Deep Learning Model to Detect Diabetic Retinopathy in 45° Images Using Ground Truth from Ultra-Widefield Imaging

Abstract

Purpose

To evaluate the performance of a deep learning (DL) model in classifying diabetic retinopathy (DR) severity using fundus images with varying fields of view and to assess whether central retinal features alone reflect overall disease burden. The study also investigates vascular biomarkers within regions contributing to model decisions to explore biological basis of inferences and enhance clinical interpretability.

Design

An observational, cross-sectional study.

Participants

A total of 2610 participants aged ≥40 years from a population-based study in South India, with ocular and systemic data, including dilated fundus images and glycated hemoglobin levels.

Methods

Diabetic retinopathy severity was graded on unmasked 200° ultra-widefield (UWF) and centrally masked 45° images, with peripheral lesions annotated in the 155° field (200°–45°). Convolutional neural network was trained on labeled images and evaluated on both datasets. Performance metrics and receiver operating characteristic (ROC) curves were calculated. Gradient-weighted class activation mapping (Grad-CAM) identified model focus. Vascular biomarkers (tortuosity, fractal dimension, and vessel density) were quantified using LWNet and Fiji software and compared between eyes with and without peripheral lesions.

Main Outcome Measures

Classification accuracy of DR severity, precision and recall, ROC curves, Grad-CAM visualization of regions of model focus, and vascular biomarkers associated with peripheral lesions.

Results

The DL model achieved high classification accuracy: 97.12% for UWF images, 97.24% for 45° masked images with re-evaluated labels, and 96.86% for masked images with original labels. Peripheral pathology, including microaneurysms (19.8%) and hemorrhages/exudates (9.1%), did not affect accuracy. Deep learning–based assessment of 45° fundus images reduced DR underestimation from 16.5% to 6.2%. Gradient-weighted class activation mapping highlighted focus on central regions, including areas without visible lesions. Vascular analysis revealed differences in vessel density and tortuosity between eyes with and without peripheral microvascular abnormalities and neovascularization, suggesting detection of subclinical vascular changes linked to peripheral disease.

Conclusions

Deep learning applied to 45° fundus images can accurately classify DR and detect subtle vascular biomarkers predictive of peripheral disease. This proof-of-concept highlights the potential of artificial intelligence (AI)-enhanced 45° imaging as a scalable tool for DR screening. Such AI-powered approaches, using accessible and affordable fundus cameras, may enable cost-effective detection and triage of high-risk cases in primary care and resource-limited settings.

Financial Disclosure(s)

The authors have no proprietary or commercial interest in any materials discussed in this article.

Diabetic retinopathy (DR) is a leading cause of vision impairment and blindness worldwide, particularly among individuals with long-standing diabetes.1, 2, 3 Early detection and timely intervention are crucial for preventing irreversible vision loss, making accurate and accessible screening methods essential.⁴ The ETDRS established a standardized grading system based on retinal lesions observed in the posterior pole, providing a structured framework for assessing DR severity.⁵ This classification system remains the clinical gold standard, facilitating research and guiding treatment decisions.⁶ However, traditional grading methods, including ETDRS seven-field imaging, primarily focus on central retinal features while overlooking potential pathology in the peripheral retina, which can lead to under-recognition of disease severity, particularly when pathology extends beyond standard imaging fields.^7,8

The advent of ultra-widefield (UWF) imaging, which captures up to 82% of the retinal surface in a single image, has significantly enhanced DR evaluation by revealing peripheral lesions not visible in standard fields.^9,10 Studies suggest that up to 40% of eyes may exhibit lesions outside ETDRS-defined regions, and in 10% of cases, these findings may alter severity grading. Moreover, the presence and extent of peripheral pathology have been linked to an increased risk of DR progression.11, 12, 13 Despite these advantages, the high cost and specialized infrastructure required for UWF imaging limit its accessibility, particularly in resource-constrained settings, where standard 45° fundus imaging remains the primary modality for DR screening.¹⁴ As a result, clinically meaningful underestimation of DR severity may persist in routine practice when relying solely on conventional imaging and human grading.

Recent advancements in artificial intelligence (AI), particularly convolutional neural networks, have enabled automated and accurate DR detection using fundus photography.^15,16 While many AI models have shown expert-level performance, most have been trained on datasets with broad retinal coverage, raising questions about their generalizability when restricted to narrower fields of view (FOVs).¹⁷ Given the global reliance on standard 45° imaging for screening, it is crucial to understand whether deep learning (DL) models can retain high performance under such constraints and whether they can leverage central retinal features to infer broader disease involvement.

Moreover, there is emerging evidence that DR induces diffuse retinal vascular changes, not restricted to visible lesions.¹⁸ These alterations in vessel density, tortuosity, and fractal dimension may serve as early biomarkers of disease progression and severity.¹⁹ If such features can be leveraged by DL models to infer the presence of peripheral lesions, including intraretinal microvascular abnormalities (IRMAs) and neovascularization elsewhere (NVE), this would significantly enhance the utility of conventional fundus cameras for both early and advanced DR screening.

In this study, the performance of a DL model for DR severity classification was evaluated using retinal images with varying FOVs. Model performance was systematically compared across full 200° UWF images, centrally masked 45° images with relabeled ground truths and masked images retaining original labels. Importantly, the analysis examined whether AI assistance could reduce the clinically relevant underestimation of DR severity observed with human grading alone. Model interpretability was assessed using gradient-weighted class activation mapping (Grad-CAM) visualizations, and vascular biomarkers within regions contributing to model decisions were quantitatively analyzed. The overall aim was to determine whether AI models can not only detect visible lesions but also infer peripheral disease from central retinal features, thereby expanding the clinical utility of 45° fundus imaging for scalable and interpretable DR screening.

