Towards Sustainable Retinal Diagnostics A Deep Learning Alternative to OCT for Macular Thickness Estimation

Towards Sustainable Retinal Diagnostics A Deep Learning Alternative to OCT for Macular Thickness Estimation | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Towards Sustainable Retinal Diagnostics A Deep Learning Alternative to OCT for Macular Thickness Estimation C A Aparna, B R Manju This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6795425/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 19 Apr, 2026 Read the published version in Discover Sustainability → Version 1 posted 14 You are reading this latest preprint version Abstract Diabetic Macular Edema (DME) is a major cause of vision impairment among diabetic patients, especially in areas lacking advanced diagnostic tools like Optical Coherence Tomography (OCT). This study proposes a cost-effective, AI-based method to estimate macular thickness—a key DME biomarker—using retinal fundus images and clinical data. A dataset of 176 colour fundus images with corresponding macular thickness and clinical information from DME patients at Amrita Institute of Medical Sciences, Kochi, India, was augmented to about 3,000 images to improve model performance. Manual macular region annotations were done using the CVAT tool, followed by segmentation using a U-Net-based deep learning model. For predicting macular thickness, three convolutional neural networks—InceptionV3, DenseNet121, and EfficientNetB0—were tested in a multimodal framework combining image features with clinical parameters like visual acuity and diabetes duration. EfficientNetB0 achieved the best performance, with a test Mean Squared Error (MSE) of 0.2518 and an R² score of 0.7752. This approach enables accurate, low-cost, and scalable diagnostics in underserved regions, aligning with Sustainable Development Goals: ensuring healthy lives (SDG 3), fostering innovation (SDG 9), and reducing inequalities (SDG 10). The results suggest that this AI-based method can serve as a viable alternative to OCT for macular thickness estimation, supporting early detection and monitoring of DME in low-resource settings. Diabetic Macular Edema (DME) Macular Thickness CVAT (Computer Vision Annotation Tool) Deep Learning U-Net Segmentation EfficientNetB0 InceptionV3 DenseNet121 Sustainable healthcare Health equity Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1. Introduction Macular edema is a pathological condition characterized by the accumulation of fluid in the macula—the central region of the retina responsible for sharp and detailed vision. This swelling distorts vision and can lead to significant visual impairment if left untreated. Macular edema can arise from several causes, including age-related macular degeneration (AMD), retinal vein occlusion (RVO), and most notably, diabetic retinopathy, where it is termed Diabetic Macular Edema (DME).Among these, DME is the most common form of macular edema, and it has become an increasingly serious public health concern, especially in countries with a high prevalence of diabetes.[1] In India, the incidence of diabetes is growing at an alarming rate, with projections estimating over 100 million diabetic individuals by 2030. This surge has led to a corresponding increase in diabetes-related eye complications. Studies suggest that approximately 7–14% of diabetic patients in India develop DME, making early detection and monitoring a critical priority for preventing vision loss.[2] In fundus imaging, microaneurysms and hard exudates are key patterns indicative of diabetic macular edema (DME). Microaneurysms are small, round outpouchings of capillaries in the retina, appearing as tiny red dots in fundus images. They represent the earliest clinically detectable signs of diabetic retinopathy and are indicative of capillary wall weakness. These microaneurysms can leak fluid, contributing to retinal swelling characteristic of DME. Their presence is crucial for early diagnosis and monitoring of disease progression.[1,3] Hard exudates are lipid-rich deposits that manifest as yellowish-white spots with sharp margins in the retina. They result from the leakage of lipoproteins and other plasma constituents from compromised retinal vessels. In fundus images, hard exudates often form a circinate pattern around areas of leakage, particularly near microaneurysms. Their accumulation, especially within the macular region, is a hallmark of DME and correlates with the severity of macular thickening and potential vision loss.[4] Retinal thickness is a crucial biomarker assessed in ophthalmology, particularly in the diagnosis and monitoring of diabetic macular edema (DME). In healthy individuals, the central macular thickness typically ranges between 250 to 280 micrometres (µm) — where 1 micrometre (µm) equals 1000 nanometres (nm). However, in patients with DME, the retinal thickness often increases significantly, frequently exceeding 300 micrometres due to fluid accumulation within the retinal layers. Monitoring changes in retinal thickness with high precision is essential for evaluating disease progression and treatment response. [5] The gold standard for diagnosing and monitoring DME is Optical Coherence Tomography (OCT), a non-invasive imaging technique that captures high-resolution cross-sectional images of the retina and provides accurate measurements of macular thickness. However, OCT devices are expensive, require skilled operators, and are often confined to specialized medical care centres, limiting their availability in rural and resource-constrained settings across India and other developing countries. In contrast, retinal fundus photography is far more accessible, cost-effective, and already used widely in diabetic screening programs.[2,6] While fundus images are inherently two-dimensional and do not offer depth information like OCT, advances in artificial intelligence (AI) and deep learning have created opportunities to infer 3D structural information from 2D images.[4,7] This study proposes a cost-effective, AI-driven alternative to OCT for estimating macular thickness—a key biomarker for DME diagnosis and progression monitoring. Our pipeline is built to mimic OCT-derived macular thickness maps using only fundus photographs and clinical metadata. The process involves several key stages to enable macular thickness estimation from fundus images. First, manual mask preparation was carried out for the macular region using the CVAT annotation tool, allowing for precise labelling of the region of interest. Next, semantic segmentation of the macula was performed using a U-Net-based deep learning model, which effectively isolated the macular area from the rest of the retinal image. Finally, the study used a modified EfficientNetB0 model that incorporates both image and clinical data to predict macular thickness.[8] This was achieved by adopting a multimodal architecture in which high-level features extracted from fundus images using EfficientNetB0 are concatenated with relevant tabular clinical features, such as visual acuity (VA) and diabetes duration. This fusion of image and clinical data enhances the model’s ability to learn complex patterns and improves the overall regression performance. In this study, modified versions of DenseNet121 and InceptionV3 were also implemented for comparison; however, the modified EfficientNetB0 model demonstrated comparatively better performance, making it the most effective choice for the proposed task.[8,9] By estimating macular thickness directly from fundus images combined with clinical data, this study aims to assist clinicians in identifying early signs of Diabetic Macular Edema (DME) without relying on OCT imaging. The proposed model leverages a multimodal deep learning framework that integrates high-level image features extracted by EfficientNetB0 with patient-specific clinical parameters such as visual acuity and diabetes duration.[10,11] This approach can potentially democratize access to retinal diagnostics, particularly in under-resourced regions, and contribute to scalable diabetic eye care in India and beyond. By offering a cost-effective, AI-driven alternative to OCT for macular thickness estimation—an essential parameter in the early detection and monitoring of DME—the study addresses the diagnostic gap in resource-constrained settings. The framework's reliance on widely available retinal fundus images, enhanced by clinical context, promotes timely intervention, improved visual outcomes, and reduced burden of diabetic eye disease, especially in low- and middle-income countries.[12,13] Beyond its technical contributions, the proposed framework aligns strongly with key goals of sustainable development. By leveraging artificial intelligence to replicate the diagnostic utility of Optical Coherence Tomography (OCT) using affordable and widely available retinal fundus images, the model supports SDG 3 (Good Health and Well-being) by improving access to early diabetic eye care in underserved regions. Furthermore, the approach exemplifies SDG 9 (Industry, Innovation, and Infrastructure) through its application of cutting-edge AI methods to strengthen healthcare infrastructure in low-resource settings without the need for high-cost diagnostic equipment. Importantly, this solution also contributes to SDG 10 (Reduced Inequalities) by bridging the gap in diagnostic access between urban tertiary hospitals and rural or economically disadvantaged communities. The study’s focus on cost-effectiveness, scalability, and clinical utility positions it as a practical innovation for advancing equitable and sustainable eye health systems globally.[14–16] 2. Literature Review Recent advancements in artificial intelligence (AI) have significantly impacted the early detection and management of diabetic macular edema (DME), particularly in resource-limited settings where access to optical coherence tomography (OCT) is constrained. Several studies have explored the potential of deep learning models to predict macular thickness and detect DME using fundus photographs and clinical data. Akça et al. (2024) conducted a comparative study using vision transformers for automated classification of choroidal neovascularization, DME, and drusen from retinal OCT images. The study highlighted the effectiveness of vision transformers in accurately classifying retinal diseases, including DME. [3] Chavan and Choubey (2024) developed a self-supervised category selective attention classifier network for DME classification. The model leveraged self-supervised learning to enhance classification performance, demonstrating its potential in scenarios with limited labelled data. [1] Qin et al. (2024) developed an ensemble deep learning model to predict the onset of centre-involved DME within a year using ultra-wide field colour fundus images. The model achieved an AUC of 0.7017 and an F1 score of 0.6512, demonstrating its potential in early DME prediction.[4] Lam et al. (2024) conducted a systematic review and meta-analysis evaluating AI performance in detecting DME from fundus photography (FP) and OCT images. Analysing 53 studies, they reported pooled sensitivity and specificity of 92.6% and 91.1% for FP-based algorithms, and 95.2% and 92.5% for OCT-based algorithms, respectively, highlighting AI's robust diagnostic capabilities. ​[17] Midena et al. (2023) validated an automated AI algorithm for quantifying major OCT parameters in DME. Their model demonstrated high accuracy in measuring central subfield thickness and macular volume, aligning closely with manual assessments, thus supporting its potential for clinical application.