first DME treatment (years), DR stage (nonproliferative DR [NPDR]/proliferative DR [PDR]), and the presence of an epiretinal membrane, defined according to Govetto et al.¹⁰ Eyes with missing core OCT or BCVA data or with inadequate OCT quality (low central map quality index) were excluded.

OCT Acquisition and AI‑Derived Biomarkers

All centers analyzed macular spectral‑domain OCT scans acquired using the Spectralis platform (Heidelberg Engineering, Heidelberg, Germany), with a standardized protocol: a 6 × 6 mm volumetric ETDRS map of 49 scans (high-speed modality, >12 automatic real-time tracking (ART), quality index of >28) centered onto the fovea and a high‑resolution linear foveal scan acquired in high resolution (>90 ART, quality index of >30).¹¹ A validated AI platform (Ophthal, Mr.Doc s.r.l., Rome, Italy, version 1.0) provided automated segmentation and quantification of the intraretinal fluid (IRF) volume (mm³), subretinal fluid (SRF) volume (mm³), IRF distribution within the 0- to 1-, 1- to 3-, and 3- to 6-mm rings (IRF1, IRF3, and IRF6, respectively, expressed as the percentage of the total IRF), and inflammatory hyper-reflective retinal foci (I-HRF) count on a high‑resolution foveal scan.¹² Outer retinal disruption was quantified as the percentage of interruption of ELM and EZ in the central 1 mm of the map central scan, with 0% indicating a continuous band and 100% complete loss, corresponding with the number of microns of nonvisualized band relative to the total scan length. The ELM and EZ were analyzed both as continuous variables (0%–100%) and dichotomized as intact vs disrupted. Combining ELM and EZ status yielded four ELM/EZ patterns: no disruption (ELM = 0 and EZ = 0), ELM interruption > EZ, EZ > ELM, and ELM = EZ (both ELM > 0 and EZ > 0, with the ELM and EZ percentages, respectively, greater than, smaller than, or identical to each other).

Statistical Analysis

Continuous variables are summarized as mean ± SD and median (interquartile range) and categorical variables as counts and percentages. The normality of continuous distributions was assessed with the Kolmogorov–Smirnov and Shapiro–Wilk tests, showing non‑normal distributions for outer retinal measures and several fluid metrics. Group comparisons between treatment‑naïve and previously treated eyes used Wilcoxon rank‑sum tests or t‑tests for continuous variables and χ² tests for categorical variables. Associations between the ELM/EZ phenotypes and AI‑OCT metrics were evaluated with linear mixed‑effects models, with each OCT or functional parameter (IRF, IRF1, IRF3, IRF6, SRF, ELM, EZ, I-HRF, BCVA, and CST) as the dependent variable and the ELM/EZ group as the fixed effect, including a patient‑level random intercept to account for correlation between fellow eyes. Type III tests of fixed effects provided global F statistics; when significant, pairwise least‑squares mean differences between groups were estimated with the Tukey–Kramer adjustment. Additional mixed‑effects models compared treated vs treatment‑naïve eyes within the ELM/EZ strata.

Pearson and Spearman correlation coefficients quantified the relationships between the ELM or EZ (continuous) and the IRF, IRF1, IRF3, IRF6, SRF, I-HRF, CST, and BCVA in the overall cohort and stratified by treatment status and by disruption status (ELM > 0, EZ > 0). An additional analysis described the ELM and EZ distributions by SRF presence (SRF = 0 vs SRF > 0). A two‑sided P value of <0.05 was considered statistically significant for all tests; P values for multiple comparisons were corrected using the Tukey–Kramer method.