Methods

Study Participants and Data Collection

This study involved a cohort of 2610 participants, all aged 18 years or older, who were recruited from population-based cross-sectional studies conducted in South India.^20,21 These cohorts encompassed individuals from both urban and rural communities, as well as tertiary eye-care hospital settings, capturing a broad spectrum of demographic and clinical characteristics representative of the regional population with diabetes. The detailed methodology has been described elsewhere.^20,21 In brief, written consent was obtained from each participant prior to sample collection, followed by a comprehensive ophthalmic evaluation and collection of demographic and systemic health information. Data collection followed a structured approach, beginning with the administration of a standardized questionnaire by a trained optometrist. This questionnaire captured key demographic details such as age, sex, and ethnicity, along with lifestyle factors including current status and past history of smoking. Additionally, participants provided a comprehensive medical history including details of systemic conditions such as diabetes, hypertension, chronic heart and lung disease, etc. Ocular history collected from the participants included previous eye diseases, surgeries, and current systemic and ocular medication use.

Following the collection of medical history, each participant underwent a complete ophthalmic examination. Visual function was assessed through visual acuity testing using a Snellen chart, followed by both objective and subjective refraction to determine the best-corrected visual acuity. Slitlamp biomicroscopy was performed to examine the anterior segment of the eye, including the cornea, conjunctiva, anterior chamber, iris, and crystalline lens. Intraocular pressure was measured using the Goldmann applanation tonometer to assess risk factors for glaucoma or ocular hypertension. Additionally, lens opacity grading was conducted based on the Lens Opacities Classification System III,²² after dilation to assess the severity of cataractous changes.

For the evaluation of the posterior segment, all participants underwent dilated fundus imaging using a UWF imaging system. Retinal images were captured using the Daytona Plus Panoramic Ophthalmoscope – P200T (Optos), which provides a 200° FOV, with a resolution format of 4000 × 4000 pixels, allowing for a comprehensive screening of various retinal pathologies, including DR. The imaging procedure was performed approximately 20 to 30 min after pharmacologic pupillary dilation using tropicamide 1% eye drops, ensuring optimal image quality and widefield visualization.

In addition to ocular health examination, systemic health parameters were recorded, including systolic and diastolic blood pressures, glycated hemoglobin levels, which were measured using the Cobas c 111 (Roche). The inclusion of glycated hemoglobin levels enabled an objective assessment of participants' glycemic control, a crucial factor in DR progression. Serum lipid profile results estimated with a Dade Behring RXL (Siemens), and spot urine albumin and creatinine estimated using Dade Behring RXL were recorded.^20,21

Ethical approval for the research was obtained from the Institutional Review Board of the Vision Research Foundation at Sankara Nethralaya, Chennai, as well as the Human Research Ethics Advisory Panel at the University of New South Wales, Sydney. All study procedures adhered to the ethical principles outlined in the Declaration of Helsinki.

Image Grading and Data Preparation

A total of 5180 fundus images were graded for DR severity by an optometrist (R.K.), following the International Clinical Diabetic Retinopathy Disease Severity Scale.²³ Diabetic retinopathy was classified into 5 stages: no DR, mild nonproliferative diabetic retinopathy (NPDR), moderate NPDR, severe NPDR, or proliferative diabetic retinopathy (PDR). The grading was performed on both 200° and 45° images. To prevent bias, all images were randomized so that the grader was unaware of whether any 45° image corresponded to the central portion of a previously graded 200° UWF image. The grader was thus masked to the pairing of images, ensuring that each 45° image was assessed independently without knowledge of its full-field counterpart or previously assigned grade.

The lesions were identified to examine what might be overlooked in the peripheral areas of the masked 45° images when only the remaining 155° were used. In the 155 peripheral region, lesion types such as microaneurysms, hemorrhages and exudates, IRMA, and NVE were individually annotated to enable a detailed assessment of their distribution beyond the central field. To ensure grading reliability, a subset of 10% of the images was regraded by both the same optometrist and a second optometrist. Intragrader and intergrader agreements were evaluated using the Cohen kappa statistic. Any discrepancies between the 2 graders were resolved through joint re-review of the images and discussion.

To assess the correspondence between central and peripheral imaging, DR grades assigned on 45° images were compared with the corresponding full 200° UWF images. This analysis quantified the extent to which grading on central crops alone might underestimate disease severity. The final DR grade for each participant was assigned based on the eye with the most advanced stage of DR. To maintain data integrity and analytical precision, fundus images that were significantly blurred, or exhibited artifacts due to media opacities such as dense cataracts, or were obscured by vitreous hemorrhage were excluded for further image processing.

Dataset Preprocessing

The dataset used for this study consisted of 5180 retinal Optos UWF fundus images. To ensure uniformity in field-of-view across different imaging modalities and facilitate direct comparisons with conventional fundus cameras, a central circular mask was applied to each UWF image. The mask radius was calculated based on a 45:200 ratio, ensuring that 45° of the central retinal area was retained, while 155° of the peripheral retina was excluded (Fig 1). This masking process generated 4 distinct datasets using different labeling approaches.