[18] Ye et al. (2023) developed an AI-based method to quantify central macular fluid volume and predict visual acuity in DME patients using OCT images. The model showed strong correlations between predicted and actual visual acuity, suggesting its utility in personalized treatment planning.[19] Li et al. (2023) performed a systematic review and meta-analysis on deep learning algorithms for DME detection in OCT images. They concluded that AI models achieved high diagnostic accuracy, with pooled sensitivity and specificity of 94% and 92%, respectively, reinforcing AI's effectiveness in OCT image analysis.[13] Abramovich et al. (2022) introduced FundusQ-Net, a deep learning algorithm for assessing fundus image quality. Trained on a large dataset, the model achieved a mean absolute error of 0.61 in quality grading, facilitating the selection of high-quality images for reliable AI-based DME detection. [20] Springer et al. (2024) developed an OCT-omics prediction model using machine learning to assess anti-VEGF treatment response in DME patients. Their model demonstrated high discriminative abilities, with the backpropagation neural network classifier achieving an AUC of 0.982, indicating its potential in guiding personalized therapy.[8] NEJM AI (2024) discussed the development of an AI-based DME screening tool designed for low-resource settings. The model incorporated techniques for privacy preservation and uncertainty estimation, aiming to improve screening accessibility and reliability in underserved areas.[9] He et al. (2025) proposed a graph neural network-based multispectral-view learning model for detecting diabetic macular ischemia (DMI) from colour fundus photographs. Their approach achieved an accuracy of 84.7% and an AUROC of 0.900, demonstrating the potential of AI in identifying DMI, a condition closely related to DME.[12] Yang et al. (2025) introduced RURANET++, an unsupervised learning framework for automated DME diagnosis. Incorporating SCSE attention mechanisms and dynamic multi-projection head clustering, the model achieved an accuracy of 84.11% and an F1-score of 83.90%, highlighting its effectiveness without the need for extensive libelled data.[7] Song et al. (2023) introduced an AI method based on multi-feature fusion for automatic macular edema classification on spectral-domain OCT images. By combining traditional and deep features, the model achieved improved accuracy in classifying various types of macular edema, including DME. [13]​​ Vinuesa et al. (2020) in [14]conducted a comprehensive expert-driven study to assess the influence of artificial intelligence (AI) on achieving the 169 individual targets within the 17 Sustainable Development Goals (SDGs). Their methodology combined a structured literature review with a consensus-based expert elicitation process involving scholars from diverse disciplines, including engineering, natural sciences, and social sciences. The evaluation focused on determining whether AI served as an enabler or inhibitor for each SDG target. The authors relied on peer-reviewed literature describing real-world AI applications, validated experimental studies, and reports from recognized institutions such as the United Nations and governmental agencies. Notably, speculative opinions, non-peer-reviewed claims, and media reports were excluded to maintain the integrity and quality of the evidence base. This rigorous approach highlighted how AI, when responsibly deployed, can significantly advance key sustainability goals—particularly those related to health (SDG 3), innovation (SDG 9), and equity (SDG 10). 3. Methodology This study presents a cost-effective, AI-driven approach for estimating macular thickness from retinal fundus images and clinical data, with the goal of aiding in the early detection and monitoring of Diabetic Macular Edema (DME), particularly in low-resource settings where Optical Coherence Tomography (OCT) is not readily available. The methodology involves three primary stages: dataset preparation, macular region segmentation, and regression-based thickness prediction. 3.1 Dataset Preparation The dataset was curated from the ophthalmology department at Amrita Institute of Medical Sciences (AIMS), Kochi, Kerala, India, and comprises 176 colour fundus images from patients clinically diagnosed with Diabetic Macular Edema. Along with the fundus images, corresponding clinical information was collected, including visual acuity and duration of diabetes (in years). Ground truth macular thickness values were obtained from OCT scans and served as regression targets during model training and evaluation. To enhance model generalizability and prevent overfitting, the original 176 fundus images were augmented to create a larger dataset of approximately 3000 images. This expanded dataset forms the foundation for developing and validating the proposed AI model, enabling the extraction of relevant features and the accurate prediction of macular thickness. 3.2 Macular Region Segmentation Using U-Net with Mask Preparation via CVAT Tool To localize the macular area from retinal fundus images, a U-Net-based segmentation model was employed. U-Net is a widely used convolutional neural network architecture in medical image analysis due to its ability to capture fine-grained spatial features through skip connections and a symmetric encoder-decoder structure. [21] Manual annotations for the macular region were generated using the CVAT (Computer Vision Annotation Tool), an open-source tool that facilitates precise annotation of medical images with an intuitive interface and flexible support for polygonal masks. These annotations served as ground truth for training the segmentation model. [22] A dataset of 70 grayscale fundus images and their corresponding masks were pre-processed by resizing them to 256×256 pixels and normalizing pixel values to the range [0, 1]. This helped standardize the input for the model and reduce computational complexity. The U-Net architecture consisted of four encoding blocks with convolutional layers followed by max pooling, a bottleneck with high-dimensional feature extraction, and four decoding blocks that used transposed convolutions for up sampling. Skip connections between the encoder and decoder facilitated the recovery of spatial details lost during down sampling. The model was trained using a custom loss function that combined Binary Cross-Entropy (BCE) with a log-transformed Dice loss to balance pixel-wise accuracy and segmentation overlap. Training was carried out for 50 epochs with a batch size of 4, using 10% of the data for validation. A model checkpoint mechanism was employed to save the best-performing model based on validation loss. This segmentation model provides binary masks of the macular region, which are subsequently used for feature extraction and thickness estimation in the regression stage of the pipeline. The U-Net model was trained for macular region segmentation on 70 grayscale fundus images paired with manually annotated masks. Training was carried out over 50 epochs using a batch size of 4, with a 90 − 10 split for training and validation. The input images and masks were resized to 256×256 pixels for computational efficiency.[21] To optimize performance, a custom loss function combining Binary Cross-Entropy and a logarithmically adjusted Dice loss was used. This approach encouraged both pixel-level accuracy and overall shape conformity in segmentation. The model achieved a training accuracy of 93.54% and a validation accuracy of 90.61% . 3.3. Hybrid Deep Learning Framework for Macular Thickness Prediction To evaluate the effectiveness of different convolutional neural network (CNN) architectures in predicting macular thickness, the study compared three deep regression models utilizing InceptionV3, EfficientNetB0, and DenseNet121 as the core image feature extractors. These models were integrated into a multimodal learning framework by combining visual information from fundus images with tabular clinical data—specifically Visual Acuity (VA) and Duration of Diabetes—for a more comprehensive prediction approach. By using the same data preprocessing, model structure, and training protocol, we ensured a fair comparison across architectures. [23] All input fundus images were uniformly resized to 128×128 pixels and processed using histogram equalization in the YUV colour space to enhance local contrast and visual clarity. These enhanced images were then normalized using model-specific preprocessing functions. Tabular features (VA and Diabetes Duration) were standardized using StandardScaler, ensuring they contributed comparably to the learning process. The target variable—macular thickness—was also standardized to stabilize training and improve convergence in the regression models. [23,24] To improve generalization and increase the dataset size, data augmentation applied using Keras’s ImageDataGenerator. This included random transformations such as small rotations, zooms, shifts, and horizontal flips. Through this augmentation process, the dataset was expanded to approximately 3,000 samples, reducing the risk of overfitting and enhancing model robustness. Each of the three model variants followed a consistent design strategy. A pre-trained CNN backbone (InceptionV3, EfficientNetB0, or DenseNet121) was loaded without its classification head and initialized with ImageNet weights. These backbones were initially frozen to retain learned low-level features. Extracted visual features were pooled using GlobalAveragePooling2D, regularized using BatchNormalization and Dropout, then concatenated with the tabular input.[25,26] This merged representation was passed through two fully connected layers (128 and 64 units, both using ReLU activation), followed by a final output layer predicting macular thickness as a single scalar value. The training setup was standardized across models using the Mean Squared Error (MSE) loss function and the Adam optimizer with a learning rate of 1e-4. Model performance was assessed based on both training and validation MSE curves. Final evaluation was conducted on a held-out test set using Test MSE and the R² score to quantify predictive strength. Additionally, scatter plots of predicted versus actual macular thickness values were generated to visually assess regression accuracy and model fit across the range of thickness values. 4. Analysis and results 4.1. InceptionV3 Regression Model Performance for Macular Thickness Prediction The Inception V3 model was trained for 30 epochs with a learning rate of 1e-4, and the training process showed consistent improvement in performance. The training loss started at 1.9761 and steadily decreased to 0.2156 by the final epoch, while the mean squared error (MSE) followed a similar downward trend. The validation loss and MSE also decreased throughout the epochs, suggesting that the model was effectively learning to predict macular thickness. After completing the training, the model was evaluated on the test set, yielding a test MSE of 0.4093 and a test R² score of 0.5859. These results indicate a moderate correlation between the predicted and true values, showing that the model performed reasonably well in predicting macular thickness. 