Deep Learning Assessment of Intraretinal Microvascular Abnormalities in Diabetic Retinopathy on Swept-Source OCT Angiography Authors: Pei Qi Lo, Ser Koon Goh, Wenjun Song, Ming Hui Colin Goh, Zhongqing Yu, Ryan Zi Lin Tan, Pavan Paniharam, Xiaofeng Shen, Jianxie Yang, Bingyao Tan, Yue Qiu, Tai Yong Loh, Juling Sia, Wei Hing Seah, Lingyi Zhuang, Charmayne Xinyi Ang, Yusra Asad, Sowkath Ali Bashajan, Ying Jie Yong, Jia Jia Mah, Min Wang, Rupesh Agrawal, Sandy Wenting Zhou Published: 2026 ABSTRACT PurposeTo develop and evaluate a deep learning algorithm for automated segmentation and quantification of intraretinal microvascular abnormalities (IRMAs) on swept-source OCT angiography (SS-OCTA) across multiple fields of view (FOVs) and explore IRMA burden as an OCT angiography (OCTA) biomarker for stratifying diabetic retinopathy (DR) severity and correlation with retinal nonperfusion.DesignCross-sectional study.SubjectsA total of 254 eyes from 181 patients with varying stages of DR, including nonproliferative DR (NPDR) and proliferative DR (PDR).MethodsEyes of different DR stages were imaged using 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm scans. A context-enhanced U-Net (CE-U-Net) incorporating nonperfusion area priors and attention-based feature recalibration was trained to segment IRMAs on en face all-retinal-layer OCTA images.Main Outcome MeasuresPerformance was evaluated using lesion-level F1 score, intraclass correlation coefficients (ICCs) for quantitative agreement with ground-truth IRMA counts, and Bland–Altman analysis. Intraretinal microvascular abnormality counts were compared across DR severity stages and correlated with nonperfusion index (NPI).ResultsA total of 254 OCTA scans were analyzed (NPDR = 60.2%, PDR = 39.8%). The CE-U-Net achieved mean F1 scores of 0.61 ± 0.08, 0.71 ± 0.02, and 0.65 ± 0.05 for IRMA segmentation on 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm scans, respectively. Algorithm-predicted IRMA counts showed strong agreement with ground-truth across all FOVs (ICC = 0.81, 0.86, and 0.82, respectively). OCT angiography–detectable IRMAs were present from mild-to-moderate NPDR and increased markedly from mild-to-moderate NPDR to severe NPDR, particularly in 29 × 24 mm scans (P < 0.001), before plateauing in PDR (P = 0.68). The strongest correlation between IRMA counts and NPI was observed in 12 × 12 mm scans across all DR stages (P = 0.01, <0.001, and <0.001, respectively).ConclusionsAutomated IRMA segmentation and quantification on SS-OCTA is feasible and reproducible across multiple FOVs, including ultra-widefield scans. By reliably capturing both clinically visible and OCTA-detectable “occult” IRMAs, this approach supports IRMA burden as a quantitative biomarker for stratifying DR severity and advances the case for biomarker-driven DR assessment.Financial Disclosure(s)Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article. Intraretinal microvascular abnormalities (IRMAs)—characterized by dilated, tortuous intraretinal capillary segments—signify severe nonproliferative diabetic retinopathy (NPDR) under the ETDRS criteria.1 As surrogate markers for retinal ischemia which cannot be directly visualized on standard color fundus photography (CFP), their presence indicates a high risk of progression to proliferative disease and subsequent vision loss.1 Despite their clinical importance, CFP is an insensitive tool to detect IRMAs.2 While fundus fluorescein angiography remains the historical gold standard for visualizing retinal vasculature with higher sensitivity in IRMA detection,3 its invasive and time-consuming nature limits its clinical utility. Furthermore, dye leakage in fundus fluorescein angiography often obscures fine structural detail, hindering the precise morphological delineation of IRMA subtypes required for advanced staging.4 OCT angiography (OCTA) has substantially advanced our understanding of IRMAs by enabling noninvasive, depth-resolved, 3-dimensional visualization of the retinal microvasculature. Compared with conventional dye–based angiography, OCTA permits more precise identification and structural characterization of IRMAs with high sensitivity and specificity.5,6 It has also challenged the traditional paradigm that IRMAs are confined to severe NPDR: recent reports have identified them even in mild-to-moderate NPDR, indicating that IRMA burden increases progressively with disease severity and may serve as a biomarker for staging and predicting diabetic retinopathy (DR) progression.7 Longitudinal data further link higher IRMA counts to progressive visual deterioration, supporting their utility as independent prognostic markers.7 Yet, most existing OCTA studies rely on relatively narrow fields of view (FOVs), despite ultra-widefield (UWF) fundus fluorescein angiography evidence that a substantial proportion of IRMAs lie in the mid-peripheral and peripheral retina.8 In addition, IRMAs exhibit highly heterogeneous morphology across studies,7 which makes manual identification and grading labor-intensive and subject to interobserver variability. The markedly larger number of IRMAs detected on OCTA further amplifies this challenge,8 limiting their current clinical applicability. Moreover, the clinical significance of IRMA burden across different DR stages remains incompletely understood. Collectively, these limitations continue to hinder the translation of IRMA quantification into a practical clinical biomarker, highlighting the need for wider-field imaging and automated, standardized analytic approaches. To overcome these limitations and enable clinically scalable quantification of IRMAs, deep learning algorithms trained to segment OCTA-detectable IRMAs have the potential to provide reproducible detection. However, algorithm-based segmentation of IRMAs remains technically challenging because of their marked morphological7 and complex relationship with surrounding capillary networks.9 This difficulty is reflected in the 2022 public benchmark challenge, in which the top-performing algorithms achieved dice similarity coefficients (DSCs) of 29.53%–47.04%.10 In addition, existing deep learning models have been developed and validated primarily on narrow-field OCTA scans, typically limited to ≤12 × 12 mm, which do not capture the mid-peripheral retina where a substantial proportion of IRMAs are known to reside.11 In this study, we address these key gaps by developing a customized deep learning algorithm for the segmentation of OCTA-detectable IRMAs across multiple FOVs, including 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm scans. We further evaluate the quantitative IRMA burden derived from this approach as a potential imaging biomarker for DR severity. Methods This cross-sectional study was conducted at Tan Tock Seng Hospital, Singapore, between November 2024 and October 2025. The study was approved by the Institutional Review Board and followed the