Results

A total of 2355 eyes of 1688 patients with DME were included. The mean IRF volume was 0.80 ± 1.17 mm³, with a predominant distribution in the 1- to 3-mm and 3- to 6-mm rings (IRF1, 18.0 ± 20.7%; IRF3, 36.8 ± 19.6%; IRF6, 45.3 ± 29.2%). SRF was present in 301 eyes (12.8%), with a mean SRF volume of 0.047 ± 0.086 mm³; the I-HRF count was 81.4 ± 28.6. Overall, the ELM was interrupted (ELM > 0) in 516 eyes (21.9%), and the EZ was interrupted (EZ > 0) in 886 eyes (37.6%). The mean ELM and EZ disruptions were 39.9 ± 33.9% and 43.7 ± 36.2%, respectively. The CST and BCVA averaged 81.4 ± 28.6, 386 ± 123 µm, and 63 ± 18 ETDRS letters. The mean patients age was 67.7 ± 10.7 years and 1507 patients (89%) were affected by type 2 diabetes. The mean diabetes duration was 15.7 ± 10.7 years, with a greater duration in patients with type 1 diabetes (25.8 ± 16.0 years) compared with patients with type 2 diabetes (14.5 ± 9.1 years). A total of 1677 eyes (71.2%) were affected by NPDR, and 678 (28.8%) eyes were affected by PDR. There were 1951 eyes that were previously treated for DME, and 404 were previously untreated. The time from the first DME treatment (DME duration) was 2.9 ± 3.0 years.¹¹

Outer retinal damage was more common and more extensive in previously treated DME eyes than in treatment‑naïve eyes. Among eyes with an ELM > 0, 16.1% of naïve eyes vs 23.1% of treated eyes showed ELM disruption, with a mean interruption of 31.0 ± 31.3% vs 41.2 ± 34.1%, respectively (mixed‑effects P = 0.03). Similarly, EZ disruption (EZ > 0) was present in 28.7% of naïve eyes and 39.5% of treated eyes, with a greater mean disruption in treated eyes (32.0 ± 32.1% vs 45.5 ± 36.4%; mixed‑effects P = 0.0003). Thus, both the prevalence and extent of the ELM and especially EZ interruption were greater in previously treated eyes, consistent with a more advanced or more chronic DME phenotype (Fig. 1).

Figure 1.

Distribution of the interruption of the ELM and EZ in previously untreated and treated eyes. Bars represent the mean extent (%) of the area showing disruption of the ELM and the EZ, comparing previously treated eyes with untreated eyes. Colors distinguish the two markers (orange, ELM; yellow, EZ); within each marker, the two bars compare untreated and previously treated eyes. Error bars denote the 95% CIs of the mean. For both markers, alterations were significantly greater in previously treated eyes. P values were derived from the type III tests of fixed effects of a mixed-effects model, which accounts for the within-patient correlation between the two eyes (ELM: F = 4.74, P = 0.03; EZ: F = 13.47, P = 0.0003). CI, confidence interval; n, number of eyes.

Eyes with an interrupted ELM showed a significantly higher IRF volume compared with eyes with an intact ELM (P < 0.0001). Similarly, IRF volume was significantly greater in eyes with EZ interruption compared with those with an intact EZ (P < 0.0001). Regarding IRF distribution, eyes with an ELM interruption had a lower proportion of IRF in the 0- to 1-mm ring and a greater proportion in the 3- to 6-mm ring (P < 0.0001); a similar pattern was observed in the EZ-based comparison, with significant differences in the 0 to 1-mm and 3 to 6-mm rings (both P < 0.0001), while the IRF percentage in the 1- to 3-mm ring was not significantly different in eyes with ELM or EZ interruption compared with eyes without an interruption. SRF volume was significantly greater in eyes an with ELM interruption and in those with an EZ interruption (P < 0.0001). As expected, the EZ disruption percentage was markedly higher in eyes with an interrupted ELM; conversely, the ELM disruption percentage was significantly higher in eyes with an EZ interruption (P < 0.0001). Finally, the I-HRF did not differ significantly between the intact and interrupted ELM groups (P = 0.1362), whereas a significant increase was observed in eyes with an EZ interruption compared with an intact EZ (P < 0.0001) (Table).

Table.

Morphological Differences in DME Eyes With EZ or ELM Integrity/Interruption

Parameter	Intact ELM	ELM Interrupted	P Value	Intact EZ	EZ Interrupted	P Value
IRF, mm3	0.575 ± 0.843	1.598 ± 1.705	<0.0001	0.461 ± 0.647	1.361 ± 1.563	<0.0001
IRF distribution, %
0–1	19.7 ± 21.9	11.8 ± 14.1	<0.0001	20.3 ± 22.6	14.1 ± 16.3	<0.0001
1–3	36.9 ± 20.2	36.4 ± 17.4	0.7563	36.5 ± 20.7	37.1 ± 17.6	0.3343
3–6	43.4 ± 30.1	51.9 ± 24.9	<0.0001	43.2 ± 30.7	48.8 ± 26.2	<0.0001
SRF, mm3	0.001 ± 0.013	0.022 ± 0.067	<0.0001	0.001 ± 0.012	0.014 ± 0.053	<0.0001
EZ, %	7.1 ± 19.8	49.7 ± 38.4	<0.0001
ELM, %				0.8 ± 5.8	22.0 ± 32.5	<0.0001
I-HRF, n	81.0 ± 28.9	82.9 ± 27.4	0.1362	79.6 ± 28.2	84.3 ± 29.0	<0.0001