4.2. EfficientNetB0 Regression Model Performance for Macular Thickness Prediction The EfficientNetB0 model was trained over 30 epochs to predict macular thickness. In the first epoch, the model showed a loss of 1.9223 and mean squared error (MSE) of 1.9223, with a validation loss of 0.9261 and a validation MSE of 0.9261. Throughout the epochs, the model improved significantly, with both loss and MSE steadily decreasing, demonstrating its learning process. By epoch 10, the model achieved a validation MSE of 0.4641. As training progressed, the model became more accurate, reaching a validation MSE of 0.2517 in the final epoch. The test evaluation confirmed this progress with a test MSE of 0.2518, and the model's R² score was 0.7752, indicating a strong fit to the data and substantial predictive capability. The model's performance shows consistent improvement throughout the training period, reflecting its ability to generalize well to unseen data. 4.3. DenseNet121 Regression Model Performance for Macular Thickness Prediction DenseNet121, a deep convolutional neural network (CNN) architecture is also used for the regression task of predicting macular thickness from fundus images and associated clinical features. The model was trained over 30 epochs, utilizing a batch size of 32. The training process started with a relatively higher loss and mean squared error (MSE) values, which progressively improved as the model trained. By epoch 30, the model achieved a final training loss of 0.2887 and a final validation loss of 0.2565, indicating a significant reduction in error. This progressive reduction in loss and MSE is indicative of the model's ability to effectively learn the relationship between the input features and the macular thickness, demonstrating the potential of DenseNet121 for medical image analysis tasks. Additionally, the learning rate was kept constant at 1.0000×10 − 4 throughout the training, suggesting that the model’s convergence was robust despite the steady learning rate. The model's performance was further evaluated on a test set, where it achieved a test MSE of 0.30198. Additionally, the R² score of 0.7254 indicates that the model was able to explain a significant portion of the variance in macular thickness prediction. These results highlight the capability of DenseNet121 in accurately predicting macular thickness from both fundus images and clinical features, with substantial potential for clinical applications. 5. Comparative Performance Analysis of Deep Learning Models for Macular Thickness Prediction To evaluate the effectiveness of different deep learning architectures in predicting macular thickness, study compared the performance of InceptionV3, EfficientNetB0, and DenseNet121. The models were assessed based on their Mean Squared Error (MSE) and R² score on the test set. A summary of the results is presented in the table below. Table 1. Comparative analysis of Deep Learning Models. Model Test MSE R² Score InceptionV3 0.4093 0.5859 EfficientNetB0 0.2518 0.7752 DenseNet121 0.3020 0.7254 Among the three deep learning models evaluated—InceptionV3, EfficientNetB0, and DenseNet121—for predicting macular thickness from fundus images and clinical features, EfficientNetB0 demonstrated the most favourable performance. It achieved the lowest test Mean Squared Error (MSE) of 0.2518 and the highest R² score of 0.7752, indicating a stronger predictive accuracy and better generalization to unseen data. In contrast, DenseNet121 showed moderate performance with a test MSE of 0.3020 and an R² score of 0.7254, while InceptionV3 yielded the highest test MSE of 0.4093 and the lowest R² score of 0.5859. These results suggest that EfficientNetB0 is more effective in capturing relevant features from fundus images for accurate regression-based macular thickness estimation compared to the other two architectures. Conclusion Diabetic Macular Edema (DME) is a significant complication of diabetes mellitus, characterized by the accumulation of fluid in the macula—the central part of the retina responsible for sharp vision. This condition arises due to the leakage of fluid from damaged blood vessels in the retina, leading to retinal thickening and vision impairment. DME is a leading cause of vision loss among diabetic patients and poses a growing public health concern, especially in countries like India, where the prevalence of diabetes is rapidly increasing. Early detection and monitoring of DME are crucial to prevent irreversible vision loss. However, the gold standard for diagnosing and monitoring DME, Optical Coherence Tomography (OCT), is expensive and often inaccessible in resource-limited settings. [27] To address this challenge, the study proposes an artificial intelligence (AI)-driven approach to estimate macular thickness—a key biomarker for DME—using retinal fundus images and clinical data. The dataset comprises 176 colour fundus images from patients diagnosed with DME at the Amrita Institute of Medical Sciences in Kochi, Kerala, India. These images were augmented to approximately 3,000 samples to enhance model robustness. Manual annotations of the macular region were performed using the Computer Vision Annotation Tool (CVAT), and a U-Net-based deep learning model was employed for macular segmentation. For macular thickness prediction, three convolutional neural network architectures—InceptionV3, DenseNet121, and EfficientNetB0—were evaluated within a multimodal framework that integrates image features with clinical parameters such as visual acuity and duration of diabetes. All models were trained using standardized preprocessing techniques, including image resizing, histogram equalization, and normalization, as well as data augmentation strategies to improve generalization. Among the evaluated models, EfficientNetB0 demonstrated the most favourable performance, achieving a test Mean Squared Error (MSE) of 0.2518 and an R² score of 0.7752, indicating strong predictive accuracy and generalization capability. DenseNet121 and InceptionV3 showed comparatively lower performance, with test MSEs of 0.3020 and 0.4093, and R² scores of 0.7254 and 0.5859, respectively. These findings suggest that the proposed AI-based method, particularly the EfficientNetB0 model, can serve as a viable alternative to OCT for macular thickness estimation, facilitating early detection and monitoring of DME in resource-constrained settings. This research not only presents a technologically sound solution for macular thickness estimation but also contributes meaningfully to sustainable healthcare goals. By reducing dependence on costly OCT infrastructure, the proposed AI-driven approach aligns with the principles of SDG 3 (Good Health and Well-being) through enhanced early detection in underserved populations. It supports SDG 9 (Industry, Innovation, and Infrastructure) by promoting accessible medical innovation and reinforces SDG 10 (Reduced Inequalities) by bridging diagnostic gaps between urban and rural healthcare systems. As such, this work represents a step toward more inclusive, affordable, and scalable eye care solutions worldwide. Future Work A key future direction involves enhancing the transparency and interpretability of the proposed deep learning model through the integration of explainable AI techniques. Tools such as Grad-CAM (Gradient-weighted Class Activation Mapping), LIME (Local Interpretable Model-agnostic Explanations), can be employed to visualize which regions of the fundus image or which clinical features most significantly influence the model's prediction of macular thickness. This interpretability is crucial for clinical acceptance, as it enables ophthalmologists to verify whether the model is focusing on relevant pathological features like microaneurysms and hard exudates associated with diabetic macular edema (DME). [28] Future work will also focus on expanding the dataset to include images from diverse populations to enhance generalizability, deploying the model in real-time clinical environments for field validation, and integrating explainable AI (XAI) techniques to improve transparency and clinician trust. In addition, exploring deployment on mobile or edge devices will further support sustainable, on-the-ground screening solutions in rural and remote regions. Declarations Author Contributions Statement Aparna C A conceptualized the study, designed the methodology, collected and curated the clinical data and fundus images. Dr Manju B R analysed the result and supervised. Competing Interests The authors declare that they have no competing interests. Conflict of interests The authors have no conflicts of interest to declare that are relevant to the content of this article. Ethics Approval This study was approved by the Institutional Ethics Committee of Amrita Institute of Medical Sciences, Kochi, India. All procedures performed and data collected were in accordance with the ethical standards of the institution. Ethics Statement The research protocol titled "Towards Sustainable Retinal Diagnostics A Deep Learning Alternative to OCT for Macular Thickness Estimation " was reviewed and approved by the Ethics Committee of Amrita School of Medicine on November 8, 2024 (Approval No. ECASM-AIMS-2024-506, Reg. No. EC/NEW/INST/2023/KL/0379), in accordance with institutional guidelines and relevant national ethical regulations. The committee found no ethical issues, and clearance was granted. All patient data used in this study were anonymized prior to analysis. Consent to Participate Informed consent was obtained from all participants or their legal guardians prior to data collection and image acquisition. Data and Code Availability The datasets generated and analysed during the current study are not publicly available due to institutional privacy policies but are available from the corresponding author on reasonable request. The code used for training and evaluation of models is available from the corresponding author. Funding No funding was received for conducting this study Clinical trial number : Not applicable. Consent to Publish declaration: Not applicable References Chavan S, Choubey N. Self-supervised category selective attention classifier network for diabetic macular edema classification. Acta Diabetol. 2024;61:879–96. Gan F, Wu F-P, Zhong Y-L. Artificial intelligence method based on multi-feature fusion for automatic macular edema (ME) classification on spectral-domain optical coherence tomography (SD-OCT) images. Front Neurosci. 2023;17:1097291. Akça S, Garip Z, Ekinci E, Atban F. Automated classification of choroidal neovascularization, diabetic macular edema, and drusen from retinal OCT images using vision transformers: a comparative study. Lasers Med Sci. 2024;39:140. Qin P, Thirunavukarasu A, Arvanitis 1 T. Deep Learning Ensemble for Predicting Diabetic Macular Edema Onset Using Ultra-Wide Field Color Fundus Image. Image-Based Prediction of Retinal Disease Progression: MICCAI Challenges, DIAMOND 2024 and MARIO 2024, Held in Conjunction with MICCAI 2024, Marrakesh, Morocco, October 10, 2024, Proceedings. 