Declaration of Helsinki. Informed consent was obtained from all patients. Data Collection We enrolled adult patients (≥18 years) with type 1 or type 2 diabetes mellitus (DM) and a diagnosis of DR based on dilated fundus examination with further independent confirmation on wide-field fundus photos (Clarus 500; Zeiss). Diabetic retinopathy severity was graded according to the International Clinical Diabetic Retinopathy Severity Scale. Discrepancies between chart diagnoses of DR grading and widefield CFP observations were resolved by an independent senior retinal specialist (S.Z.W). Exclusion criteria were as follows: (1) Poor image quality, defined as signal strength index <7 or presence of significant artifacts; or (2) presence of other confounding ocular comorbidities and coexisting retinal diseases likely to confound lesion segmentation including retinal vein or artery occlusions or other retinal vascular pathologies; or (3) coexisting ocular pathology or retinal disease: high myopia (>6.0 D), significant media opacity such as dense vitreous hemorrhage, or visually significant cataracts that impair image acquisition, and glaucoma. All patients underwent a comprehensive ophthalmic examination. Electronic medical records were reviewed including patient demographics, DM type and duration, diabetic macular edema, glycosylated hemoglobin levels and comorbidities such as hypertension, hyperlipidemia, ischemic heart disease, cerebrovascular accident, and chronic kidney disease. Swept-Source OCTA Image Acquisition and Data Preparation All eyes were imaged using a 400 kHz widefield swept-source OCTA (SS-OCTA) device (VG200; Intalight Imaging, Ltd), operating at a central wavelength of 1050 nm. Retinal vascular plexuses were segmented automatically using the device software, with manual correction performed when automated segmentation errors were identified. For each eye, fovea-centered images were acquired at 3 scan protocols (6 × 6 mm, 12 × 12 mm, and 29 × 24 mm). Key acquisition parameters for each scan protocol are summarized in Table 1. Scanning was performed by trained photographers, and scans were repeated where necessary to ensure adequate image quality.Table 1Acquisition Parameters for OCT Angiography Scan ProtocolsScan ProtocolsVolume (A-Scans × B-Scans)RepetitionsSpacing between Adjacent A-Scans or B-Scans6 × 6 mm512 × 512411.7 μm12 × 12 mm1024 × 1024211.7 μm29 × 24 mm1536 × 1024219 μm Table 1Acquisition Parameters for OCT Angiography Scan ProtocolsScan ProtocolsVolume (A-Scans × B-Scans)RepetitionsSpacing between Adjacent A-Scans or B-Scans6 × 6 mm512 × 512411.7 μm12 × 12 mm1024 × 1024211.7 μm29 × 24 mm1536 × 1024219 μm Acquisition Parameters for OCT Angiography Scan Protocols Considering that IRMAs can be detected in both the superficial and deep capillary plexus slabs, with slightly different detection rates, the use of an all-retinal-layer (ARL) angiographic slab extending from the internal limiting membrane (ILM) to the retinal pigment epithelium was chosen for IRMA segmentation to minimize the risk of missing lesions. This was supplemented by the vitreoretinal interface (VRI) slab, with an upper boundary 200 μm above the ILM to a lower boundary 30 μm below the ILM. Bulk eye movement artifacts and projection artifacts were automatically removed by the device’s in-built deep learning algorithm. Cross-sectional OCT B-scans with flow overlay were used to exclude neovascularizations (NVs) by confirming whether the flow signal breached the ILM. Manual Annotation and Intergrader Agreement Two trained graders (L.P.Q. and G.S.K.) independently delineated IRMAs on en face ARL slab across all FOVs. Intraretinal microvascular abnormalities were defined as dilated or tortuous intraretinal vascular abnormalities, typically adjacent to nonperfused regions, and were confirmed on coregistered OCT B-scans with flow overlay to ensure no ILM breach. The en face VRI slab was additionally reviewed to aid exclusion of NVs; lesions demonstrating flow signal extending beyond the ILM were classified as NVs and annotated separately for the NV areas to be excluded. Annotation criteria and subtype classification were guided by published OCTA-based IRMA grading frameworks from previous publications.7,12 Intraretinal microvascular abnormalities were categorized into 5 morphologic subtypes shown in Figure 1. Dilated-trunk IRMAs were identified as enlarged intraretinal channels with caliber clearly greater than adjacent capillaries, often terminating without an obvious draining branch and appearing relatively straight or mildly curved. Looped IRMAs were defined by a circular or ring-like vascular configuration that originates from and returns to the same vessel segment. Sea-fan IRMAs exhibited a compact, triangular branching pattern arising from a narrow base, with feeding and draining vessels clustered at the base. Net-shaped IRMAs showed a more complex, mesh-like configuration with multiple branches and poorly distinguishable draining vessels, resulting in a broader, irregular, or rectangular network. Tufts were defined as focal “bud-like” clusters of dense, newly formed microvessels located at the distal ends of IRMAs, frequently near the border of nonperfusion.7,12Figure 1Different IRMA subtypes identified during manual labelling of ARL images, consisting of dilated trunk, loop, net-shaped, sea fan, and tufted subtypes. ARL = all-retinal-layer; IRMA = intraretinal microvascular abnormality. Different IRMA subtypes identified during manual labelling of ARL images, consisting of dilated trunk, loop, net-shaped, sea fan, and tufted subtypes. ARL = all-retinal-layer; IRMA = intraretinal microvascular abnormality. Intergrader agreement for IRMA quantification was assessed using intraclass correlation coefficients (ICCs) and was good to excellent across scan sizes, with the highest reliability in 12 × 12 mm scans (ICC 0.84, 95% confidence interval [CI] 0.74–0.90), followed by 6 × 6 mm (ICC 0.82, 95% CI 0.43–0.93) and 29 × 24 mm (ICC 0.82, 95% CI 0.74–0.88) scans (Table 4). A senior retina specialist (S.Z.W.) reviewed all annotations and adjudicated discrepancies to generate a consensus reference standard, which served as ground-truth for model training and evaluation.Table 2Baseline Demographic and Clinical Characteristics of Study Participants Stratified by Fields of ViewFOV6 × 6 mm12 × 12 mm29 × 24 mmNo. of eyes985898No. of patients713971Male, n (%)48 (67.6)22 (56.4)48 (67.6)Age, mean (SD)57.7 (11.3)55.0 (10.7)57.7 (11.3)Eyes with DME, n (%)36 (36.7)18 (31.0)36 (36.7)Patients with HTN, n (%)51 (71.8)17 (43.6)51 (71.8)Patients with HLD, n (%)55 (77.5)20 (51.3)55 (77.5)Patients with IHD, n (%)15 (21.1)2 (5.1)15 (21.1)Patients with CVA, n (%)16 (22.5)5 (12.8)16 (22.5)Patients with CKD, n (%)22 (31.0)4 (10.3)22 (31.0)CKD = chronic kidney disease; CVA = cerebrovascular accident; DME = diabetic macular edema; FOV = field of view; HLD = hyperlipidemia; HTN = hypertension; IHD = ischemic heart disease; NPDR = nonproliferative diabetic retinopathy; OCTA = OCT angiography; PDR = proliferative diabetic retinopathy; SD = standard deviation.Demographic and clinical variables are presented for eyes imaged using 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm OCTA scans, stratified by mild-to-moderate NPDR, severe NPDR, and PDR. Continuous variables are reported as mean (SD), and categorical variables are reported as counts (percentages).Table 3Algorithm Performance for All FOVs across DR SeverityFOVF1 Scores across DR SeverityOverall F1 ScoreP ValueMild-to-moderate NPDRSevere NPDRPDR6 × 6 mm0.58 ± 0.150.65 ± 0.160.61 ± 0.160.61 ± 0.080.4212 × 12 mm0.69 ± 0.140.70 ± 0.040.71 ± 0.040.71 ± 0.020.5629 × 24 mm0.66 ± 0.120.63 ± 0.070.66 ± 0.070.65 ± 0.050.07DR = diabetic retinopathy; FOV = field of view; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy.F1 scores were reported as mean ± standard deviation. The bold formatting highlights the overall F1 score of our algorithm, demonstrating its consistent performance across the different FOVs and its high accuracy in IRMA segmentation.Table 4IRMA Counts and ICC Across Different DR Severities and FOV SizesFOVDR SeverityMean Ground-Truth IRMA Count (SD)Mean Prediction IRMA Count (SD)Mean Difference (SD)Prediction vs. Ground-Truth ICC (95% CI)Expert Grader 1 vs. Expert Grader 2 ICC (95% CI)6 × 6 mmMild-to-moderate NPDR2.26 (1.97)1.51 (1.85)–0.76 (1.38)0.69 (0.40 – 0.84)0.62 (0.06 – 0.84)Severe NPDR5.65 (4.58)4.57 (4.20)–1.08 (3.50)0.67 (0.35 – 0.85)0.72 (0.32 – 0.89)PDR8.98 (6.37)9.37 (6.73)0.40 (4.71)0.75 (0.57 – 0.86)0.78 (0.18 – 0.92)Overall5.69 (5.56)5.34 (5.96)–0.35 (3.53)0.81 (0.73 – 0.87)0.82 (0.43 – 0.93)12 × 12 mmMild-to-moderate NPDR7.75 (8.48)9.00 (7.50)1.25 (2.82)0.93 (0.73 – 0.99)0.79 (0.26 – 0.95)Severe NPDR15.55 (10.90)15.10 (11.33)–0.45 (5.72)0.88 (0.61 – 0.97)0.82 (0.46 – 0.95)PDR26.42 (15.28)26.14 (13.62)–0.28 (9.08)0.81 (0.66 – 0.90)0.79 (0.64 – 0.89)Overall21.70 (15.33)21.60 (14.06)–0.10 (7.85)0.86 (0.77 – 0.92)0.84 (0.74 – 0.90)29 × 24 mmMild-to-moderate NPDR6.91 (7.47)5.65 (7.40)–1.26 (3.58)0.87 (0.76 – 0.94)0.80 (0.62 – 0.90)Severe NPDR25.89 (12.38)21.42 (12.16)–4.47 (7.41)0.77 (0.43 – 0.91)0.73 (0.44 – 0.89)PDR28.40 (15.23)23.40 (13.70)–4.99 (11.04)0.67 (0.43 – 0.82)0.69 (0.45 – 0.83)Overall20.04 (15.80)16.52 (14.05)–3.52 (8.37)0.82 (0.71 – 0.89)0.82 (0.74 – 0.88)CI = confidence interval; DR = diabetic retinopathy; FOV = field of view; ICC = intraclass correlation coefficient; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy; SD = standard deviation.Mean difference is computed by subtracting ground-truth from prediction values. Agreement between manual grading and algorithm predictions was assessed using Bland–Altman mean differences and ICC. Intraclass correlation coefficient between manual expert graders is also reported. Table 2Baseline Demographic and Clinical Characteristics of Study Participants Stratified by Fields of ViewFOV6 × 6 mm12 × 12 mm29 × 24 mmNo. of eyes985898No. of patients713971Male, n (%)48 (67.6)22 (56.4)48 (67.6)Age, mean (SD)57.7 (11.3)55.0 (10.7)57.7 (11.3)Eyes with DME, n (%)36 (36.7)18 (31.0)36 (36.7)Patients with HTN, n (%)51 (71.8)17 (43.6)51 (71.8)Patients with HLD, n (%)55 (77.5)20 (51.3)55 (77.5)Patients with IHD, n (%)15 (21.1)2 (5.1)15 (21.1)Patients with CVA, n (%)16 (22.5)5 (12.8)16 (22.5)Patients with CKD, n (%)22 (31.0)4 (10.3)22 (31.0)CKD = chronic kidney disease; CVA = cerebrovascular accident; DME = diabetic macular edema; FOV = field of view; HLD = hyperlipidemia; HTN = hypertension; IHD = ischemic heart disease; NPDR = nonproliferative diabetic retinopathy; OCTA = OCT angiography; PDR = proliferative diabetic retinopathy; SD = standard deviation.Demographic and clinical variables are presented for eyes imaged using 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm OCTA scans, stratified by mild-to-moderate NPDR, severe NPDR, and PDR. Continuous variables are reported as mean (SD), and categorical variables are reported as counts (percentages). Baseline Demographic and Clinical Characteristics of Study Participants Stratified by Fields of View CKD = chronic kidney disease; CVA = cerebrovascular accident; DME = diabetic macular edema; FOV = field of view; HLD = hyperlipidemia; HTN = hypertension; IHD = ischemic heart disease; NPDR = nonproliferative diabetic retinopathy; OCTA = OCT angiography; PDR = proliferative diabetic retinopathy; SD = standard deviation. Demographic and clinical variables are presented for eyes imaged using 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm OCTA scans, stratified by mild-to-moderate NPDR, severe NPDR, and PDR. Continuous variables are reported as mean (SD), and categorical variables are reported as counts (percentages). Table 3Algorithm Performance for All FOVs across DR SeverityFOVF1 Scores across DR SeverityOverall F1 ScoreP ValueMild-to-moderate NPDRSevere NPDRPDR6 × 6 mm0.58 ± 0.150.65 ± 0.160.61 ± 0.160.61 ± 0.080.4212 × 12 mm0.69 ± 0.140.70 ± 0.040.71 ± 0.040.71 ± 0.020.5629 × 24 mm0.66 ± 0.120.63 ± 0.070.66 ± 0.070.65 ± 0.050.07DR = diabetic retinopathy; FOV = field of view; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy.F1 scores were reported as mean ± standard deviation. The bold formatting highlights the overall F1 score of our algorithm, demonstrating its consistent performance across the different FOVs and its high accuracy in IRMA segmentation. Algorithm Performance for All FOVs across DR Severity DR = diabetic retinopathy; FOV = field of view; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy. F1 scores were reported as mean ± standard deviation. The bold formatting highlights the overall F1 score of our algorithm, demonstrating its consistent performance across the different FOVs and its high accuracy in IRMA segmentation. Table 4IRMA Counts and ICC Across Different DR Severities and FOV SizesFOVDR SeverityMean Ground-Truth IRMA Count (SD)Mean Prediction IRMA Count (SD)Mean Difference (SD)Prediction vs. Ground-Truth ICC (95% CI)Expert Grader 1 vs. Expert Grader 2 ICC (95% CI)6 × 6 mmMild-to-moderate NPDR2.26 (1.97)1.51 (1.85)–0.76 (1.38)0.69 (0.40 – 0.84)0.62 (0.06 – 0.84)Severe NPDR5.65 (4.58)4.57 (4.20)–1.08 (3.50)0.67 (0.35 – 0.85)0.72 (0.32 – 0.89)PDR8.98 (6.37)9.37 (6.73)0.40 (4.71)0.75 (0.57 – 0.86)0.78 (0.18 – 0.92)Overall5.69 (5.56)5.34 (5.96)–0.35 (3.53)0.81 (0.73 – 0.87)0.82 (0.43 – 0.93)12 × 12 mmMild-to-moderate NPDR7.75 (8.48)9.00 (7.50)1.25 (2.82)0.93 (0.73 – 0.99)0.79 (0.26 – 0.95)Severe NPDR15.55 (10.90)15.10 (11.33)–0.45 (5.72)0.88 (0.61 – 0.97)0.82 (0.46 – 0.95)PDR26.42 (15.28)26.14 (13.62)–0.28 (9.08)0.81 (0.66 – 0.90)0.79 (0.64 – 0.89)Overall21.70 (15.33)21.60 (14.06)–0.10 (7.85)0.86 (0.77 – 0.92)0.84 (0.74 – 0.90)29 × 24 mmMild-to-moderate NPDR6.91 (7.47)5.65 (7.40)–1.26 (3.58)0.87 (0.76 – 0.94)0.80 (0.62 – 0.90)Severe NPDR25.89 (12.38)21.42 (12.16)–4.47 (7.41)0.77 (0.43 – 0.91)0.73 (0.44 – 0.89)PDR28.40 (15.23)23.40 (13.70)–4.99 (11.04)0.67 (0.43 – 0.82)0.69 (0.45 – 0.83)Overall20.04 (15.80)16.52 (14.05)–3.52 (8.37)0.82 (0.71 – 0.89)0.82 (0.74 – 0.88)CI = confidence interval; DR = diabetic retinopathy; FOV = field of view; ICC = intraclass correlation coefficient; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy; SD = standard deviation.Mean difference is computed by subtracting ground-truth from prediction values. Agreement between manual grading and algorithm predictions was assessed using Bland–Altman mean differences and ICC. Intraclass correlation coefficient between manual expert graders is also reported. IRMA Counts and ICC Across Different DR Severities and FOV Sizes CI = confidence interval; DR = diabetic retinopathy; FOV = field of view; ICC = intraclass correlation coefficient; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy; SD = standard deviation. Mean difference is computed by subtracting ground-truth from prediction values. Agreement between manual grading and algorithm predictions was assessed using Bland–Altman mean differences and ICC. Intraclass correlation coefficient between manual expert graders is also reported. Retinal images were preprocessed with contrast-limited adaptive histogram equalization and Gaussian blurring to enhance image quality. Extensive data augmentation strategies were implemented, including geometric transformations, illumination adjustments, and image quality simulation to improve model robustness and generalizability. Deep Learning Model for IRMA Segmentation (CE-U-Net) Three coregistered en face images were used as inputs (Fig 2): A ground-truth–labeled en face ARL slab used as the primary input for training IRMA segmentation, an en face VRI slab used to exclude the area of NVs in proliferative DR (PDR) cases and an algorithm-generated nonperfusion area (NPA) mask. The NPA mask is generated by our self-developed and previously validated NPA segmentation algorithm13 and was incorporated as a spatial prior to bias the model toward regions along the perfused–nonperfused boundary, where IRMAs frequently arise. This prior encoded a soft anatomical preference rather than a hard constraint, because IRMAs are not confined exclusively to these boundaries. All inputs were preprocessed using a consistent pipeline (e.g. resizing to a fixed resolution and intensity normalization) prior to model training and inference. The encoder adopts a pretrained convolutional backbone to extract multiscale feature representations. At each skip connection, encoder features were fused with the corresponding NPA map through a context fusion block. Briefly, the NPA map was resized to match the feature resolution and projected to the same channel dimension, after which a learnable attention gate modulated the contribution of ischemic-context features to the skip connection. Fused skip features were further recalibrated using channel attention to enhance informative vascular channels and suppress background noise. The decoder then progressively upsampled the fused representations to produce the final IRMA segmentation map. To improve localization of small vascular structures, spatial–channel attention was applied within decoder blocks. To improve robustness across scan sizes and image quality variations, we applied data augmentation including random rotations and flips, scaling and translation, intensity/contrast perturbations, and simulated image-quality degradation (e.g., blur/noise and local signal drop-off). The model was trained end-to-end on manually annotated reference masks, with performance monitored on a held-out validation set.Figure 2Overview of the proposed CE-U-Net. ARL = all-retinal-layer; CE-U-Net = context-enhanced U-Net; IRMA = intraretinal microvascular abnormality; NPA = nonperfusion area; NV = neovascularization; VRI = vitreoretinal interface. Overview of the proposed CE-U-Net. ARL = all-retinal-layer; CE-U-Net = context-enhanced U-Net; IRMA = intraretinal microvascular abnormality; NPA = nonperfusion area; NV = neovascularization; VRI = vitreoretinal interface. Model Evaluation Intraretinal microvascular abnormality segmentation performance was evaluated at each IRMA lesion level using the F1 score. For each algorithm-generated IRMA segmentation, its geometric center was computed. A prediction was counted as a true positive (TP) if the center point lay within a manually annotated IRMA region; predictions not falling within any annotated region were counted as false positives (FP); and manually annotated IRMA regions without a corresponding prediction were counted as false negatives (FN). Duplicate detections within the same annotated IRMA region were counted only once. The F1 score was computed as:F1=2×TP2×TP+FP+FN Algorithm performance was evaluated using stratified fivefold cross-validation. In each iteration, the model was trained from scratch on 4 folds and validated on the fifth. Results are reported as the mean ± standard deviation across the 5 validation folds. The reported standard deviation reflects the variability of performance across the 5 cross-validation folds, rather than variability at the individual image level. The quantitative performance of the model was analyzed using ICC of IRMA counts between ground-truth and algorithm prediction. Comparisons among each FOV were calculated using Kruskal–Wallis test. To assess algorithm-predicted treatment-related change of IRMA burden, baseline and follow-up scans (6 weeks post-treatment) were registered to ensure alignment of vessels between 2 visits using in-house software. To ensure accurate nonperfusion index (NPI) values in our analysis comparing pretreatment and post-treatment, manual adjustment was made to the NPA segmentation done by our previously validated NPA segmentation algorithm.13 All manually corrected segmentations were independently reviewed by 2 trained graders, with disagreements resolved through adjudication by a senior retinal specialist following the grading protocol used in our DRIVE study.14 Nonperfusion index was subsequently derived by calculating the ratio of NPA to the total gradable retinal area. Statistical Analysis All analyses were performed in R (version 4.5.0, R Foundation for Statistical Computing). Continuous variables were presented as mean ± standard deviation and categorical variables as counts (%). To evaluate associations between baseline demographics and DR biomarkers (NPI and IRMA counts), multivariable linear regression was performed. To account for inclusion of bilateral eyes, comparisons across DR severity and FOVs (6 × 6 mm, 12 × 12 mm, and 29 × 24 mm) were performed using linear mixed-effects models with a patient-level random intercept. Post hoc pairwise comparisons were adjusted for multiple testing using Bonferroni correction. Intergrader reliability and agreement were assessed using ICC and Bland–Altman analysis. Intraclass