Values are mean ± SD. Boldface entries indicate statistical significance.

When the ELM and EZ were combined, four patterns emerged: no interruption (1413 eyes [60.0%]), EZ > ELM (701 eyes [29.8%]), ELM > EZ (200 eyes [8.5%]), and ELM = EZ (41 eyes [1.7%]) (Fig. 2). The EZ > ELM pattern was by far the most frequent group, whereas ELM > EZ and ELM = EZ were relatively uncommon. In the EZ > ELM group, the mean EZ disruption was 44.2 ± 36.5% and the mean ELM interruption 11.5 ± 22.9%; in the ELM > EZ group, the mean ELM interruption was 45.3 ± 32.6% and the EZ interruption was 21.4 ± 24.8%. ELM = EZ eyes showed severe and symmetric damage (the mean ELM and EZ interruptions were both 84.1 ± 33.0%).

Figure 2.

Representative examples of ELM and EZ interruption patterns on OCT. In each scan, the ELM interruption is highlighted in orange and the EZ interruption in yellow; vertical lines indicate the measurement references. (A) Pattern of no interruption. (B) Pattern interruption of ELM > EZ. (C) Pattern interruption of EZ > ELM. (D) Pattern interruption of ELM = EZ.

The four groups were associated with distinct AI‑OCT and clinical profiles. The IRF volume was significantly higher in all disrupted groups compared with he no interruption group (P < 0.0001): the mean IRF was 0.44 mm³ in the no interruption eyes, 1.33 mm³ in the ELM > EZ eyes, 1.34 mm³ in the EZ > ELM eyes, and 1.31 mm³ in the ELM = EZ eyes. The IRF tended to be lower in the central ring and higher in the outer ring in the disrupted groups, indicating a relative shift of fluid toward the outer macula (IRF1 ≈ 12%–15% vs 20.5% in no alteration; IRF6 ≈ 47%–52% vs 42.9% in no interruption; overall for IRF in the outer ring P = 0.0002).

SRF showed a marked gradient across patterns (global P < 0.0001). The highest SRF volumes were observed in ELM > EZ eyes (0.024 ± 0.073 mm³), followed by ELM = EZ (≈0.01 mm³) and EZ > ELM, while the no interruption group had almost no SRF (≈0.001 mm³). Consistently, the dedicated SRF stratified analysis showed that both the ELM and the EZ percentages were higher in eyes with SRF > 0 than in those without SRF (mean ELM, 20.0% vs 7.1%; mean EZ, 26.9% vs 14.9%), and the proportion of eyes with an ELM > 0 or an EZ > 0 was also higher in the SRF‑positive group.

The I-HRF counts differed significantly across patterns (P = 0.0003). I-HRFs were lower in the no interruption group (79.7 ± 28.4) and the ELM > EZ group (79.5 ± 23.7), and higher in the EZ > ELM group (85.2 ± 29.6) and the ELM = EZ group (84.4 ± 33.5).

Eyes without ELM/EZ disruption had a better VA (69 ETDRS letters), while ELM = EZ eyes had the worst (51 ETDRS letters), and both the ELM > EZ and EZ > ELM groups had an intermediate BCVA (55–57 ETDRS letters), significantly inferior to the no interruption group (P < 0.0001), but similar to each other in the whole population, as well as when considering both untreated and treated eyes (P > 0.05).

When analyzed as continuous variables in eyes with disruption, both the ELM and the EZ percentages correlated positively with each other (P < 0.005; r > 0.4). A significant week negative correlation was also found between the ELM and visual function (P < 0.0001; r = −0.325) and the EZ and visual function (P < 0.0001; r = −0.336). When considering eyes not previously treated, a moderate negative correlation was found only for the ELM and visual function (P < 0.0001; r = −0.542 for ELM; r = −0.314 for EZ).