2025;15503:64. Arcadu F, Benmansour F, Maunz A, Michon J, Haskova Z, McClintock D, et al. Deep learning predicts OCT measures of diabetic macular thickening from color fundus photographs. Invest Ophthalmol Vis Sci. 2019;60:852–7. Hasan MM, Phu J, Wang H, Sowmya A, Meijering E, Kalloniatis M. Predicting visual field global and local parameters from OCT measurements using explainable machine learning. Sci Rep. 2025;15:5685. Yang W, Zhu Y, Shen J, Tang Y, Pan C, He H, et al. RURANET++: An Unsupervised Learning Method for Diabetic Macular Edema Based on SCSE Attention Mechanisms and Dynamic Multi-Projection Head Clustering. arXiv preprint arXiv:250220224. 2025; Meng Z, Chen Y, Li H, Zhang Y, Yao X, Meng Y, et al. Machine learning and optical coherence tomography-derived radiomics analysis to predict persistent diabetic macular edema in patients undergoing anti-VEGF intravitreal therapy. J Transl Med. 2024;22:358. Tan T-E, Teo ZL, Vujosevic S, Ting DSW. AI-Based Diabetic Macular Edema Screening for Improved Care in Low-Resource Settings. NEJM AI. Massachusetts Medical Society; 2024. p. AIe2400698. Kusuma S, Udayan JD. Machine learning and deep learning methods in heart disease (HD) research. International Journal of Pure and Applied Mathematics. 2018;119:1483–96. Aloysius N, Geetha M. A review on deep convolutional neural networks. 2017 international conference on communication and signal processing (ICCSP). IEEE; 2017. p. 588–92. He Q, Jiang H, Fang D, Yang D, Nguyen TX, Ran A, et al. A graph neural network-based multispectral-view learning model for diabetic macular ischemia detection from color fundus photographs. arXiv preprint arXiv:250217886. 2025; Li H-Y, Wang D-X, Dong L, Wei W-B. Deep learning algorithms for detection of diabetic macular edema in OCT images: A systematic review and meta-analysis. Eur J Ophthalmol. 2023;33:278–90. Vinuesa R, Azizpour H, Leite I, Balaam M, Dignum V, Domisch S, et al. The role of artificial intelligence in achieving the Sustainable Development Goals. Nat Commun. 2020;11:233. Nedungadi P, Salethoor SN, Puthiyedath R, Nair VK, Kessler C, Raman R. Ayurveda research: Emerging trends and mapping to sustainable development goals. J Ayurveda Integr Med. 2023;14:100809. Kirankumar M, Saikiran P, Sabarivasan N, Kumaran M, Thangavel SK, Somasundaram K, et al. SoulNourish: A Comprehensive Machine learning Based Recommendation Yoga Prediction System Based on Health and Lifestyle Data. 2024 4th International Conference on Sustainable Expert Systems (ICSES). IEEE; 2024. p. 1828–35. Lam C, Wong YL, Tang Z, Hu X, Nguyen TX, Yang D, et al. Performance of artificial intelligence in detecting diabetic macular edema from fundus photography and optical coherence tomography images: a systematic review and meta-analysis. Diabetes Care. 2024;47:304–19. Midena E, Toto L, Frizziero L, Covello G, Torresin T, Midena G, et al. Validation of an automated artificial intelligence algorithm for the quantification of major OCT parameters in diabetic macular edema. J Clin Med. 2023;12:2134. Ye X, Gao K, He S, Zhong X, Shen Y, Wang Y, et al. Artificial intelligence-based quantification of central macular fluid volume and VA prediction for diabetic macular edema using OCT images. Ophthalmol Ther. 2023;12:2441–52. Abramovich O, Pizem H, Van Eijgen J, Oren I, Melamed J, Stalmans I, et al. FundusQ-Net: A regression quality assessment deep learning algorithm for fundus images quality grading. Comput Methods Programs Biomed. 2023;239:107522. Krithika Alias AnbuDevi M, Suganthi K. Review of semantic segmentation of medical images using modified architectures of UNET. Diagnostics. 2022;12:3064. Procino E. CVAT meets transformers: accelerating semantic segmentation labeling in industrial applications. Tamil Priya D, Divya Udayan J. Transfer learning techniques for emotion classification on visual features of images in the deep learning network. Int J Speech Technol. 2020;23:361–72. Kusuma S, Udayan JD. Machine learning and deep learning methods in heart disease (HD) research. International Journal of Pure and Applied Mathematics. 2018;119:1483–96. Neena A, Geetha M. Image classification using an ensemble-based deep CNN. Recent Findings in Intelligent Computing Techniques: Proceedings of the 5th ICACNI 2017, Volume 3. Springer; 2018. p. 445–56. Vijay Gopal D V, Gopakumar G. A deep learning approach to image splicing using depth map. Advances in Distributed Computing and Machine Learning: Proceedings of ICADCML 2022. Springer; 2022. p. 401–11. Meng Z, Chen Y, Li H, Zhang Y, Yao X, Meng Y, et al. Machine learning and optical coherence tomography-derived radiomics analysis to predict persistent diabetic macular edema in patients undergoing anti-VEGF intravitreal therapy. J Transl Med. 2024;22:358. Hasan MM, Phu J, Wang H, Sowmya A, Meijering E, Kalloniatis M. Predicting visual field global and local parameters from OCT measurements using explainable machine learning. Sci Rep. 2025;15:5685. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 19 Apr, 2026 Read the published version in Discover Sustainability → Version 1 posted Editorial decision: Revision requested 21 Jul, 2025 Reviews received at journal 15 Jul, 2025 Reviews received at journal 09 Jul, 2025 Reviews received at journal 30 Jun, 2025 Reviewers agreed at journal 30 Jun, 2025 Reviewers agreed at journal 30 Jun, 2025 Reviewers agreed at journal 29 Jun, 2025 Reviews received at journal 26 Jun, 2025 Reviewers agreed at journal 25 Jun, 2025 Reviewers invited by journal 25 Jun, 2025 Editor invited by journal 18 Jun, 2025 Editor assigned by journal 15 Jun, 2025 Submission checks completed at journal 13 Jun, 2025 First submitted to journal 13 Jun, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6795425","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":478012208,"identity":"dba8c669-1fbe-472b-bb7d-eefaf663e93d","order_by":0,"name":"C A Aparna","email":"","orcid":"","institution":"Amrita Vishwa Vidyapeetham","correspondingAuthor":false,"prefix":"","firstName":"C","middleName":"A","lastName":"Aparna","suffix":""},{"id":478012209,"identity":"83fcb412-517b-49e3-9e50-c9d085262cd9","order_by":1,"name":"B R Manju","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA6ElEQVRIiWNgGAWjYJACA8YGBjkoWwKJJKDFGKZFgigtDEAtiQ3I1uDVott+9kAx74576f3TDjB+urnHoo6/gfngbR48WszO5CUY854pzp1xO4FZOueZhITEAbZka7xaDuQYGPO2JeQ23E5gkM45APTLAR4zabxazr8Ba0mXB9ryG6RF/gD/N/xabkBsSTC4ncAGtsXgAA8bAS1vDAznnkkw3Hg7sc0aqEVy42E2Y8s5eB2WY2bwdkeCvNzt5MO3cw7U8csdb3544w0eLUDAZgChgbEDBsz4lYOVPCCsZhSMglEwCkY0AAA900n4PVGM4gAAAABJRU5ErkJggg==","orcid":"","institution":"Amrita Vishwa Vidyapeetham","correspondingAuthor":true,"prefix":"","firstName":"B","middleName":"R","lastName":"Manju","suffix":""}],"badges":[],"createdAt":"2025-06-01 11:38:18","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6795425/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6795425/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s43621-026-03218-5","type":"published","date":"2026-04-19T15:58:31+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":85743656,"identity":"25290b49-17b0-4afe-809a-9ab076044cbe","added_by":"auto","created_at":"2025-07-01 09:11:29","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":201728,"visible":true,"origin":"","legend":"\u003cp\u003eFramework for Macular Thickness Prediction Using Deep Learning and Clinical Data Integration\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6795425/v1/11ac8f1ca3f14f69ca507ea3.png"},{"id":85743654,"identity":"ccddb2c5-81cc-41e0-bcdf-49a145bad9a8","added_by":"auto","created_at":"2025-07-01 09:11:29","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":85772,"visible":true,"origin":"","legend":"\u003cp\u003eA random fundus image and its predicted mask\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6795425/v1/67b17eb14d76e68c34b98410.png"},{"id":85744831,"identity":"3d166252-bc14-418a-a95d-14111eda5f0f","added_by":"auto","created_at":"2025-07-01 09:19:29","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":339201,"visible":true,"origin":"","legend":"\u003cp\u003eMethodology for Macular Thickness Prediction Using Fundus Images and Clinical Data.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6795425/v1/f3f4484e8376d08edb65d00e.png"},{"id":85743657,"identity":"54a88e3c-f12f-4459-a716-7a8e1038b5dc","added_by":"auto","created_at":"2025-07-01 09:11:29","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":117064,"visible":true,"origin":"","legend":"\u003cp\u003eRegression results for Inception V3\u003c/p\u003e\n\u003cp\u003e(a) Training vs Validation MSE Curve for InceptionV3-Based Macular Thickness Regression.\u003c/p\u003e\n\u003cp\u003e(b) Predicted vs True Macular Thickness for InceptionV3 with R\u003csup\u003e2\u003c/sup\u003e = 0.586\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6795425/v1/321f2059a0053b9d7d32a87e.png"},{"id":85746160,"identity":"bfc616be-018e-46c3-abc0-3db31942a6be","added_by":"auto","created_at":"2025-07-01 09:27:29","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":106607,"visible":true,"origin":"","legend":"\u003cp\u003eRegression results of EfficientNetB0\u003c/p\u003e\n\u003cp\u003e(a) Training vs Validation MSE Curve for EfficientNetB0-Based Macular Thickness Regression.\u003c/p\u003e\n\u003cp\u003e(b) Predicted vs True Macular Thickness for EfficientNetB0 with R\u003csup\u003e2\u003c/sup\u003e = 0.7752\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6795425/v1/fcfed049563cc81e8aa64974.png"},{"id":85743659,"identity":"20d909a0-294c-44d9-9ab1-15b5dd5628b6","added_by":"auto","created_at":"2025-07-01 09:11:29","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":126867,"visible":true,"origin":"","legend":"\u003cp\u003eRegression results of DenseNet121\u003c/p\u003e\n\u003cp\u003e(a) Training vs Validation MSE Curve for DenseNet121-Based Macular Thickness Regression.\u003c/p\u003e\n\u003cp\u003e(b) Predicted vs True Macular Thickness for DenseNet121 with R\u003csup\u003e2\u003c/sup\u003e = 0.7254\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6795425/v1/cd08c4f96c770d92f4bb7720.png"},{"id":107351640,"identity":"447c9ea3-10b5-455d-b281-034885e297d0","added_by":"auto","created_at":"2026-04-20 16:11:44","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1400897,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6795425/v1/1e822da4-aa32-4f85-a103-030c31019d78.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Towards Sustainable Retinal Diagnostics A Deep Learning Alternative to OCT for Macular Thickness Estimation","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eMacular edema is a pathological condition characterized by the accumulation of fluid in the macula\u0026mdash;the central region of the retina responsible for sharp and detailed vision. This swelling distorts vision and can lead to significant visual impairment if left untreated. Macular edema can arise from several causes, including age-related macular degeneration (AMD), retinal vein occlusion (RVO), and most notably, diabetic retinopathy, where it is termed Diabetic Macular Edema (DME).Among these, DME is the most common form of macular edema, and it has become an increasingly serious public health concern, especially in countries with a high prevalence of diabetes.[1] In India, the incidence of diabetes is growing at an alarming rate, with projections estimating over 100\u0026nbsp;million diabetic individuals by 2030. This surge has led to a corresponding increase in diabetes-related eye complications. Studies suggest that approximately 7\u0026ndash;14% of diabetic patients in India develop DME, making early detection and monitoring a critical priority for preventing vision loss.