• a.Full-field 200° UWF dataset: This dataset comprises unmasked UWF retinal images capturing up to a 200° FOV. The DR severity labels were assigned based on the full extent of the retina, incorporating both central and peripheral retinal features.
• b.Central 45° FOV (masked periphery) with re-evaluated labels: In this dataset, the peripheral retina is masked to simulate a conventional 45° FOV. A qualified expert reassessed the DR severity grades by evaluating only the visible central retinal region. This allows for an independent classification of DR severity, excluding potential influence from peripheral lesions.
• c.Central 45° FOV (masked periphery) with original labels: This dataset also contains simulated 45° FOV (peripheral regions masked) but retains the original DR severity labels assigned based on the full 200° field. It enables assessment of the classification performance when only central retinal features are available, but labels reflect the complete retinal pathology.
• d.Peripheral 155° FOV (masked central): In this dataset, the central 45° region is masked, isolating the peripheral 155° field from the original UWF image. Manual annotations of individual DR lesion types (e.g., microaneurysms, hemorrhages, hard exudates, IRMA, and NVE) were performed to support detailed spatial analysis of lesion distribution outside the central retina.

Figure 1.

Overview of the 4 retinal image datasets used for DR classification. A, Original unmasked 200° UWF image graded as PDR. B, Central 45° FOV image with masked periphery and re-evaluated DR severity based solely on central retinal features, graded as moderate NPDR. C, The same centrally masked 45° image as in (B), retaining the original full-field DR label from (A) as PDR. D, Manual annotations of individual DR lesions, with microaneurysms indicated by blue circles and NVE marked by yellow circles. DR = diabetic retinopathy; FOV = field of view; NPDR = nonproliferative diabetic retinopathy; NVE = neovascularization elsewhere; PDR = proliferative diabetic retinopathy; UWF = ultra-widefield.

Given the substantial class imbalance in the severity levels of DR, with certain stages being underrepresented in the dataset, a mixed generative augmentation method was employed.²⁴ This method created additional composite images by combining both central and peripheral regions of the retina to help balance the dataset. This augmentation approach was designed to enhance intraclass diversity while preserving the anatomical and pathological features essential for DR classification. The synthetic images were generated by combining 2 class-matched images, denoted as IM1 and IM2, using a weighted image-mixing formula. A randomly selected coefficient, K, with values ranging between 0 and 1, was applied to generate the new image NM as follows:

$$NM=IM1\times K+IM2\times \left(1–K\right)$$

This generative approach allowed for the creation of new images that retained realistic anatomical structures while incorporating variations in intensity and texture. The process was iteratively applied until each DR severity class contained 5000 images, resulting in a final dataset of 25 000 images with balanced class representation. Representative examples of the augmented images demonstrating preserved anatomical and pathological features are shown in Figure S1 (available at www.ophthalmologyscience.org).

Model Development and Training

For DR classification, a convolutional neural network based on the VGG16 architecture,²⁵ was employed due to its proven effectiveness in image classification tasks. VGG16's deep architecture, with multiple stacked layers of small 3 × 3 convolutional kernels, allows it to extract complex hierarchical features from retinal images. This makes it particularly effective for identifying subtle patterns and lesions associated with different DR severity levels, and well-suited for robust feature extraction compared to shallower or more fragmented architectures. The model was initialized with pretrained ImageNet weights to utilize the hierarchical feature representations learned from large-scale image datasets, thereby enhancing feature extraction efficiency. In addition, the model incorporated both the pretrained features and the severity grading of retinal images to improve the accuracy of DR classification. The model architecture was modified to accommodate the 5 levels of DR. The final fully connected layer was replaced with a fully connected layer with 5 neurons and a SoftMax activation function.²⁶ This function outputs probability scores for each of the five DR severity classes, allowing the model to classify the images accordingly.

For training, cross-entropy loss was employed, a standard objective function for multiclass classification tasks, which measures how well the model predicts the correct DR severity level. Two optimization algorithms, Adam and AdamW, were evaluated to assess their impact on model convergence and generalization.²⁷ These optimizers were applied continuously throughout the training process as part of the model's standard learning procedure. They were not triggered or employed conditionally based on the cross-entropy loss reaching a specific threshold. The AdamW optimizer was ultimately selected due to its incorporation of weight decay regularization, which helps prevent overfitting by penalizing large weights across all features. This regularization encourages the model to avoid relying too heavily on any individual feature, rather than specifically on neighboring or surrounding features, thereby promoting better generalization.

The dataset was split into 2 parts: 80% for training the model and 20% for testing it. This split ensured that all levels of DR were adequately represented in both parts. Internal cross-validation was performed during model development to optimize hyperparameters and enhance generalization. To further enhance performance, early stopping was implemented, meaning that training was stopped when the model's performance on the test data plateaued. This approach saved time and minimized overfitting. Finally, the model's performance was measured using accuracy, precision, recall, and F1-score, providing a comprehensive evaluation of its ability to correctly identify and classify different DR severity levels. Additionally, receiver operating characteristic (ROC) curves were generated to assess the model's discriminative power across DR classification thresholds. An overview of the entire methodological workflow is illustrated in Figure 2.

Figure 2.

Overview of the methodological workflow for DR severity classification and model evaluation. DR = diabetic retinopathy; FOV = field of view; Grad-CAM = gradient-weighted class activation mapping; IRMA = intraretinal microvascular abnormality; NPDR = nonproliferative diabetic retinopathy; NVE = neovascularization elsewhere; PDR = proliferative diabetic retinopathy.

Model Explainability and Interpretability

To enhance the interpretability of the DL model and provide insights into its decision-making process, Grad-CAM was applied to the final convolutional layer.²⁸ Gradient-weighted class activation mapping generated high-resolution heatmaps that visually highlighted the regions within the retinal images that contributed most significantly to the model's classification decisions. These heatmaps were overlaid onto the original fundus images, enabling a clear visualization of the model's attention focus. The heatmaps used a color scale in which areas with higher predictive importance appeared in warmer colors such as red and yellow, while regions of lower significance were displayed in cooler shades such as blue and green.