The duration of diabetes and DME, age, DR stage, epiretinal membrane presence, and diabetes type showed milder but significant differences across the ELM/EZ groups. EZ > ELM eyes tended to have a slightly longer diabetes duration (16.3 ± 11.1 years) and a greater prevalence of type 2 diabetes (∼89%) and PDR (∼32%) compared with the no interruption group. ELM = EZ eyes, despite representing a small subgroup, showed very severe outer retinal disruption with the worst VA and relatively high rates of epiretinal membrane presence and PDR.

Discussion

The relevant role of the integrity of the outer retinal layers in DME has long been recognized, particularly from a functional perspective. Several clinical series have established EZ integrity as a robust predictor of VA and treatment response in DME, both at baseline and after intravitreal therapy.^13–15 Maheshwary et al¹⁶ showed that each percent increase in EZ disruption was associated with approximately 0.3 ETDRS‑letter loss after adjusting for macular volume in DME, highlighting the tight quantitative coupling between the EZ and vision. In parallel, other studies have suggested that ELM integrity may be an even better predictor of visual recovery after treatment than the EZ or CST, with each increment in ELM integrity yielding a measurable visual gain.^17–19 This finding underlines the relevance of the interactions between Müller cells and photoreceptors in maintaining visual function, as suggested by a study analyzing the status of the ELM in the central macula of individuals with DME.²⁰ Moreover, it has also been shown that the interphotoreceptor matrix undergoes light-dependent structural reorganization. In the setting of DME, disruption of the outer retinal structure and interphotoreceptor matrix organization, leading to an altered ELM–RPE thickness, may contribute to visual dysfunction.²¹ However, to date, the evaluation of the ELM and EZ is difficult to standardize and is often assessed subjectively and nonquantitatively.^20,22 Modern AI models are able to perform an automated segmentation of the outer retina with high precision. The real innovation lies in AI's ability to provide objective and repeatable quantification, calculating the amount of the intact area of the EZ and the ELM and tracing structural microvariations in response to therapies (anti-VEGF or corticosteroids) with extreme sensitivity. This approach allows the provision of comparable data between different operators and clinical centers, setting new standards for precision medicine in ophthalmology.¹²

Recently, distinct DME clusters have been identified through the statistical combination of quantified morphological OCT parameters, allowing for a standardized classification of the disease. Previous AI-assisted analyses have shown that alterations of the ELM and EZ are heterogeneously distributed and exhibit different patterns across the various DME clusters.¹¹ The present study confirms that outer retinal integrity, captured by quantitative ELM and EZ interruption, is a central structural determinant in DME morphology and refines the relative roles of these two bands within a large, real‑world multicenter cohort. We found a greater prevalence and extent of ELM and EZ damage in previously treated eyes, suggesting that DME therapies, although effective in reducing fluid, have a limited impact on the inflammatory component, often begin in eyes with already advanced structural injury, and may not fully prevent cumulative outer retinal damage over time. This observation is aligned with various prospective cohorts studies, which identified baseline EZ disruption as a predictor of limited visual recovery despite anatomical improvement after anti‑VEGF treatment.^7,23–25 Our data emphasizes that, in addition to the EZ, early preservation of the ELM should be a therapeutic target; once ELM ≥ EZ damage is established, BCVA appears to reach lower levels.

We also found that EZ disruption was more prevalent than ELM disruption (37.6% vs 21.9%) and that eyes with EZ > ELM interruption represented the most frequent pattern of outer retinal damage, whereas the ELM ≥ EZ pattern (ELM > EZ or ELM = EZ) was relatively uncommon. The current literature suggests that the pattern of damage varies according to the stage of the disease, the type of edema, and the response to treatment.^1,26 Clinical and experimental data support a sequence in outer retinal damage in DME, where the EZ appears as more extensively damaged in DME development, but an intact ELM is a prerequisite for subsequent restoration of the EZ after therapy.^1,17 In this context, our observation that the ELM = EZ eyes (severe, symmetric ELM/EZ disruption) have the lowest BCVA reinforces the concept that the ELM status encodes the “structural reserve” of the photoreceptor apparatus.²⁷ However, a minority of cases may show a greater disruption in the ELM compared with the EZ, as visualized on OCT scans.²⁶ The apical processes of Müller cells are attached to one another and to the inner segments of photoreceptor cells by continuous heterotypic adherens junctions.²⁸ The reflectivity of the ELM band is due to Müller cell junctional complexes, although Müller cell microvilli may also contribute to the width of the band.²⁹ In DR, glial Müller cells are swollen and they lose their occludin content at the ELM level, leading to the disruption of normal retinal morphology and cyst formation.²⁸ When Müller cell involvement is particularly prominent, such as in the presence of marked cytotoxin-induced edema with cellular swelling and loss of junctional proteins at the ELM level, morphological alterations of the ELM may be disproportionately severe relative to those of the EZ, resulting in poorer OCT visualization of the ELM.