[2]\u003c/p\u003e \u003cp\u003eIn fundus imaging, microaneurysms and hard exudates are key patterns indicative of diabetic macular edema (DME). Microaneurysms are small, round outpouchings of capillaries in the retina, appearing as tiny red dots in fundus images. They represent the earliest clinically detectable signs of diabetic retinopathy and are indicative of capillary wall weakness. These microaneurysms can leak fluid, contributing to retinal swelling characteristic of DME. Their presence is crucial for early diagnosis and monitoring of disease progression.[1,3] Hard exudates are lipid-rich deposits that manifest as yellowish-white spots with sharp margins in the retina. They result from the leakage of lipoproteins and other plasma constituents from compromised retinal vessels. In fundus images, hard exudates often form a circinate pattern around areas of leakage, particularly near microaneurysms. Their accumulation, especially within the macular region, is a hallmark of DME and correlates with the severity of macular thickening and potential vision loss.[4]\u003c/p\u003e \u003cp\u003eRetinal thickness is a crucial biomarker assessed in ophthalmology, particularly in the diagnosis and monitoring of diabetic macular edema (DME). In healthy individuals, the central macular thickness typically ranges between 250 to 280 micrometres (\u0026micro;m) \u0026mdash; where 1 micrometre (\u0026micro;m) equals 1000 nanometres (nm). However, in patients with DME, the retinal thickness often increases significantly, frequently exceeding 300 micrometres due to fluid accumulation within the retinal layers. Monitoring changes in retinal thickness with high precision is essential for evaluating disease progression and treatment response. [5]\u003c/p\u003e \u003cp\u003eThe gold standard for diagnosing and monitoring DME is Optical Coherence Tomography (OCT), a non-invasive imaging technique that captures high-resolution cross-sectional images of the retina and provides accurate measurements of macular thickness. However, OCT devices are expensive, require skilled operators, and are often confined to specialized medical care centres, limiting their availability in rural and resource-constrained settings across India and other developing countries. In contrast, retinal fundus photography is far more accessible, cost-effective, and already used widely in diabetic screening programs.[2,6] While fundus images are inherently two-dimensional and do not offer depth information like OCT, advances in artificial intelligence (AI) and deep learning have created opportunities to infer 3D structural information from 2D images.[4,7]\u003c/p\u003e \u003cp\u003eThis study proposes a cost-effective, AI-driven alternative to OCT for estimating macular thickness\u0026mdash;a key biomarker for DME diagnosis and progression monitoring. Our pipeline is built to mimic OCT-derived macular thickness maps using only fundus photographs and clinical metadata. The process involves several key stages to enable macular thickness estimation from fundus images. First, manual mask preparation was carried out for the macular region using the CVAT annotation tool, allowing for precise labelling of the region of interest. Next, semantic segmentation of the macula was performed using a U-Net-based deep learning model, which effectively isolated the macular area from the rest of the retinal image. Finally, the study used a modified EfficientNetB0 model that incorporates both image and clinical data to predict macular thickness.[8] This was achieved by adopting a multimodal architecture in which high-level features extracted from fundus images using EfficientNetB0 are concatenated with relevant tabular clinical features, such as visual acuity (VA) and diabetes duration. This fusion of image and clinical data enhances the model\u0026rsquo;s ability to learn complex patterns and improves the overall regression performance. In this study, modified versions of DenseNet121 and InceptionV3 were also implemented for comparison; however, the modified EfficientNetB0 model demonstrated comparatively better performance, making it the most effective choice for the proposed task.[8,9]\u003c/p\u003e \u003cp\u003eBy estimating macular thickness directly from fundus images combined with clinical data, this study aims to assist clinicians in identifying early signs of Diabetic Macular Edema (DME) without relying on OCT imaging. The proposed model leverages a multimodal deep learning framework that integrates high-level image features extracted by EfficientNetB0 with patient-specific clinical parameters such as visual acuity and diabetes duration.[10,11] This approach can potentially democratize access to retinal diagnostics, particularly in under-resourced regions, and contribute to scalable diabetic eye care in India and beyond. By offering a cost-effective, AI-driven alternative to OCT for macular thickness estimation\u0026mdash;an essential parameter in the early detection and monitoring of DME\u0026mdash;the study addresses the diagnostic gap in resource-constrained settings. The framework's reliance on widely available retinal fundus images, enhanced by clinical context, promotes timely intervention, improved visual outcomes, and reduced burden of diabetic eye disease, especially in low- and middle-income countries.[12,13]\u003c/p\u003e \u003cp\u003eBeyond its technical contributions, the proposed framework aligns strongly with key goals of sustainable development. By leveraging artificial intelligence to replicate the diagnostic utility of Optical Coherence Tomography (OCT) using affordable and widely available retinal fundus images, the model supports \u003cb\u003eSDG 3 (Good Health and Well-being)\u003c/b\u003e by improving access to early diabetic eye care in underserved regions. Furthermore, the approach exemplifies \u003cb\u003eSDG 9 (Industry, Innovation, and Infrastructure)\u003c/b\u003e through its application of cutting-edge AI methods to strengthen healthcare infrastructure in low-resource settings without the need for high-cost diagnostic equipment. Importantly, this solution also contributes to \u003cb\u003eSDG 10 (Reduced Inequalities)\u003c/b\u003e by bridging the gap in diagnostic access between urban tertiary hospitals and rural or economically disadvantaged communities. The study\u0026rsquo;s focus on cost-effectiveness, scalability, and clinical utility positions it as a practical innovation for advancing equitable and sustainable eye health systems globally.[14\u0026ndash;16]\u003c/p\u003e"},{"header":"2. Literature Review","content":"\u003cp\u003eRecent advancements in artificial intelligence (AI) have significantly impacted the early detection and management of diabetic macular edema (DME), particularly in resource-limited settings where access to optical coherence tomography (OCT) is constrained. Several studies have explored the potential of deep learning models to predict macular thickness and detect DME using fundus photographs and clinical data. \u003cb\u003eAk\u0026ccedil;a et al. (2024)\u003c/b\u003e conducted a comparative study using vision transformers for automated classification of choroidal neovascularization, DME, and drusen from retinal OCT images. The study highlighted the effectiveness of vision transformers in accurately classifying retinal diseases, including DME. [3] \u003cb\u003eChavan and Choubey (2024)\u003c/b\u003e developed a self-supervised category selective attention classifier network for DME classification. The model leveraged self-supervised learning to enhance classification performance, demonstrating its potential in scenarios with limited labelled data. [1]\u003c/p\u003e \u003cp\u003e \u003cb\u003eQin et al. (2024)\u003c/b\u003e developed an ensemble deep learning model to predict the onset of centre-involved DME within a year using ultra-wide field colour fundus images. The model achieved an AUC of 0.7017 and an F1 score of 0.6512, demonstrating its potential in early DME prediction.[4] \u003cb\u003eLam et al. (2024)\u003c/b\u003e conducted a systematic review and meta-analysis evaluating AI performance in detecting DME from fundus photography (FP) and OCT images. Analysing 53 studies, they reported pooled sensitivity and specificity of 92.6% and 91.1% for FP-based algorithms, and 95.2% and 92.5% for OCT-based algorithms, respectively, highlighting AI's robust diagnostic capabilities. ​[17]\u003c/p\u003e \u003cp\u003e \u003cb\u003eMidena et al. (2023)\u003c/b\u003e validated an automated AI algorithm for quantifying major OCT parameters in DME. Their model demonstrated high accuracy in measuring central subfield thickness and macular volume, aligning closely with manual assessments, thus supporting its potential for clinical application.[18] \u003cb\u003eYe et al. (2023)\u003c/b\u003e developed an AI-based method to quantify central macular fluid volume and predict visual acuity in DME patients using OCT images. The model showed strong correlations between predicted and actual visual acuity, suggesting its utility in personalized treatment planning.[19] \u003cb\u003eLi et al. (2023)\u003c/b\u003e performed a systematic review and meta-analysis on deep learning algorithms for DME detection in OCT images. They concluded that AI models achieved high diagnostic accuracy, with pooled sensitivity and specificity of 94% and 92%, respectively, reinforcing AI's effectiveness in OCT image analysis.[13] \u003cb\u003eAbramovich et al. (2022)\u003c/b\u003e introduced FundusQ-Net, a deep learning algorithm for assessing fundus image quality. Trained on a large dataset, the model achieved a mean absolute error of 0.61 in quality grading, facilitating the selection of high-quality images for reliable AI-based DME detection. [20]\u003c/p\u003e \u003cp\u003e \u003cb\u003eSpringer et al. (2024)\u003c/b\u003e developed an OCT-omics prediction model using machine learning to assess anti-VEGF treatment response in DME patients. Their model demonstrated high discriminative abilities, with the backpropagation neural network classifier achieving an AUC of 0.982, indicating its potential in guiding personalized therapy.[8] \u003cb\u003eNEJM AI (2024)\u003c/b\u003e discussed the development of an AI-based DME screening tool designed for low-resource settings. The model incorporated techniques for privacy preservation and uncertainty estimation, aiming to improve screening accessibility and reliability in underserved areas.[9] \u003cb\u003eHe et al. (2025)\u003c/b\u003e proposed a graph neural network-based multispectral-view learning model for detecting diabetic macular ischemia (DMI) from colour fundus photographs. Their approach achieved an accuracy of 84.7% and an AUROC of 0.900, demonstrating the potential of AI in identifying DMI, a condition closely related to DME.[12] \u003cb\u003eYang et al. (2025)\u003c/b\u003e introduced RURANET++, an unsupervised learning framework for automated DME diagnosis. Incorporating SCSE attention mechanisms and dynamic multi-projection head clustering, the model achieved an accuracy of 84.11% and an F1-score of 83.90%, highlighting its effectiveness without the need for extensive libelled data.