Vascular Segmentation and Biomarker Extraction

Following the generation of Grad-CAM attention maps and thresholding to highlight salient regions, vascular structures were subsequently segmented from the fundus images. This step was essential to facilitate the extraction of detailed vascular biomarkers. For segmentation, LWNet was employed, a lightweight DL architecture specifically designed for efficient and accurate retinal vessel segmentation.²⁹

LWNet features a compact and optimized architecture suitable for medical image analysis. It employs a dual-pathway design, integrating both local features that capture fine vessel details and global contextual information for a comprehensive understanding of the retinal vasculature.^29,30 The model uses depth-wise separable convolutions to significantly reduce computational cost, ensuring rapid inference without sacrificing performance. Attention mechanisms incorporated within the architecture enhance vessel boundary delineation by focusing on high-saliency regions, improving accuracy in distinguishing between arteries and veins. LWNet was trained and validated on widely used retinal imaging datasets such as digital retinal images for vessel extraction and structured analysis of the retina, ensuring robustness and generalizability across varying image qualities and pathologies.^31,32

The retinal vasculature output of LWNet was processed using the Fiji software (a free software available at https://fiji.sc).³³ The segmented image was skeletonized using Skeletonize (2D/3D) function.³⁴ The AnalyzeSkeleton plugin in Fiji was then applied to the skeletonized images to extract detailed branch information. Local tortuosity for each vascular segment was calculated as the ratio of the arc length to the Euclidean distance between branch points. The global tortuosity index was computed as the total arc length divided by the total Euclidean distance across all branches, while the median tortuosity was obtained as the median of all local tortuosity values. Fractal dimension was calculated using the box-counting method on the skeletonized images. Vessel density was measured using Fiji's “Measure” function and defined as the ratio of vessel pixels to the total image area. The schematic of the segmentation and processing pipeline is shown in Figure 3.

Figure 3.

Overview of segmentation and processing pipeline.

Statistical Analysis

Descriptive statistics were used to summarize the frequency and distribution of DR lesion types in the peripheral 155° field, reported as counts and percentages. Chi-square (χ²) tests were performed to evaluate the association between the presence or absence of peripheral lesions and categorical variables. For continuous retinal vascular parameters, mean and standard deviation values were calculated. Independent samples t-tests were conducted to compare these parameters between groups with and without specific peripheral lesions. Statistical significance was set at P < 0.05. All analyses were carried out using SPSS version 21.0 (IBM Corp).

Results

A total of 2610 participants were included in the study, with 5180 retinal images collected. The median age of participants was 58.8 years (interquartile range: 31–85), with 62% identifying as male. The median duration of diabetes was 17 years (interquartile range: 0.59–31), and the median hemoglobin A1c was 9% (interquartile range: 3.9–18.2). Additional key variables are presented in Table 1.

Table 1.

Basic Characteristics of the Participants

Variables	Summary Statistic	Dispersion/Range
No. of participants	2610	-
Total no. of images	5180	-
Age (years), median (IQR)	58.8	8.8 [31–85]
Male, n (%)	1618 (62%)	-
Duration of diabetes (years), mean (SD range)	17	6.9 [0.59–31]
Glycated hemoglobin, HbA1c (%), mean (SD range)	9	2.1 [3.9–18.2]
Low-density lipoprotein (mg/dL), mean (SD range)	112.6	39.2 [32–358]
High-density lipoprotein (mg/dL), mean (SD range)	49.8	38.3 [14–378]
Triglycerides (mg/dL), mean (SD range)	163.5	100.0 [20.1–450]
Albuminuria (%)	32	-
Hypertension (%)	39	-

HbA1c = glycated hemoglobin; IQR = interquartile range; SD = standard deviation.

Image Grading

There was a substantial degree of both intragrader reliability (kappa = 0.94 [95% confidence interval: 0.92–0.97]) and intergrader reliability (kappa = 0.91 [95% confidence interval: 0.89–0.93]) for DR grading on fundus photographs. Comparison of DR grades assigned on 45° images with their corresponding full 200° UWF images showed that agreement decreased with increasing disease severity, reflecting peripheral lesions missed in the central field. The detailed comparison is summarized in Table 2.

Table 2.

Agreement between 45° Central Fundus Images and Full 200° UWF DR Grading

Full UWF DR Grade	Number of Images (n)	45° Grade Matches UWF, n (%)	45° Grade Underestimates, n (%)	45° Grade Overestimates, n (%)
No DR	1177	1177 (100%)	0 (0%)	0 (0%)
Mild NPDR	801	758 (94.6%)	43 (5.4%)	0 (0%)
Moderate NPDR	962	809 (84.1%)	153 (15.9%)	0 (0%)
Severe NPDR	700	507 (72.4%)	193 (27.6%)	0 (0%)
Proliferative DR	1540	1075 (69.8%)	465 (30.2%)	0 (0%)
Total	5180	4326 (83.5%)	854 (16.5%)	0 (0%)

DR = diabetic retinopathy; NPDR = nonproliferative diabetic retinopathy; UWF = ultra-widefield.

These results indicate that while 45° central images capture most lesions in early stages, peripheral pathology increasingly contributes to higher DR grades in moderate to proliferative disease, with 16.5% of eyes showing underestimation when only the central 45° image was used, highlighting the importance of peripheral evaluation in certain cases.