Looking at the other evaluated morphological parameters, this analysis shows clear morphological differences between eyes with DME according to the integrity of the outer retinal layers. Eyes with disruption of either the ELM or EZ showed a significantly greater IRF burden compared with eyes with preserved structures. In particular, IRF volume was almost threefold higher in eyes with ELM interruption and similarly increased in eyes with EZ disruption, suggesting that outer retinal damage is associated with more severe IRF accumulation. Consistently, eyes with altered outer retinal layers showed a lower proportion of minimal IRF (0–1 mm³) and a greater prevalence of more extensive IRF (3–6 mm³), indicating that structural damage of the photoreceptor–Muller cells complex is associated with more advanced morphological disease. A similar pattern was observed for the SRF, which was significantly more represented in eyes with a disrupted ELM or EZ. Indeed, the relationship of outer retinal integrity with DME seems to be bidirectional. Chronic edema can induce photoreceptor injury through mechanical and metabolic stress, while microvascular dysfunction and retinal inflammation may further compromise mitochondrial function.³⁰ Conversely, the preservation of EZ integrity after anti-VEGF treatment has consistently been associated with better visual outcomes, supporting its role as a biomarker of photoreceptor health rather than merely an anatomical feature on OCT.⁷

Notably, the disruption of one outer retinal layer was strongly associated with alterations of the other one, confirming the close structural and functional relationship between the ELM and the EZ.³¹ As regards the I-HRF, a significantly higher number in eyes with EZ disruption documents a close association between inflammatory activity and photoreceptor damage. I-HRFs are widely considered a marker of activated microglia and neuroinflammation in DME.³² Their increased prevalence in eyes with EZ impairment could therefore reflect a microenvironment characterized by sustained inflammatory stress leading to photoreceptor dysfunction and structural disruption of EZ. Inflammatory activity contributes to outer retinal damage, but is not exclusively linked to ELM alterations, which may reflect a more complex inflammatory component of DME, characterized also by mechanical or structural changes related to Müller cell impairment. This may be related to the fundamental role of the glioneurovascular unit, in which Müller cells play a key role in maintaining retinal homeostasis.³³ Consequently, ELM disruption may represent a more reliable indicator of global retinal integrity and potential visual recovery than other outer retinal structures.

Looking at the groups of eyes with ELM > EZ, or EZ > ELM, EZ = ELM, or both intact, outer retinal damage seems to be tightly linked with both the amount and the topography of the IRF. All of the disrupted groups showed a higher IRF than eyes with an intact ELM/EZ, with a relative shift of the IRF from the central 1‑mm ring toward the 3- to 6‑mm ring, suggesting that the alterations in the outer retinal layers are unlikely to result exclusively from the mechanical stress generated by fluid accumulation in the central fovea. Rather, these findings support the presence of a more generalized disturbance of macular homeostasis, involving the broader macular region and reflecting a complex interplay between fluid accumulation, retinal structural damage, and underlying inflammatory and microvascular mechanisms.