correlation coefficient (2,1) (2-way random-effects, absolute agreement, single measurement) was calculated for (1) grader-to-grader agreement and (2) agreement between algorithm-predicted and reference-standard (ground-truth) IRMA counts. Intraclass correlation coefficients were reported with 95% CIs, overall and stratified by FOV and DR stage. Bland–Altman plots were used to evaluate systematic bias and variability, with mean difference calculated for each FOV (Supplementary Figure S1, available at www.ophthalmologyscience.org). A 2-sided P < 0.05 was considered statistically significant. Results Patient Characteristics A total of 254 OCTA images (98 6 × 6 mm, 58 12 × 12 mm, and 98 29 × 24 mm scans) from 254 eyes (181 participants) were analyzed. Among these, 60.2% had NPDR and 39.8% had PDR. Detailed demographics including age, gender, status of diabetic macular edema, and other comorbidities are given in Table 2. Segmentation Performance across Different FOVs and DR Severity The deep learning algorithm achieved an overall F1 score of 0.61 ± 0.08 for 6 × 6 mm scans, 0.71 ± 0.02 for 12 × 12 mm scans, and 0.65 ± 0.05 for 29 × 24 mm scans (Table 3). Across FOVs, there are significant pairwise differences in F1 scores between 6 × 6 mm scan versus 12 × 12 mm scans (P < 0.001) and versus 29 × 24 mm scans (P < 0.001), while there is no significant difference between 12 × 12 mm and 29 × 24 mm scans. Across DR severity, F1 scores remained relatively stable within each FOV, with no statistically significant differences observed across severity stages for 6 × 6 mm (P = 0.42), 12 × 12 mm (P = 0.56), and 29 × 24 mm scans (P = 0.07) (Table 3). Post hoc pairwise comparison demonstrated no significant differences between DR severity stages within all FOVs. Figure 3 illustrates example cases of different DR severities across all FOVs showing high agreement between manual ground-truth and algorithm prediction.Figure 3Intraretinal microvascular abnormality segmentation of both ground-truth and algorithm prediction on en face OCTA ARL images. (A) to (F) shows 6 × 6 mm scans, (G) to (L) shows 12 × 12 mm scans and (M) to (R) shows 29 × 24 mm scans. Ground-truth IRMA segmentation is denoted as purple annotations on each of the images, while algorithm-predicted IRMA segmentation is denoted as green annotations. Green boxes on right side of OCTA ARL images denote a magnified view of the IRMA and the corresponding B-scan, showing that there is no breach of ILM. Red annotations denote an overlay of the NV lesion from VRI scans. Red boxes on left side of OCTA ARL image denote a magnified view of NV and its corresponding B-scan, showing a clear flow signal above the ILM. High agreement between ground-truth and algorithm prediction is shown in various DR severities across all FOVs. ARL = all-retinal-layer; DR = diabetic retinopathy; FOV = field of view; ILM = internal limiting membrane; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; NV = neovascularization; OCTA = OCT angiography; PDR = proliferative diabetic retinopathy; VRI = vitreoretinal interface. Intraretinal microvascular abnormality segmentation of both ground-truth and algorithm prediction on en face OCTA ARL images. (A) to (F) shows 6 × 6 mm scans, (G) to (L) shows 12 × 12 mm scans and (M) to (R) shows 29 × 24 mm scans. Ground-truth IRMA segmentation is denoted as purple annotations on each of the images, while algorithm-predicted IRMA segmentation is denoted as green annotations. Green boxes on right side of OCTA ARL images denote a magnified view of the IRMA and the corresponding B-scan, showing that there is no breach of ILM. Red annotations denote an overlay of the NV lesion from VRI scans. Red boxes on left side of OCTA ARL image denote a magnified view of NV and its corresponding B-scan, showing a clear flow signal above the ILM. High agreement between ground-truth and algorithm prediction is shown in various DR severities across all FOVs. ARL = all-retinal-layer; DR = diabetic retinopathy; FOV = field of view; ILM = internal limiting membrane; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; NV = neovascularization; OCTA = OCT angiography; PDR = proliferative diabetic retinopathy; VRI = vitreoretinal interface. In an illustrative PDR case (Fig 4) treated with panretinal photocoagulation and anti-VEGF injection, both algorithm-predicted and ground-truth–labeled IRMA counts showed similar reductions after treatment across all FOVs. Ground-truth IRMA counts decreased from 6 to 3 (50.0%) in 6 × 6 mm scans, from 51 to 29 (43.1%) in 12 × 12 mm scans, and from 75 to 38 (49.3%) in 29 × 24 mm scans. Similarly, algorithm-predicted IRMA counts decreased from 12 to 5 (58.3%), from 38 to 20 (47.4%), and from 57 to 32 (43.9%) across the respective FOVs. In contrast, NPI showed only minimal change after treatment, decreasing from 0.032 to 0.029 (9.9%) in 6 × 6 mm scans, from 0.071 to 0.064 (8.8%) in 12 × 12 mm scans, and from 0.087 to 0.084 (4.3%) in 29 × 24 mm scans.Figure 4Change in absolute IRMA counts illustrated in pretreatment and 6 weeks post-treatment across all FOVs. Orange annotations (A, B, E, F, I, J) denote manual human grading of IRMAs, while light green annotations (C, D, G, H, K, L) denote algorithm prediction of IRMAs. Dark green annotations (C, D, G, H, K, L) denote algorithm prediction with manual adjustment of NPA. FOV = field of view; IRMA = intraretinal microvascular abnormality; NPA = nonperfusion area. Change in absolute IRMA counts illustrated in pretreatment and 6 weeks post-treatment across all FOVs. Orange annotations (A, B, E, F, I, J) denote manual human grading of IRMAs, while light green annotations (C, D, G, H, K, L) denote algorithm prediction of IRMAs. Dark green annotations (C, D, G, H, K, L) denote algorithm prediction with manual adjustment of NPA. FOV = field of view; IRMA = intraretinal microvascular abnormality; NPA = nonperfusion area. Quantitative Performance across Different FOVs Quantification of algorithm-predicted and ground-truth–labeled IRMA counts demonstrated strong agreement in various DR severities across all FOVs (Table 4). Bland–Altman analysis showed overall agreement between algorithm-predicted and ground-truth IRMA counts. Consistent performance was achieved across all FOVs, with an ICC of 0.81 (95% CI 0.73–0.87) for 6 × 6 mm, 0.86 (95% CI 0.77–0.92) for 12 × 12 mm, and 0.82 (95% CI 0.71–0.89) for 29 × 24 mm indicating good-to-excellent reliability. Overall, the algorithm showed a slight tendency toward underestimation, with mean differences (predicted – ground truth) of –0.35, –0.10, and –3.52 IRMAs for the 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm scans, respectively. These values were closely comparable to intergrader agreement between the 2 human expert graders (6 × 6 mm: ICC = 0.82, 95% CI 0.43–0.93; 12 × 12mm: ICC = 0.84, 95% CI 0.74–0.90; 29 × 24 mm: ICC = 0.82, 95% CI 0.74–0.88). Figure 5 