[7] \u003cb\u003eSong et al. (2023)\u003c/b\u003e introduced an AI method based on multi-feature fusion for automatic macular edema classification on spectral-domain OCT images. By combining traditional and deep features, the model achieved improved accuracy in classifying various types of macular edema, including DME. [13]​​\u003c/p\u003e \u003cp\u003e \u003cb\u003eVinuesa et al.\u003c/b\u003e (2020) in [14]conducted a comprehensive expert-driven study to assess the influence of artificial intelligence (AI) on achieving the 169 individual targets within the 17 Sustainable Development Goals (SDGs). Their methodology combined a structured literature review with a consensus-based expert elicitation process involving scholars from diverse disciplines, including engineering, natural sciences, and social sciences. The evaluation focused on determining whether AI served as an enabler or inhibitor for each SDG target. The authors relied on peer-reviewed literature describing real-world AI applications, validated experimental studies, and reports from recognized institutions such as the United Nations and governmental agencies. Notably, speculative opinions, non-peer-reviewed claims, and media reports were excluded to maintain the integrity and quality of the evidence base. This rigorous approach highlighted how AI, when responsibly deployed, can significantly advance key sustainability goals\u0026mdash;particularly those related to health (SDG 3), innovation (SDG 9), and equity (SDG 10).\u003c/p\u003e"},{"header":"3. Methodology","content":"\u003cp\u003eThis study presents a cost-effective, AI-driven approach for estimating macular thickness from retinal fundus images and clinical data, with the goal of aiding in the early detection and monitoring of Diabetic Macular Edema (DME), particularly in low-resource settings where Optical Coherence Tomography (OCT) is not readily available. The methodology involves three primary stages: dataset preparation, macular region segmentation, and regression-based thickness prediction.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Dataset Preparation\u003c/h2\u003e \u003cp\u003eThe dataset was curated from the ophthalmology department at Amrita Institute of Medical Sciences (AIMS), Kochi, Kerala, India, and comprises 176 colour fundus images from patients clinically diagnosed with Diabetic Macular Edema. Along with the fundus images, corresponding clinical information was collected, including visual acuity and duration of diabetes (in years). Ground truth macular thickness values were obtained from OCT scans and served as regression targets during model training and evaluation. To enhance model generalizability and prevent overfitting, the original 176 fundus images were augmented to create a larger dataset of approximately 3000 images. This expanded dataset forms the foundation for developing and validating the proposed AI model, enabling the extraction of relevant features and the accurate prediction of macular thickness.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Macular Region Segmentation Using U-Net with Mask Preparation via CVAT Tool\u003c/h2\u003e \u003cp\u003eTo localize the macular area from retinal fundus images, a U-Net-based segmentation model was employed. U-Net is a widely used convolutional neural network architecture in medical image analysis due to its ability to capture fine-grained spatial features through skip connections and a symmetric encoder-decoder structure. [21] Manual annotations for the macular region were generated using the CVAT (Computer Vision Annotation Tool), an open-source tool that facilitates precise annotation of medical images with an intuitive interface and flexible support for polygonal masks. These annotations served as ground truth for training the segmentation model. [22]\u003c/p\u003e \u003cp\u003eA dataset of 70 grayscale fundus images and their corresponding masks were pre-processed by resizing them to 256\u0026times;256 pixels and normalizing pixel values to the range [0, 1]. This helped standardize the input for the model and reduce computational complexity. The U-Net architecture consisted of four encoding blocks with convolutional layers followed by max pooling, a bottleneck with high-dimensional feature extraction, and four decoding blocks that used transposed convolutions for up sampling. Skip connections between the encoder and decoder facilitated the recovery of spatial details lost during down sampling.\u003c/p\u003e \u003cp\u003eThe model was trained using a custom loss function that combined Binary Cross-Entropy (BCE) with a log-transformed Dice loss to balance pixel-wise accuracy and segmentation overlap. Training was carried out for 50 epochs with a batch size of 4, using 10% of the data for validation. A model checkpoint mechanism was employed to save the best-performing model based on validation loss. This segmentation model provides binary masks of the macular region, which are subsequently used for feature extraction and thickness estimation in the regression stage of the pipeline. The U-Net model was trained for macular region segmentation on 70 grayscale fundus images paired with manually annotated masks. Training was carried out over 50 epochs using a batch size of 4, with a 90\u0026thinsp;\u0026minus;\u0026thinsp;10 split for training and validation. The input images and masks were resized to 256\u0026times;256 pixels for computational efficiency.[21]\u003c/p\u003e \u003cp\u003eTo optimize performance, a custom loss function combining Binary Cross-Entropy and a logarithmically adjusted Dice loss was used. This approach encouraged both pixel-level accuracy and overall shape conformity in segmentation.\u003c/p\u003e \u003cp\u003eThe model achieved a \u003cb\u003etraining accuracy of 93.54%\u003c/b\u003e and a \u003cb\u003evalidation accuracy of 90.61%\u003c/b\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Hybrid Deep Learning Framework for Macular Thickness Prediction\u003c/h2\u003e \u003cp\u003eTo evaluate the effectiveness of different convolutional neural network (CNN) architectures in predicting macular thickness, the study compared three deep regression models utilizing InceptionV3, EfficientNetB0, and DenseNet121 as the core image feature extractors. These models were integrated into a multimodal learning framework by combining visual information from fundus images with tabular clinical data\u0026mdash;specifically Visual Acuity (VA) and Duration of Diabetes\u0026mdash;for a more comprehensive prediction approach. By using the same data preprocessing, model structure, and training protocol, we ensured a fair comparison across architectures. [23]\u003c/p\u003e \u003cp\u003eAll input fundus images were uniformly resized to 128\u0026times;128 pixels and processed using histogram equalization in the YUV colour space to enhance local contrast and visual clarity. These enhanced images were then normalized using model-specific preprocessing functions. Tabular features (VA and Diabetes Duration) were standardized using StandardScaler, ensuring they contributed comparably to the learning process. The target variable\u0026mdash;macular thickness\u0026mdash;was also standardized to stabilize training and improve convergence in the regression models. [23,24]\u003c/p\u003e \u003cp\u003eTo improve generalization and increase the dataset size, data augmentation applied using Keras\u0026rsquo;s ImageDataGenerator. This included random transformations such as small rotations, zooms, shifts, and horizontal flips. Through this augmentation process, the dataset was expanded to approximately 3,000 samples, reducing the risk of overfitting and enhancing model robustness. Each of the three model variants followed a consistent design strategy. A pre-trained CNN backbone (InceptionV3, EfficientNetB0, or DenseNet121) was loaded without its classification head and initialized with ImageNet weights. These backbones were initially frozen to retain learned low-level features. Extracted visual features were pooled using GlobalAveragePooling2D, regularized using BatchNormalization and Dropout, then concatenated with the tabular input.[25,26] This merged representation was passed through two fully connected layers (128 and 64 units, both using ReLU activation), followed by a final output layer predicting macular thickness as a single scalar value.\u003c/p\u003e \u003cp\u003eThe training setup was standardized across models using the Mean Squared Error (MSE) loss function and the Adam optimizer with a learning rate of 1e-4. Model performance was assessed based on both training and validation MSE curves. Final evaluation was conducted on a held-out test set using Test MSE and the R\u0026sup2; score to quantify predictive strength. Additionally, scatter plots of predicted versus actual macular thickness values were generated to visually assess regression accuracy and model fit across the range of thickness values.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. Analysis and results","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003e4.1. InceptionV3 Regression Model Performance for Macular Thickness Prediction\u003c/h2\u003e\n \u003cp\u003eThe Inception V3 model was trained for 30 epochs with a learning rate of 1e-4, and the training process showed consistent improvement in performance. The training loss started at 1.9761 and steadily decreased to 0.2156 by the final epoch, while the mean squared error (MSE) followed a similar downward trend. The validation loss and MSE also decreased throughout the epochs, suggesting that the model was effectively learning to predict macular thickness. After completing the training, the model was evaluated on the test set, yielding a test MSE of 0.4093 and a test R\u0026sup2; score of 0.5859. These results indicate a moderate correlation between the predicted and true values, showing that the model performed reasonably well in predicting macular thickness.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n \u003ch2\u003e4.2. EfficientNetB0 Regression Model Performance for Macular Thickness Prediction\u003c/h2\u003e\n \u003cp\u003eThe EfficientNetB0 model was trained over 30 epochs to predict macular thickness. In the first epoch, the model showed a loss of 1.9223 and mean squared error (MSE) of 1.9223, with a validation loss of 0.9261 and a validation MSE of 0.9261. Throughout the epochs, the model improved significantly, with both loss and MSE steadily decreasing, demonstrating its learning process. By epoch 10, the model achieved a validation MSE of 0.4641. As training progressed, the model became more accurate, reaching a validation MSE of 0.2517 in the final epoch. The test evaluation confirmed this progress with a test MSE of 0.2518, and the model\u0026apos;s R\u0026sup2; score was 0.7752, indicating a strong fit to the data and substantial predictive capability. The model\u0026apos;s performance shows consistent improvement throughout the training period, reflecting its ability to generalize well to unseen data.