Comparison of Model Performance across Different FOV

The performance of the DL model was evaluated across 3 distinct datasets, each simulating different retinal imaging conditions by varying the FOV and masking. The datasets comprised: (1) unmasked UWF images with a 200° FOV, (2) masked 45° images labeled based on re-evaluation of the masked images alone, and (3) masked 45° images that retained the original labels from the corresponding unmasked UWF images.

• (i)Performance on the unmasked 200° FOV dataset

In the first experimental condition, the model was trained and tested using unmasked UWF (200°) retinal images, providing the most comprehensive view of the retina. This dataset included both central and peripheral retinal features, enabling the model to leverage a complete representation of DR manifestations. The model achieved a training accuracy of 98.97% and a testing accuracy of 97.12%, indicating a strong capacity to generalize to unseen data. These results demonstrate that the model performed well when trained on full-field images, effectively capturing relevant features across the retina to support accurate DR severity classification.

The near-perfect training accuracy indicates that the model was able to learn the underlying patterns in the dataset effectively. However, the slightly lower testing accuracy suggests a minor degree of overfitting, implying that while the model performed exceptionally well on training data, minor variations in unseen test images led to slight deviations in its predictions. Despite this, the overall classification performance remained robust, supporting the hypothesis that UWF imaging provides valuable diagnostic information for automated DR grading. The ROC curves and confusion matrix are presented in Figure 4A, B, respectively, while the detailed performance metrics are provided in Table 3.

• (ii)Performance on the masked 45° FOV dataset with re-evaluated labels

Figure 4.

A, Receiver operating characteristic curves and (B) confusion matrix for the unmasked 200° FOV dataset. AI = artificial intelligence; FOV = field of view; ROC = receiver operating characteristic.

Table 3.

Performance Metrics for All Datasets

Dataset/DR Stage	Accuracy	Precision	Recall	F1-Score
Unmasked 200° dataset
No DR	0.97	0.97	0.98	0.97
Mild NPDR	0.97	0.96	0.97	0.96
Moderate NPDR	0.97	1.00	0.98	0.99
Severe NPDR	0.97	0.96	0.99	0.97
Proliferative DR	0.97	0.97	0.94	0.96
Masked 45° dataset (reevaluated labels)
No DR	0.97	0.97	1.00	0.97
Mild NPDR	0.97	0.97	0.97	0.96
Moderate NPDR	0.97	0.98	0.98	0.99
Severe NPDR	0.97	0.97	0.99	0.97
Proliferative DR	0.97	0.97	0.94	0.96
Masked 45° dataset (original labels)
No DR	0.97	0.97	0.96	0.96
Mild NPDR	0.97	0.98	0.97	0.97
Moderate NPDR	0.97	0.95	1.00	0.97
Severe NPDR	0.97	0.96	0.98	0.97
Proliferative DR	0.97	0.98	0.93	0.96

DR = diabetic retinopathy; NPDR = nonproliferative diabetic retinopathy.

To simulate the imaging conditions of conventional fundus cameras, which typically capture a 45° central FOV, the second dataset was created by masking the peripheral regions of UWF images, retaining only the central retina, followed by relabeling. Using this dataset, the model achieved a training accuracy of 99.22% and a testing accuracy of 97.24%. Notably, despite the absence of peripheral retinal information, the model demonstrated slightly improved testing accuracy compared to the unmasked 200° dataset. The ROC curves and confusion matrix are presented in Figure 5A, B, respectively, while the detailed performance metrics are provided in Table 3.

Figure 5.

A, Receiver operating characteristic curves and (B) confusion matrix for the masked dataset with reevaluated labels. AI = artificial intelligence; ROC = receiver operating characteristic.

This result suggests that central retinal features alone are sufficient for DR classification, at least within the framework of the International Clinical Diabetic Retinopathy Disease Severity Scale. The minor improvement in testing accuracy indicates that while UWF imaging provides additional details, the exclusion of peripheral regions did not substantially hinder the model's classification performance. One possible explanation is that key DR-related features such as microaneurysms, hemorrhages, and cotton-wool spots are frequently located in the central retina, making the 45° view an effective reference for grading severity. The slight improvement in accuracy may also reflect the reduced complexity of the dataset, as peripheral artifacts, variations in illumination, and lens-induced distortions, more common in UWF imaging, were eliminated through masking.

These findings highlight an important implication for clinical practice: while UWF imaging enhances diagnostic capabilities, conventional fundus cameras capturing a 45° field can still provide highly reliable assessments for DR severity grading when combined with DL models.

Performance on the Masked 45° FOV Dataset with Original Unmasked Labels

The third dataset was designed to assess whether the features visible in the masked images alone can predict the original labels from the corresponding unmasked images. It uses the same masked images as the second dataset but applies the labels from the full, unmasked images. This helps evaluate if central retinal features suffice for accurate classification. The model trained on this dataset achieved the highest training accuracy among all 3 configurations, reaching 99.33%, while the testing accuracy was slightly lower at 96.86%. The increased training accuracy suggests that the model was able to learn the available features effectively; however, the slightly reduced testing accuracy indicates a minor drop in generalization capability. The decreased performance in the test set, compared to both the unmasked and 45° masked datasets, suggests a potential case of overfitting. The ROC curves and confusion matrix are presented in Figure 6A, B, respectively, while the detailed performance metrics are provided in Table 3.

Figure 6.

A, Receiver operating characteristic curves and (B) confusion matrix for the masked dataset with original labels. AI = artificial intelligence; ROC = receiver operating characteristic.

Additionally, the drop in testing accuracy relative to the unmasked 200° dataset suggests that the model is not provided with the peripheral features to learn. When compared with the second dataset, which used re-evaluated labels, nearly the same accuracy was observed, implying that the model can learn and utilize the central retinal features. This result reinforces the notion that peripheral retinal features contribute valuable information to DR classification. Our models can perform predictions with reasonable accuracy even without these peripheral retinal features.