In addition to a high IRF volume, ELM > EZ eyes had the highest SRF volumes. The ELM, composed of zonulae adherents between Müller cells and photoreceptors, functions primarily as a size-selective diffusion barrier, allowing the free passage of water molecules, differently from the RPE tight-junction–based outer blood–retinal barrier.⁵ The ELM maintains the integrity of the interphotoreceptor matrix, restricting the diffusion of macromolecules; therefore, when the ELM is disrupted, these proteins can redistribute across the boundary, altering local oncotic gradients and promoting osmotically driven water accumulation.³⁴ This mechanism is consistent with the observation by Chung et al³⁵ that SRF in DME results from fluid movement through a weakened and permeable ELM or from increased permeability of the choriocapillaris through a dysfunctional RPE. Moreover, because the ELM is formed by junctional complexes at the apices of Müller cells, its disruption signals compromised Müller cell integrity.²⁸ Müller cell morphological and functional alterations contribute to the development of cystoid macular edema through multiple mechanisms, resulting in water influx, cellular swelling, and necrosis.³⁶ Thus, ELM disruption may serve as a structural biomarker of impaired glial water homeostasis rather than a direct cause of water movement.^36–39

By contrast, I-HRFs were more abundant in the EZ > ELM and ELM = EZ groups than in no the interruption or ELM > EZ groups, pointing to a stronger link between EZ‑predominant damage and neuroinflammatory activity, as previously mentioned, and supporting a model where I-HRFs reflect ongoing disease activity, whereas ELM disruption captures the cumulative structural burden of chronic edema and neuroinflammation.²² This interpretation is supported by mechanistic data showing that Müller cells, whose endfeet form the ELM, regulate barrier permeability, oxidative stress, and inflammatory signaling in DR and that the loss of junctional proteins such as occludin at the ELM level accompanies chronic DME.⁴⁰ ELM disruption in DME reflects a convergence of inflammatory, metabolic, and mechanical insults on the Müller cells, with inflammatory/barrier-related injury likely playing the predominant role.^40,41 The ELM, therefore, appears to integrate information on both fluid dynamics and glial–photoreceptor health, making it a plausible biomarker for structural prognosis. Conversely, the EZ seems to remain sensitive to more dynamic changes in photoreceptor outer segments and disease activity: the mechanisms underlying the appearance of EZ band disruption on OCT are likely heterogeneous and may reflect impaired photoreceptor integrity or mitochondrial loss or disorganization, although photoreceptor misalignment may also contribute in some cases.^37,42,43 Therefore, the ELM does not appear as merely redundant with the EZ, but represents a more advanced level of structural injury: in our cohort, a significant correlation with VA was detected only with ELM damage.²⁷

A limitation of this study may be that it relies exclusively on imaging data. Although current devices allow for excellent image quality, and the images collected in this study met predefined quality parameters that were high relative to standard clinical practice, OCT scans may not always be able to represent the underlying anatomy in detail, particularly in the presence of structural damage. This factor is particularly relevant in eyes with significant macular edema or clusters of hard exudates, where posterior shadowing may limit the visualization of the outer bands for both AI-based and human clinical assessment. It is nonetheless true that current clinical practice, as well as research in the field of DME, necessarily relies primarily on imaging modalities, and on OCT in particular. We, therefore, consider this study a fundamental first step toward a better understanding of the behavior of the ELM and the EZ in DME, based on quantitative and reproducible data obtained from high-quality OCT images. This is particularly true, looking at the continuous technological advances in OCT imaging, with the development of high-resolution devices capable of visualizing an increasing number of retinal layers and microstructural details.⁴⁴ Another main limitation of the present analysis is its cross-sectional design. Future studies based on these findings may help to clarify the longitudinal and reciprocal behavior of these structures over time, as well as their evolution throughout the course of the disease and in response to treatment. Moreover, longitudinal studies with standardized AI‑based quantification of the ELM/EZ will be essential to determine whether early changes in the ELM or shifts between ELM/EZ patterns may be used as surrogate endpoints or criteria to intensify or modify DME therapy. Finally, the ELM and EZ were analyzed only on the central scan passing through the fovea, and this approach may not capture the full spatial extent of outer retinal disruption across the macula, particularly considering potential inaccuracies in automated fovea detection. However, this approach was dictated by the options provided by the software and allows automated, reproducible measurements. Moreover, the central scan targets the most clinically relevant location for correlating photoreceptor integrity with visual function. Advances in imaging and analytic technology are likely to enable a more comprehensive and quantitative characterization of outer retinal layer disruption in the future.

Conclusions

Integrating the traditional assessment of fluid burden with the evaluation of neural tissue integrity, particularly within the outer retina, represents a promising paradigm in the management of DME. This integrated approach refines our prognostic capability and may support a more personalized therapeutic strategy aimed not only at resolving edema but also at preserving photoreceptor function.