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n \u003ch2\u003e4.3. DenseNet121 Regression Model Performance for Macular Thickness Prediction\u003c/h2\u003e\n \u003cp\u003eDenseNet121, a deep convolutional neural network (CNN) architecture is also used for the regression task of predicting macular thickness from fundus images and associated clinical features. The model was trained over 30 epochs, utilizing a batch size of 32. The training process started with a relatively higher loss and mean squared error (MSE) values, which progressively improved as the model trained. By epoch 30, the model achieved a final training loss of 0.2887 and a final validation loss of 0.2565, indicating a significant reduction in error. This progressive reduction in loss and MSE is indicative of the model\u0026apos;s ability to effectively learn the relationship between the input features and the macular thickness, demonstrating the potential of DenseNet121 for medical image analysis tasks. Additionally, the learning rate was kept constant at 1.0000\u0026times;10\u0026thinsp;\u0026minus;\u0026thinsp;4 throughout the training, suggesting that the model\u0026rsquo;s convergence was robust despite the steady learning rate.\u003c/p\u003e\n \u003cp\u003eThe model\u0026apos;s performance was further evaluated on a test set, where it achieved a test MSE of 0.30198. Additionally, the R\u0026sup2; score of 0.7254 indicates that the model was able to explain a significant portion of the variance in macular thickness prediction. These results highlight the capability of DenseNet121 in accurately predicting macular thickness from both fundus images and clinical features, with substantial potential for clinical applications.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"5. Comparative Performance Analysis of Deep Learning Models for Macular Thickness Prediction","content":"\u003cp\u003eTo evaluate the effectiveness of different deep learning architectures in predicting macular thickness, study compared the performance of InceptionV3, EfficientNetB0, and DenseNet121. The models were assessed based on their Mean Squared Error (MSE) and R\u0026sup2; score on the test set. A summary of the results is presented in the table below.\u003c/p\u003e\n\u003cp\u003eTable 1. Comparative analysis of Deep Learning Models.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 101px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTest MSE\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 136px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eR\u0026sup2; Score\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 101px;\"\u003e\n \u003cp\u003eInceptionV3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e0.4093\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 136px;\"\u003e\n \u003cp\u003e0.5859\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 101px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eEfficientNetB0\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.2518\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 136px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.7752\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 101px;\"\u003e\n \u003cp\u003eDenseNet121\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e0.3020\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 136px;\"\u003e\n \u003cp\u003e0.7254\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAmong the three deep learning models evaluated\u0026mdash;InceptionV3, EfficientNetB0, and DenseNet121\u0026mdash;for predicting macular thickness from fundus images and clinical features, EfficientNetB0 demonstrated the most favourable performance. It achieved the lowest test Mean Squared Error (MSE) of 0.2518 and the highest R\u0026sup2; score of 0.7752, indicating a stronger predictive accuracy and better generalization to unseen data. In contrast, DenseNet121 showed moderate performance with a test MSE of 0.3020 and an R\u0026sup2; score of 0.7254, while InceptionV3 yielded the highest test MSE of 0.4093 and the lowest R\u0026sup2; score of 0.5859. These results suggest that EfficientNetB0 is more effective in capturing relevant features from fundus images for accurate regression-based macular thickness estimation compared to the other two architectures.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eDiabetic Macular Edema (DME) is a significant complication of diabetes mellitus, characterized by the accumulation of fluid in the macula\u0026mdash;the central part of the retina responsible for sharp vision. This condition arises due to the leakage of fluid from damaged blood vessels in the retina, leading to retinal thickening and vision impairment. DME is a leading cause of vision loss among diabetic patients and poses a growing public health concern, especially in countries like India, where the prevalence of diabetes is rapidly increasing. Early detection and monitoring of DME are crucial to prevent irreversible vision loss. However, the gold standard for diagnosing and monitoring DME, Optical Coherence Tomography (OCT), is expensive and often inaccessible in resource-limited settings. [27] To address this challenge, the study proposes an artificial intelligence (AI)-driven approach to estimate macular thickness\u0026mdash;a key biomarker for DME\u0026mdash;using retinal fundus images and clinical data. The dataset comprises 176 colour fundus images from patients diagnosed with DME at the Amrita Institute of Medical Sciences in Kochi, Kerala, India. These images were augmented to approximately 3,000 samples to enhance model robustness.\u003c/p\u003e\n\u003cp\u003eManual annotations of the macular region were performed using the Computer Vision Annotation Tool (CVAT), and a U-Net-based deep learning model was employed for macular segmentation. For macular thickness prediction, three convolutional neural network architectures\u0026mdash;InceptionV3, DenseNet121, and EfficientNetB0\u0026mdash;were evaluated within a multimodal framework that integrates image features with clinical parameters such as visual acuity and duration of diabetes. All models were trained using standardized preprocessing techniques, including image resizing, histogram equalization, and normalization, as well as data augmentation strategies to improve generalization. Among the evaluated models, EfficientNetB0 demonstrated the most favourable performance, achieving a test Mean Squared Error (MSE) of 0.2518 and an R\u0026sup2; score of 0.7752, indicating strong predictive accuracy and generalization capability. DenseNet121 and InceptionV3 showed comparatively lower performance, with test MSEs of 0.3020 and 0.4093, and R\u0026sup2; scores of 0.7254 and 0.5859, respectively. These findings suggest that the proposed AI-based method, particularly the EfficientNetB0 model, can serve as a viable alternative to OCT for macular thickness estimation, facilitating early detection and monitoring of DME in resource-constrained settings.\u003c/p\u003e\n\u003cp\u003eThis research not only presents a technologically sound solution for macular thickness estimation but also contributes meaningfully to sustainable healthcare goals. By reducing dependence on costly OCT infrastructure, the proposed AI-driven approach aligns with the principles of \u003cstrong\u003eSDG 3 (Good Health and Well-being)\u003c/strong\u003e through enhanced early detection in underserved populations. It supports \u003cstrong\u003eSDG 9 (Industry, Innovation, and Infrastructure)\u003c/strong\u003e by promoting accessible medical innovation and reinforces \u003cstrong\u003eSDG 10 (Reduced Inequalities)\u003c/strong\u003e by bridging diagnostic gaps between urban and rural healthcare systems. As such, this work represents a step toward more inclusive, affordable, and scalable eye care solutions worldwide.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFuture Work\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA key future direction involves enhancing the transparency and interpretability of the proposed deep learning model through the integration of explainable AI techniques. Tools such as Grad-CAM (Gradient-weighted Class Activation Mapping), LIME (Local Interpretable Model-agnostic Explanations), can be employed to visualize which regions of the fundus image or which clinical features most significantly influence the model\u0026apos;s prediction of macular thickness. This interpretability is crucial for clinical acceptance, as it enables ophthalmologists to verify whether the model is focusing on relevant pathological features like microaneurysms and hard exudates associated with diabetic macular edema (DME). [28] Future work will also focus on expanding the dataset to include images from diverse populations to enhance generalizability, deploying the model in real-time clinical environments for field validation, and integrating explainable AI (XAI) techniques to improve transparency and clinician trust. In addition, exploring deployment on mobile or edge devices will further support sustainable, on-the-ground screening solutions in rural and remote regions.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor Contributions Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAparna C A conceptualized the study, designed the methodology, collected and curated the clinical data and fundus images. Dr Manju B R analysed the result and supervised.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflict of interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors have no conflicts of interest to declare that are relevant to the content of this article.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics Approval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Institutional Ethics Committee of Amrita Institute of Medical Sciences, Kochi, India. All procedures performed and data collected were in accordance with the ethical standards of the institution.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe research protocol titled \u0026quot;Towards Sustainable Retinal Diagnostics A Deep Learning Alternative to OCT for Macular Thickness Estimation \u0026quot; was reviewed and approved by the Ethics Committee of Amrita School of Medicine on November 8, 2024 (Approval No. ECASM-AIMS-2024-506, Reg. No. EC/NEW/INST/2023/KL/0379), in accordance with institutional guidelines and relevant national ethical regulations. The committee found no ethical issues, and clearance was granted. All patient data used in this study were anonymized prior to analysis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to Participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eInformed consent was obtained from all participants or their legal guardians prior to data collection and image acquisition.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData and Code Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets generated and analysed during the current study are not publicly available due to institutional privacy policies but are available from the corresponding author on reasonable request. The code used for training and evaluation of models is available from the corresponding author.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNo funding was received for conducting this study\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical trial number\u003c/strong\u003e: Not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to Publish declaration:\u0026nbsp;\u003c/strong\u003eNot applicable\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eChavan S, Choubey N. Self-supervised category selective attention classifier network for diabetic macular edema classification. Acta Diabetol. 2024;61:879\u0026ndash;96. \u003c/li\u003e\n\u003cli\u003eGan F, Wu F-P, Zhong Y-L. Artificial intelligence method based on multi-feature fusion for automatic macular edema (ME) classification on spectral-domain optical coherence tomography (SD-OCT) images. Front Neurosci. 