Overall Performance Comparison and Implications

Across all 3 datasets, the DL model exhibited consistently high classification performance, with testing accuracies ranging between 96.86% and 97.24%. Despite differences in retinal coverage, the model demonstrated strong reliability in DR staging, as evidenced by the high average precision, recall, and F1-score of 0.97 across all conditions. The findings suggest that while UWF imaging provides additional diagnostic details, central retinal features alone remain highly informative for automated DR grading.

The highest testing accuracy was observed for the masked 45° dataset with reevaluated labels, implying that the central retina holds sufficient pathological information for effective classification. Meanwhile, the highest training accuracy was observed in the masked 45° FOV dataset with original labels, which also exhibited a slight drop in testing performance. The unmasked UWF dataset performed robustly, reinforcing the value of comprehensive retinal imaging for clinical applications.

To assess model robustness, a sensitivity analysis was conducted using only real retinal images. On the unmasked 200° dataset, performance was excellent, with accuracy ranging from 0.99 to 1.00 across DR severity levels and consistently high precision, recall, and F1-scores. When applied to the 45° dataset, the original labels corresponding to the unmasked 200° dataset were accurately recovered, achieving accuracy between 0.98 and 0.99, with similarly high precision, recall, and F1-scores (Table 4). These results demonstrate that model performance remains robust and reliable on real retinal images, independent of generative augmentation.

Table 4.

Model Performance on Real Retinal Images

Dataset/DR Stage	Accuracy	Precision	Recall	F1-Score
Unmasked 200° dataset
No DR	0.99	1.00	1.00	1.00
Mild NPDR	0.99	0.96	1.00	0.98
Moderate NPDR	0.99	1.00	1.00	1.00
Severe NPDR	0.99	0.97	1.00	0.99
Proliferative DR	0.99	1.00	0.98	0.99
Masked 45° dataset
No DR	0.98	1.00	1.00	1.00
Mild NPDR	0.98	0.98	1.00	0.99
Moderate NPDR	0.98	0.96	1.00	0.98
Severe NPDR	0.98	0.98	0.98	0.98
Proliferative DR	0.98	0.99	0.98	0.99

DR = diabetic retinopathy; NPDR = nonproliferative diabetic retinopathy.

Figure 7 presents Grad-CAM visualizations for cases of DR, comparing the model's attention across 200° UWF and 45° retinal images. In Figure 7A–C, the model focuses on different regions of interest, capturing both peripheral and central features in the 200° FOV images (i) and primarily central features in the 45° images (ii), yet consistently classifies the condition as mild, moderate, and PDR. This indicates that the model can identify relevant pathological cues across varying FOVs to arrive at a consistent diagnosis. However, in Figure 7D, the model arrives at a different classification, suggesting that FOV and the corresponding visual context can, in certain cases, influence diagnostic outcomes. To determine which features are most relevant for achieving an accurate diagnosis, it is notable that vascular parameters are key in guiding the model's attention, highlighting their crucial role in the accurate classification of DR across varying FOVs. To extract these parameters, quantification of retinal vessel features was performed using image processing techniques.

Figure 7.

Gradient-weighted class activation mapping visualizations comparing model attention between 200° UWF and simulated 45° retinal images across different stages of DR. Panels correspond to disease severity as follows: (A) mild NPDR, (B) moderate NPDR, (C) PDR, and (D) severe NPDR. Within each panel, (i) denotes 200° UWF images, while (ii) represents simulated 45° FOV images. DR = diabetic retinopathy; NPDR = nonproliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy; UWF = ultra-widefield.

To show an example of AI-assisted detection of DR, Figure 8 presents a full 200° UWF fundus image (panel A) graded as PDR with extensive peripheral neovascularization (yellow circles). The black circle in panel A indicates the area cropped to generate the corresponding 45° central image (panel B), connected by the red arrow. Peripheral neovascularization is not visible in this central crop, which was graded as moderate NPDR, illustrating how conventional central-field grading can underestimate disease severity. The DL model output (panel C) correctly classifies the image as PDR, demonstrating AI's ability to infer peripheral pathology from central-field features.

Figure 8.

Illustration of DL-based detection of peripheral DR. (A) 200° UWF fundus image graded as PDR, (B) corresponding simulated 45° central field image graded as moderate NPDR, and (C) DL model output classifying the image as PDR. AI = artificial intelligence; DL = deep learning; DR = diabetic retinopathy; NPDR = nonproliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy; UWF = ultra-widefield.

When DR on 45° fundus images against the reference standard of 200° UWF images, human graders underestimated disease severity in 16.5% of cases (Table 2). In the DL-based assessment, this underestimation was reduced to 6.2%, with consistent improvement across all DR stages. Agreement for mild NPDR increased from 94.6% with human grading to 95.3% with AI, while moderate NPDR improved from 84.1% to 93.3%. For severe NPDR, agreement rose from 72.4% to 90.2%, and in PDR, it increased from 69.8% to 90.3%. Both human and AI grading maintained 100% accuracy for eyes without DR. These findings show that AI can substantially enhance DR severity assessment, particularly in advanced stages, by capturing subtle retinal features that may be overlooked on standard 45° images.