illustrates the high agreement in not only IRMA quantification but also spatial distribution for both manual grading and algorithm prediction across DR severities for all FOVs.Figure 5Comparison of manual grading and algorithm-predicted IRMA counts across DR severity and FOVs. For each severity category, manual human grading (purple) is shown alongside algorithm prediction (green) across 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm OCTA scans. DR = diabetic retinopathy; FOV = field of view; GT = ground truth; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; OCTA = OCT angiography; PDR = proliferative diabetic retinopathy. Comparison of manual grading and algorithm-predicted IRMA counts across DR severity and FOVs. For each severity category, manual human grading (purple) is shown alongside algorithm prediction (green) across 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm OCTA scans. DR = diabetic retinopathy; FOV = field of view; GT = ground truth; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; OCTA = OCT angiography; PDR = proliferative diabetic retinopathy. IRMA Burden across DR Severity in Different FOVs Figure 6 illustrates the increase in IRMA counts from mild-to-moderate NPDR to severe NPDR across all FOVs for both manual grading and algorithm prediction, with only 29 × 24 mm scans demonstrating statistical significance (ground-truth: P < 0.001; prediction: P < 0.001. From severe NPDR to PDR, IRMA counts continued to rise significantly in the 6 × 6 mm scans (ground-truth: P = 0.04; prediction: P = 0.01) while both the increase in 12 × 12 mm (ground-truth: P = 0.15; prediction: P = 0.08) and the plateau in 29 × 24 mm scans (ground-truth: P = 0.72; prediction: P = 0.68) did not amount to statistical significance. Algorithm-predicted values slightly underestimate IRMA counts but still closely parallel ground-truth IRMA count trends.Figure 6Median IRMA counts with interquartile range (IQR) are shown for mild-to-moderate NPDR, severe NPDR, and PDR across 6 × 6 mm (yellow), 12 × 12 mm (gray), and 29 × 24 mm (pink) OCTA scans for both manual grading (dotted) and algorithm prediction (solid). DR = diabetic retinopathy; FOV = field of view; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; OCTA = OCT angiography; PDR = proliferative diabetic retinopathy. Median IRMA counts with interquartile range (IQR) are shown for mild-to-moderate NPDR, severe NPDR, and PDR across 6 × 6 mm (yellow), 12 × 12 mm (gray), and 29 × 24 mm (pink) OCTA scans for both manual grading (dotted) and algorithm prediction (solid). DR = diabetic retinopathy; FOV = field of view; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; OCTA = OCT angiography; PDR = proliferative diabetic retinopathy. Mixed effect model analysis (Supplementary Table S1, available at www.ophthalmologyscience.org) demonstrated that only age showed significant negative association with ground-truth IRMA counts. Other demographic variables and comorbidities such as gender, hypertension, hyperlipidemia, ischemic heart disease, cerebrovascular accident, chronic kidney disease, DM, and glycosylated hemoglobin were not significantly associated with ground-truth IRMA counts. Correlation between IRMA Burden and NPI across DR Severity in Different FOVs A linear regression analysis conducted to assess the relationship between the number of IRMAs and NPI showed that algorithm-predicted IRMA counts demonstrated a statistically significant positive correlation with algorithm-generated NPI across all FOVs (6 × 6 mm: P < 0.001, 12 × 12 mm: P < 0.001, 29 × 24 mm: P < 0.001). Stage-specific analyses revealed that in the 12 × 12 mm scans, significant positive correlations between IRMA count and NPI were observed across all DR severities, with the strongest association seen in PDR. In the 29 × 24 mm scans, a significant association was detected only in severe NPDR, while correlations in other stages did not reach statistical significance. In contrast, the 6 × 6 mm scans demonstrated overall weaker associations, with no statistically significant stage-specific correlations observed (Table 5). Figure 7 demonstrates the spatial distribution of IRMA counts and the NPA across the various DR severities.Table 5Stage-Specific Linear Regression Analysis of the Association between IRMA Counts and NPI across Different OCTA Fields of ViewFOVDR StageNPI β (95% CI)P Value6 × 6 mmMild-to-moderate NPDR0 (–0.002 to 0.007)0.23Severe NPDR0.02 (–0.016 to 0.050)0.23PDR0.01 (–0.038 to 0.060)0.6512 × 12 mmMild-to-moderate NPDR0.03 (0.013 – 0.055)0.01Severe NPDR0.03 (0.034 – 0.035)<0.001PDR0.05 (0.027 – 0.070)<0.00129 × 24 mmMild-to-moderate NPDR0.01 (–0.010 to 0.037)0.24Severe NPDR0.06 (0.004 – 0.106)0.04PDR0.02 (–0.031 to 0.077)0.34CI = confidence interval; DR = diabetic retinopathy; FOV = field of view; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; NPI = nonperfusion index; OCTA = OCT angiography; PDR = proliferative diabetic retinopathy.Regression coefficients (β) with 95% CIs and corresponding P values are presented. Positive β values indicate an increase in IRMA count with increasing NPI.Figure 7Spatial distribution of IRMA count and NPA across all DR severities and FOVs. For each severity category, IRMA counts (orange) are overlapped with NPA (gray) across 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm images. Increasing IRMA count is associated with increasing NPA across DR severities. DR = diabetic retinopathy; FOV = field of view; IRMA = intraretinal microvascular abnormality; NPA = nonperfusion area; NPDR = nonproliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy. Table 5Stage-Specific Linear Regression Analysis of the Association between IRMA Counts and NPI across Different OCTA Fields of ViewFOVDR StageNPI β (95% CI)P Value6 × 6 mmMild-to-moderate NPDR0 (–0.002 to 0.007)0.23Severe NPDR0.02 (–0.016 to 0.050)0.23PDR0.01 (–0.038 to 0.060)0.6512 × 12 mmMild-to-moderate NPDR0.03 (0.013 – 0.055)0.01Severe NPDR0.03 (0.034 – 0.035)<0.001PDR0.05 (0.027 – 0.070)<0.00129 × 24 mmMild-to-moderate NPDR0.01 (–0.010 to 0.037)0.24Severe NPDR0.06 (0.004 – 0.106)0.04PDR0.02 (–0.031 to 0.077)0.34CI = confidence interval; DR = diabetic retinopathy; FOV = field of view; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; NPI = nonperfusion index; OCTA = OCT angiography; PDR = proliferative diabetic retinopathy.Regression coefficients (β) with 95% CIs and corresponding P values are presented. Positive β values indicate an increase in IRMA count with increasing NPI. Stage-Specific Linear Regression Analysis of the Association between IRMA Counts and NPI across Different OCTA Fields of View CI = confidence interval; DR = diabetic retinopathy; FOV = field of view; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; NPI = nonperfusion index; OCTA = OCT angiography; PDR = proliferative diabetic retinopathy. Regression coefficients (β) with 95% CIs and corresponding P values