2023;17:1097291. \u003c/li\u003e\n\u003cli\u003eAk\u0026ccedil;a S, Garip Z, Ekinci E, Atban F. Automated classification of choroidal neovascularization, diabetic macular edema, and drusen from retinal OCT images using vision transformers: a comparative study. Lasers Med Sci. 2024;39:140. \u003c/li\u003e\n\u003cli\u003eQin P, Thirunavukarasu A, Arvanitis\u003csup\u003e1\u003c/sup\u003e T. Deep Learning Ensemble for Predicting Diabetic Macular Edema Onset Using Ultra-Wide Field Color Fundus Image. Image-Based Prediction of Retinal Disease Progression: MICCAI Challenges, DIAMOND 2024 and MARIO 2024, Held in Conjunction with MICCAI 2024, Marrakesh, Morocco, October 10, 2024, Proceedings. 2025;15503:64. \u003c/li\u003e\n\u003cli\u003eArcadu F, Benmansour F, Maunz A, Michon J, Haskova Z, McClintock D, et al. Deep learning predicts OCT measures of diabetic macular thickening from color fundus photographs. Invest Ophthalmol Vis Sci. 2019;60:852\u0026ndash;7. \u003c/li\u003e\n\u003cli\u003eHasan MM, Phu J, Wang H, Sowmya A, Meijering E, Kalloniatis M. Predicting visual field global and local parameters from OCT measurements using explainable machine learning. Sci Rep. 2025;15:5685. \u003c/li\u003e\n\u003cli\u003eYang W, Zhu Y, Shen J, Tang Y, Pan C, He H, et al. RURANET++: An Unsupervised Learning Method for Diabetic Macular Edema Based on SCSE Attention Mechanisms and Dynamic Multi-Projection Head Clustering. arXiv preprint arXiv:250220224. 2025; \u003c/li\u003e\n\u003cli\u003eMeng Z, Chen Y, Li H, Zhang Y, Yao X, Meng Y, et al. Machine learning and optical coherence tomography-derived radiomics analysis to predict persistent diabetic macular edema in patients undergoing anti-VEGF intravitreal therapy. J Transl Med. 2024;22:358. \u003c/li\u003e\n\u003cli\u003eTan T-E, Teo ZL, Vujosevic S, Ting DSW. AI-Based Diabetic Macular Edema Screening for Improved Care in Low-Resource Settings. NEJM AI. Massachusetts Medical Society; 2024. p. AIe2400698. \u003c/li\u003e\n\u003cli\u003eKusuma S, Udayan JD. Machine learning and deep learning methods in heart disease (HD) research. International Journal of Pure and Applied Mathematics. 2018;119:1483\u0026ndash;96. \u003c/li\u003e\n\u003cli\u003eAloysius N, Geetha M. A review on deep convolutional neural networks. 2017 international conference on communication and signal processing (ICCSP). IEEE; 2017. p. 588\u0026ndash;92. \u003c/li\u003e\n\u003cli\u003eHe Q, Jiang H, Fang D, Yang D, Nguyen TX, Ran A, et al. A graph neural network-based multispectral-view learning model for diabetic macular ischemia detection from color fundus photographs. arXiv preprint arXiv:250217886. 2025; \u003c/li\u003e\n\u003cli\u003eLi H-Y, Wang D-X, Dong L, Wei W-B. Deep learning algorithms for detection of diabetic macular edema in OCT images: A systematic review and meta-analysis. Eur J Ophthalmol. 2023;33:278\u0026ndash;90. \u003c/li\u003e\n\u003cli\u003eVinuesa R, Azizpour H, Leite I, Balaam M, Dignum V, Domisch S, et al. The role of artificial intelligence in achieving the Sustainable Development Goals. Nat Commun. 2020;11:233. \u003c/li\u003e\n\u003cli\u003eNedungadi P, Salethoor SN, Puthiyedath R, Nair VK, Kessler C, Raman R. Ayurveda research: Emerging trends and mapping to sustainable development goals. J Ayurveda Integr Med. 2023;14:100809. \u003c/li\u003e\n\u003cli\u003eKirankumar M, Saikiran P, Sabarivasan N, Kumaran M, Thangavel SK, Somasundaram K, et al. SoulNourish: A Comprehensive Machine learning Based Recommendation Yoga Prediction System Based on Health and Lifestyle Data. 2024 4th International Conference on Sustainable Expert Systems (ICSES). IEEE; 2024. p. 1828\u0026ndash;35. \u003c/li\u003e\n\u003cli\u003eLam C, Wong YL, Tang Z, Hu X, Nguyen TX, Yang D, et al. Performance of artificial intelligence in detecting diabetic macular edema from fundus photography and optical coherence tomography images: a systematic review and meta-analysis. Diabetes Care. 2024;47:304\u0026ndash;19. \u003c/li\u003e\n\u003cli\u003eMidena E, Toto L, Frizziero L, Covello G, Torresin T, Midena G, et al. Validation of an automated artificial intelligence algorithm for the quantification of major OCT parameters in diabetic macular edema. J Clin Med. 2023;12:2134. \u003c/li\u003e\n\u003cli\u003eYe X, Gao K, He S, Zhong X, Shen Y, Wang Y, et al. Artificial intelligence-based quantification of central macular fluid volume and VA prediction for diabetic macular edema using OCT images. Ophthalmol Ther. 2023;12:2441\u0026ndash;52. \u003c/li\u003e\n\u003cli\u003eAbramovich O, Pizem H, Van Eijgen J, Oren I, Melamed J, Stalmans I, et al. FundusQ-Net: A regression quality assessment deep learning algorithm for fundus images quality grading. Comput Methods Programs Biomed. 2023;239:107522. \u003c/li\u003e\n\u003cli\u003eKrithika Alias AnbuDevi M, Suganthi K. Review of semantic segmentation of medical images using modified architectures of UNET. Diagnostics. 2022;12:3064. \u003c/li\u003e\n\u003cli\u003eProcino E. CVAT meets transformers: accelerating semantic segmentation labeling in industrial applications. \u003c/li\u003e\n\u003cli\u003eTamil Priya D, Divya Udayan J. Transfer learning techniques for emotion classification on visual features of images in the deep learning network. Int J Speech Technol. 2020;23:361\u0026ndash;72. \u003c/li\u003e\n\u003cli\u003eKusuma S, Udayan JD. Machine learning and deep learning methods in heart disease (HD) research. International Journal of Pure and Applied Mathematics. 2018;119:1483\u0026ndash;96. \u003c/li\u003e\n\u003cli\u003eNeena A, Geetha M. Image classification using an ensemble-based deep CNN. Recent Findings in Intelligent Computing Techniques: Proceedings of the 5th ICACNI 2017, Volume 3. Springer; 2018. p. 445\u0026ndash;56. \u003c/li\u003e\n\u003cli\u003eVijay Gopal D V, Gopakumar G. A deep learning approach to image splicing using depth map. Advances in Distributed Computing and Machine Learning: Proceedings of ICADCML 2022. Springer; 2022. p. 401\u0026ndash;11. \u003c/li\u003e\n\u003cli\u003eMeng Z, Chen Y, Li H, Zhang Y, Yao X, Meng Y, et al. Machine learning and optical coherence tomography-derived radiomics analysis to predict persistent diabetic macular edema in patients undergoing anti-VEGF intravitreal therapy. J Transl Med. 2024;22:358. \u003c/li\u003e\n\u003cli\u003eHasan MM, Phu J, Wang H, Sowmya A, Meijering E, Kalloniatis M. Predicting visual field global and local parameters from OCT measurements using explainable machine learning. Sci Rep. 2025;15:5685. \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"discover-sustainability","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"disu","sideBox":"Learn more about [Discover Sustainability](https://www.springer.com/43621)","snPcode":"","submissionUrl":"","title":"Discover Sustainability","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Diabetic Macular Edema (DME), Macular Thickness, CVAT (Computer Vision Annotation Tool), Deep Learning, U-Net Segmentation, EfficientNetB0, InceptionV3, DenseNet121, Sustainable healthcare, Health equity","lastPublishedDoi":"10.21203/rs.3.rs-6795425/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6795425/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDiabetic Macular Edema (DME) is a major cause of vision impairment among diabetic patients, especially in areas lacking advanced diagnostic tools like Optical Coherence Tomography (OCT). This study proposes a cost-effective, AI-based method to estimate macular thickness\u0026mdash;a key DME biomarker\u0026mdash;using retinal fundus images and clinical data. A dataset of 176 colour fundus images with corresponding macular thickness and clinical information from DME patients at Amrita Institute of Medical Sciences, Kochi, India, was augmented to about 3,000 images to improve model performance. Manual macular region annotations were done using the CVAT tool, followed by segmentation using a U-Net-based deep learning model. For predicting macular thickness, three convolutional neural networks\u0026mdash;InceptionV3, DenseNet121, and EfficientNetB0\u0026mdash;were tested in a multimodal framework combining image features with clinical parameters like visual acuity and diabetes duration. EfficientNetB0 achieved the best performance, with a test Mean Squared Error (MSE) of 0.2518 and an R\u0026sup2; score of 0.7752. This approach enables accurate, low-cost, and scalable diagnostics in underserved regions, aligning with Sustainable Development Goals: ensuring healthy lives (SDG 3), fostering innovation (SDG 9), and reducing inequalities (SDG 10). The results suggest that this AI-based method can serve as a viable alternative to OCT for macular thickness estimation, supporting early detection and monitoring of DME in low-resource settings.\u003c/p\u003e","manuscriptTitle":"Towards Sustainable Retinal Diagnostics A Deep Learning Alternative to OCT for Macular Thickness Estimation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-01 09:11:24","doi":"10.21203/rs.3.rs-6795425/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-07-21T18:24:02+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-15T20:42:16+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-09T04:43:15+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-30T16:14:27+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"59757451782639928058015105717476867721","date":"2025-06-30T15:50:51+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"143482788449604552746794721776322340838","date":"2025-06-30T15:11:23+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"226628242435608436555155288832226329393","date":"2025-06-29T15:31:28+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-26T10:06:33+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"67661227733874064822987059868053409429","date":"2025-06-25T15:47:38+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-06-25T15:08:09+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-06-18T18:32:48+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-06-15T12:00:56+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-06-13T14:28:23+00:00","index":"","fulltext":""},{"type":"submitted","content":"Discover Sustainability","date":"2025-06-13T12:23:34+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"discover-sustainability","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"disu","sideBox":"Learn more about [Discover Sustainability](https://www.springer.com/43621)","snPcode":"","submissionUrl":"","title":"Discover Sustainability","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"7044693d-5e04-4443-a337-30ace7de0fe3","owner":[],"postedDate":"July 1st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-04-20T16:09:21+00:00","versionOfRecord":{"articleIdentity":"rs-6795425","link":"https://doi.org/10.1007/s43621-026-03218-5","journal":{"identity":"discover-sustainability","isVorOnly":false,"title":"Discover Sustainability"},"publishedOn":"2026-04-19 15:58:31","publishedOnDateReadable":"April 19th, 2026"},"versionCreatedAt":"2025-07-01 09:11:24","video":"","vorDoi":"10.1007/s43621-026-03218-5","vorDoiUrl":"https://doi.org/10.1007/s43621-026-03218-5","workflowStages":[]},"version":"v1","identity":"rs-6795425","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6795425","identity":"rs-6795425","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}