The presence or absence of DR lesions was assessed in the peripheral 155° field. Table 5 shows the frequency distribution of each lesion type. Peripheral microaneurysms were observed in 19.8% of eyes, hemorrhages and exudates in 9.1%, IRMA in 13.7%, and NVE in 13.8%. Chi-square analysis showed that microaneurysms (χ² = 120.19, P < 0.001) and hemorrhages/exudates (χ² = 92.73, P < 0.001) were significantly more likely to occur in specific image groups. In contrast, no significant differences were observed in the frequency of IRMA (χ² = 0.6, P = 0.896) or NVE (χ² = 0.31, P = 0.958). These findings indicate that early lesions are more frequently detected in the periphery compared to more advanced changes.

Table 5.

Frequency of Lesion Types in the Peripheral 155° Field

Total Images (n) = 5180
Lesion Type	Present (n, %)	Absent (n, %)	χ2 Value	P Value
Microaneurysms	1025 (19.8%)	4155 (80.2%)	120.19	<0.001
Hemorrhages and exudates	469 (9.1%)	4711 (90.9%)	92.73	<0.001
IRMA	708 (13.7%)	4472 (86.3%)	0.6	0.896
NVE	714 (13.8%)	4466 (86.2%)	0.31	0.958

IRMA = intraretinal microvascular abnormality; NVE = neovascularization elsewhere.

Following image processing, quantitative analysis of retinal vascular parameters revealed significant differences between eyes with and without specific DR lesions (Table 6). Vessel density was consistently and significantly lower in eyes with microaneurysms (0.476 ± 0.010 vs. 0.482 ± 0.012, P = 0.000), hemorrhages and exudates (0.470 ± 0.015 vs. 0.494 ± 0.014, P = 0.000), IRMA (0.468 ± 0.017 vs. 0.490 ± 0.011, P < 0.001), and NVE (0.471 ± 0.011 vs. 0.487 ± 0.013, P = 0.000), indicating reduced retinal perfusion in the presence of these lesions.

Table 6.

Comparison of Retinal Vascular Parameters in the Presence and Absence of DR Lesions

Vascular Parameter	Lesion Type	Absent (Mean ± SD)	Present (Mean ± SD)	P Value
Vessel density (%)	Microaneurysms	0.482 ± 0.012	0.476 ± 0.010	0.000
	Hemorrhages and exudates	0.494 ± 0.014	0.470 ± 0.015	0.000
	IRMA	0.490 ± 0.011	0.468 ± 0.017	<0.001
	NVE	0.492 ± 0.015	0.471 ± 0.011	0.000
Fractal index	Microaneurysms	1.021 ± 0.276	1.003 ± 0.200	0.046
	Hemorrhages and exudates	1.017 ± 0.260	1.004 ± 0.210	0.279
	IRMA	1.022 ± 0.289	1.002 ± 0.203	0.069
	NVE	1.018 ± 0.285	1.007 ± 0.204	0.317
Tortuosity (unitless)	Microaneurysms	1.091 ± 0.026	1.095 ± 0.019	<0.001
	Hemorrhages and exudates	1.093 ± 0.019	1.098 ± 0.021	<0.001
	IRMA	1.095 ± 0.020	1.093 ± 0.019	0.046
	NVE	1.098 ± 0.020	1.097 ± 0.020	0.317
Median tortuosity (unitless)	Microaneurysms	1.064 ± 0.017	1.080 ± 0.010	0.000
	Hemorrhages and exudates	1.060 ± 0.019	1.076 ± 0.016	<0.001
	IRMA	1.058 ± 0.019	1.075 ± 0.012	0.000
	NVE	1.061 ± 0.022	1.079 ± 0.016	0.000

DR = diabetic retinopathy; IRMA = intraretinal microvascular abnormality; NVE = neovascularization elsewhere; SD = standard deviation.

The fractal index, a measure of vascular branching complexity, was slightly lower in eyes with lesions. A statistically significant difference was observed only in the presence of microaneurysms (P = 0.046), while changes associated with other lesions were not significant.

Both mean and median tortuosity were significantly higher in eyes with lesion presence, reflecting increased vascular distortion. Mean tortuosity was elevated in eyes with microaneurysms (1.095 ± 0.019 vs. 1.091 ± 0.026, P < 0.001) and hemorrhages (1.098 ± 0.021 vs. 1.093 ± 0.019, P < 0.001), while differences with IRMA and NVE were marginal or nonsignificant. In contrast, median tortuosity was significantly higher across all lesion types (all P ≤ 0.001), suggesting a consistent pattern of increased vascular curvature in affected eyes. Together, these findings suggest that the presence of peripheral lesions is associated with measurable central vascular alterations, including decreased perfusion and increased vessel tortuosity, highlighting their potential role as noninvasive biomarkers in DR progression.

Discussion

This study evaluated the performance of a DL model in classifying DR severity using retinal fundus images with varying FOV from 200° UWF images to centrally masked 45° views simulating standard fundus photography. The model achieved high testing accuracies across all 3 datasets: 97.12% for full UWF images, 97.24% for 45° masked images with re-evaluated labels, and 96.86% for masked images retaining original full-field labels. While the model achieved strong testing accuracies, the near-perfect training accuracies (up to 99.3%) suggest mild overfitting. These findings demonstrate the robustness of AI-based DR classification and confirm that central retinal information alone can yield reliable diagnostic performance, reinforcing the practical utility of 45° imaging in diverse clinical settings, especially in low-resource environments. Importantly, the sensitivity analysis performed exclusively on real retinal images further supports the robustness and clinical relevance of the proposed approach.