are presented. Positive β values indicate an increase in IRMA count with increasing NPI. Spatial distribution of IRMA count and NPA across all DR severities and FOVs. For each severity category, IRMA counts (orange) are overlapped with NPA (gray) across 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm images. Increasing IRMA count is associated with increasing NPA across DR severities. DR = diabetic retinopathy; FOV = field of view; IRMA = intraretinal microvascular abnormality; NPA = nonperfusion area; NPDR = nonproliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy. Discussion In this study, we developed and validated a deep learning–based framework for automated segmentation and quantification of IRMAs on SS-OCTA across multiple FOVs. To our knowledge, this is the first study to demonstrate robust IRMA detection and quantification from 6 × 6 mm to UWF 29 × 24 mm scans, achieving moderately strong performance for this morphologically challenging lesion across different FOVs. Beyond technical validation, we demonstrate the clinical relevance of automated IRMA quantification by showing that the algorithm-derived IRMA burden closely mirrors manual grading—where IRMA counts increase progressively with worsening DR severity. These findings support automated IRMA quantification as a potential OCTA metric for assessing disease severity and activity in DR. Technical Performance and Validation of Automated Deep Learning–Based IRMA Segmentation We selected the F1 score as our primary evaluation metric because our objective was accurate detection and localization of IRMAs rather than pixel-perfect segmentation. Clinically, IRMA number and distribution are more informative for DR risk stratification and disease monitoring14 than pixel-level boundary agreement, consistent with how IRMA burden is used in DR staging and monitoring.15,16 The progressive rise in IRMA count with advancing DR severity is well established7,17 and is supported by our own prior work.14 Pixel-overlap metrics such as the DSC are less suited to this task: given the pleomorphic morphology, gradual tapering, and indistinct margins of IRMAs,18,19 minor contour disagreements between expert graders can substantially depress DSC even when lesion presence and location are concordant. This limitation is evident in the Diabetic Retinopathy Analysis Challenge benchmark,10 where IRMAs recorded the lowest segmentation performance of any lesion type (top-10 models: DSC 29.53%–47.04%, versus 46.59%–69.26% for NPAs and 46.73%–65.71% for NVs)—reflecting the intrinsic difficulty of pixel-level agreement for this lesion rather than a failure of detection. The F1 score therefore provides a more clinically meaningful measure of performance for lesion detection and quantification. To ensure a reliable reference standard, we developed a standardized IRMA grading framework based on previously published morphological classifications,7,12 with manual annotation performed by trained graders using predefined criteria. The resulting high intergrader agreement underpinned the consistency of the labels used for training and validation, and our model achieved F1 scores of 0.61, 0.71, and 0.65 for the 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm scans, respectively. A key novel component of our approach is the incorporation of our previously validated NPA segmentation algorithm13 to guide IRMA detection. Rather than treating IRMAs as isolated abnormalities, the model explicitly integrates ischemic context by relating them to adjacent areas of capillary nonperfusion—reflecting the established observation that IRMAs arise at the junction between perfused and nonperfused retina, functioning as intraretinal shunt vessels in response to ischemia.15,20 Operationally, the algorithm first generates an automated NPA mask that serves as structural prior, localizing regions at risk of IRMA formation along the ischemic border. Anchoring detection to this biologically meaningful interface improves discrimination of true IRMAs from confounders such as vessel tortuosity or microaneurysms. Consistent with this detection-oriented design, algorithm-derived IRMA quantification showed good-to-excellent agreement with manual grading across all 3 FOVs, with overall ICC values ranging from 0.81 to 0.86, minimal mean differences between predicted and ground-truth counts, and strong spatial concordance. Performance was consistent and comparable across the 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm scans, with no significant difference between scan protocols; the minor numerical variation in ICC observed across disease stages was small and should be interpreted with caution as a difference in performance. Together, these findings support the reliability of the model for clinically interpretable and reproducible assessment of IRMA burden and location across different OCTA scan protocols. Furthermore, our algorithm was developed using meticulously segmented OCTA scans in which NVs were identified on the VRI slab and excluded prior to model training. This VRI-guided approach represents a novel key strength of our algorithm training, as NVs can be reliably distinguished from IRMAs by the presence of an ILM breach.9 In contrast, when viewed on the ARL slab alone, NV and IRMAs may coexist and exhibit similar vascular morphology, particularly if NV is erroneously segmented onto the ARL slab,21,22 increasing the risk of misclassification. To evaluate the performance of our algorithm in routine clinical practice, where the VRI slab may not be routinely reviewed alongside the ARL slab, we performed an additional analysis without prior exclusion of NV lesions. Under these conditions, false-positive rates for misclassifying NV as IRMA were 15.9%, 8.2%, and 6.8% for the 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm scan protocols respectively. These rates represent the error that the VRI-based exclusion step is specifically designed to prevent and thus underscore the value of incorporating VRI slab information rather than relying on the ARL slab alone. In clinical practice, this source of error can be readily mitigated by reviewing the VRI slab for evidence of ILM breach,22 allowing suspected NVs to be correctly identified and excluded from IRMA quantification. Additionally, our illustrative pretreatment and post-treatment PDR example serves as a proof of concept that the IRMA algorithm can be used to track treatment response through longitudinal changes in IRMA burden. In this example, algorithm-predicted and manually graded IRMA counts showed similar relative reductions after treatment across all FOVs, whereas NPI changed minimally, consistent with previous reports.23,24 Traditionally, IRMAs have served as surrogate markers of retinal ischemia when nonperfusion cannot be directly visualized on CFP or clinical examination.17 However, OCTA frequently reveals substantially more IRMAs, often