In the early stages of DR, characterized by microaneurysms, dot/blot hemorrhages, and hard exudates, the model performed exceptionally well. In the 45° masked dataset with re-evaluated labels, the model achieved precision and recall of 0.97 to 0.98 for both mild and moderate NPDR, indicating that central lesions were well captured and effectively classified. Notably, even in eyes with peripheral lesions beyond the 45° field, the model preserved its classification accuracy. Peripheral microaneurysms and hemorrhages/exudates were present in 19.8% and 9.1% of cases, respectively, yet the model maintained 97.24% accuracy, highlighting its ability to detect early disease even when some peripheral features are masked. This suggests that the model uses nuanced structural and vascular features in the central retina to infer DR severity, which is especially important for large-scale screening, where widefield imaging is unavailable.

Beyond early disease, the model also demonstrated the ability to infer advanced DR stages, including severe NPDR and PDR, despite restricted FOV. In clinical practice, accurate grading of DR severity is critical for timely referral and intervention. Human grading of 45° fundus images underestimated disease severity in 16.5% of cases compared with UWF imaging, primarily due to undetected peripheral lesions. Using the DL-based assessment, this underestimation was reduced to 6.2%, with the most pronounced improvement observed in advanced stages, where agreement increased from 72.4% to 90.2% in severe NPDR and from 69.8% to 90.3% in PDR. These findings indicate that the DL-based assessment can detect subtle vascular biomarkers in the central retina that reflect peripheral pathology, allowing reliable severity grading even when only standard fundus images are available.

Gradient-weighted class activation mapping visualizations showed that the model focused on central retinal regions without overt lesions upon clinical inspection. Vascular analysis of these areas revealed significant differences between eyes with and without peripheral pathology. For instance, vessel density in eyes with IRMA was 0.468 ± 0.017 compared to 0.490 ± 0.011 in those without IRMA (P < 0.001); similarly, median tortuosity in eyes with NVE was 1.079 ± 0.016 versus 1.061 ± 0.022 in unaffected eyes (P < 0.001). These findings support the concept that the model detects subtle, subclinical microvascular changes reflective of peripheral disease activity. Such inference is clinically meaningful; it means that even without visualizing peripheral IRMA or neovascularization, AI can flag eyes that may warrant further evaluation or early intervention. Vessel density and tortuosity extend beyond the traditional lesion-based classification of DR. These continuous vascular biomarkers reflect diffuse endothelial dysfunction, altered blood flow, and early angiogenic remodeling, processes that may occur before or alongside visible lesions.³⁵ While vessel density and tortuosity demonstrated consistent and statistically significant differences across most lesion types, the fractal index showed more subtle changes for IRMA and NVE. This variability reflects the complementary nature of different vascular features, suggesting that multiple biomarkers together can provide a richer and more nuanced assessment of retinal microvascular health. Artificial intelligence-driven quantification of these subtle features thus provides insight into microvascular health and identifies predictive markers that are often imperceptible in standard clinical examination. Importantly, these findings provide important proof of concept that 45° fundus images contain vascular biomarkers predictive of peripheral DR, thereby supporting their potential role as a clinically viable surrogate for UWF imaging in both screening and diagnostic settings.

This capability is further strengthened by the model's interpretability. Gradient-weighted class activation mapping heatmaps revealed attention in areas of the central retina that overlapped with vascular abnormalities, even when no clear lesions were visible. These attention maps enhance clinical confidence in AI predictions and align with the understanding that DR represents a diffuse retinal vascular disease, where central structural biomarkers may indirectly reflect peripheral pathology.

From a practical standpoint, these results reinforce the value of AI-powered DR screening using standard 45° fundus cameras, which are more accessible and affordable than UWF systems. In primary care and rural settings, where UWF imaging is not feasible, such models can provide accurate grading for early DR and serve as a triage tool to identify advanced cases for referral. This dual utility, combining precise detection of mild disease with inferred prediction of high-risk peripheral involvement, positions AI as a transformative and highly scalable tool in diabetic eye care.

This study presents a novel method that classifies DR and detects peripheral lesions using central retinal biomarkers, but few limitations exist. Data originated from a single South Indian eye care center, which may limit generalizability to other ethnic groups and disease profiles; however, inclusion of systemic variables and vascular biomarker analysis helps enhance applicability. The exact FOV was not precisely measured, and factors such as focusing errors or eyelid/eyelash obstruction may have reduced the effective retinal area imaged. Real 45° central-field images were not used, with cropped UWF regions serving as a simulation; future work will employ real 45° images to better represent conventional imaging. Images were graded by an optometrist following a standardized protocol with assessed intergrader reliability, though future studies will include retinal specialists to further enhance grading accuracy. Gradient-weighted class activation mapping was used to visualize model focus, providing region-level insight but limited pixel-level interpretability. This study primarily serves as a simulation framework, with plans to scale to larger, multiethnic datasets. Future multicenter studies incorporating more diverse demographic and clinical populations will be essential to rigorously evaluate the robustness, generalizability, and external validity of the proposed approach across varied health care settings. Additionally, future research may explore alternative DL models, vision transformers, or multimodal imaging such as OCT and fluorescein angiography to improve classification performance and feature extraction.

Conclusion

This study demonstrates that DL models can achieve high accuracy in DR classification using standard 45° fundus images, not only for detecting visible early-stage DR lesions but also for inferring peripheral advanced disease through central vascular biomarkers. With classification accuracies exceeding 97% and clear associations between vascular changes and peripheral pathology, this AI-driven approach offers a scalable, interpretable, and cost-effective solution for DR screening. It holds promise for expanding access to early detection and triage in low-resource settings, while supporting real-time clinical decision-making in routine ophthalmic care. Future work should focus on broader external validation and longitudinal follow-up to further establish predictive value and clinical integration.

Manuscript no. XOPS-D-25-00748.