adjacent to areas of nonperfusion, which are not detectable on conventional imaging. These OCTA-detectable or "occult" IRMAs may represent earlier or more subtle manifestations of retinal microvascular remodeling, providing additional information beyond that available from CFP alone. By objectively and reproducibly quantifying both visible and “occult” IRMAs, our deep learning algorithm may enable sensitive, low-variability assessment of treatment-induced microvascular change that CFP-based tracking cannot achieve (Supplementary Figure S2, available at www.ophthalmologyscience.org). Rather than serving solely as a marker of DR severity, longitudinal changes in IRMA burden and morphology may represent a quantitative biomarker of treatment response. This potential application warrants validation in larger prospective longitudinal cohorts. Clinical Relevance of IRMA Quantification across Different FOVs in all Stages of DR The IRMA segmentation algorithm enabled more efficient and reproducible analysis of IRMA burden across all stages of DR. Intraretinal microvascular abnormality counts increased progressively with DR severity in all scan protocols, consistent with the known association between retinal ischemia and intraretinal microvascular remodeling,3 although the magnitude and pattern of change varied with FOVs. The UWF 29 × 24 mm scans showed the largest increase between mild-to-moderate and severe NPDR, suggesting that IRMA formation may occur preferentially in the mid-peripheral retina as disease advances—a region captured more effectively by wider imaging fields.25 In contrast, the smaller 6 × 6 mm and 12 × 12 mm scans, which primarily sample the posterior pole, showed a more gradual increase across stages, consistent with field-dependent sampling effects in OCTA biomarker quantification.25 The relative plateau in the 29 × 24 mm scans between severe NPDR and PDR may reflect saturation of detectable peripheral lesions once widespread ischemia is present,26 leaving less viable retina for new IRMA formation.14 Notably, algorithm-predicted counts closely paralleled manually graded trends across all FOVs despite slight underestimation, indicating that the model preserved the biological gradient of IRMA burden across DR severity. The significant negative association between age and ground-truth IRMA counts may further reflect a shift toward more established ischemia and less active vascular remodeling with longer diabetes duration. Further supporting the ischemia-driven nature of IRMA formation, IRMA counts demonstrated a positive correlation with NPI across scan protocols.27 The 12 × 12 mm scans showed the most consistent stage-specific associations in our study, suggesting that this FOV may provide a stable representation of the spatial coupling between ischemic retina and intraretinal microvascular remodeling. By capturing both the posterior pole and adjacent mid-peripheral retina while maintaining adequate spatial resolution, this scan size likely enables concurrent visualization of NPAs and IRMAs, which frequently develop along the borders of ischemic zones.28 In contrast, correlations were weaker in the 6 × 6 mm scans, likely reflecting the limited sampling area and relative under-representation of peripheral ischemia within the macula-centered field.29 Although the 29 × 24 mm scans provide broader peripheral coverage, the reduced apparent size of IRMAs and lower effective spatial resolution at this scale may introduce variability in both NPA segmentation and IRMA detection, thereby weakening the stage-specific coupling between these biomarkers despite the larger retinal area sampled. In addition, in advanced disease such as PDR, progressive expansion of ischemic retina may leave less viable tissue for further IRMA development, which could further attenuate the correlation between IRMA burden and NPA extent in this stage.26 Limitations Our study has several limitations. First, direct comparison of performance between our study and previously published studies should be interpreted with caution. Although our model demonstrated improved IRMA segmentation performance, our model was evaluated using a F1 score, which emphasizes accurate lesion detection, localization and quantification, whereas other existing studies report DSC, a metric that primarily evaluates pixel-wise overlap. As these metrics evaluate different aspects of segmentation performance, direct comparisons across studies may not be directly equivalent. Second, the study was conducted with a relatively small sample size, which may limit the generalizability of the findings. Future work could expand our cohort size by adopting a multicenter study design or having a longer time frame to recruit more potential patients for our study. Third, although our detection-centric evaluation framework better reflects clinical priorities compared to strict pixel-level segmentation metrics, it does not fully capture morphological subtypes or vascular remodeling patterns of IRMAs. Future work integrating morphological phenotyping and temporal trajectory analysis may further enhance the prognostic utility of the model. Furthermore, longitudinal validation was limited to a single-treated case as a proof of concept. Although the observed reduction in IRMA burden after anti-VEGF and panretinal photocoagulation treatment supports biological plausibility, larger longitudinal studies with standardized image registration and repeated follow-up intervals are required to confirm sensitivity to temporal change and to establish minimal detectable change thresholds for clinical application. Another limitation pertains to the assessment of DM duration. As this information was collected retrospectively from endocrinology clinical records, DM duration was frequently documented as categorical variables (e.g. 5-10 years) rather than exact durations. Consequently, DM duration was analyzed as a categorical variable as modeling it as a continuous variable would require assumptions regarding the actual duration and potentially introduce measurement error. Future studies with prospectively collected data would allow more robust evaluation of its relationship with IRMA burden. Conclusions In this study, we developed and validated a novel deep learning algorithm (CE-U-Net) for the automated segmentation of IRMAs on SS-OCTA, including UWF SS-OCTA. Our model achieved good performance for this morphologically challenging task and showed strong agreement with manually annotated ground-truth. These findings support the potential clinical utility of automated IRMA quantification in DR severity stratification and treatment response tracking. Future work will focus on broader external and longitudinal validation of the algorithm, as well as extension toward IRMA morphological subtyping, with the aim of improving prognostic stratification and supporting quantitative, biomarker-driven management of DR.