CNN Identifies Exudative Macular Disease on OCT in AMD and DME Patients with 97% Accuracy

Read the paper →

AI-Powered OCT Analysis Detects Exudative Macular Disease with 97% Accuracy

New convolutional neural network aids retina specialists by automatically identifying intraretinal fluid and macular edema from OCT scans, promising faster and more accurate diagnosis of vision-threatening conditions.

Introduction

Diseases affecting the macula, such as age-related macular degeneration (AMD) and diabetic macular edema (DME), are leading causes of vision loss worldwide. Rapid identification of exudative changes and fluid accumulation in the retina is crucial for timely intervention and preventing irreversible damage. Optical coherence tomography (OCT) is the preferred imaging technique for detecting these retinal abnormalities, yet interpreting OCT scans is time-consuming and requires specialized expertise.

Despite advances in imaging, the reliance on manual review of OCT images poses challenges, especially in busy clinical settings or underserved areas where retina specialists may not be readily available. Artificial intelligence (AI), particularly deep learning, has emerged as a promising tool to augment OCT interpretation and improve diagnostic speed and accuracy.

In this study, a novel convolutional neural network (CNN) was developed specifically to detect intraretinal fluid and macular edema on OCT B-scans. The model demonstrated remarkable performance, achieving an area under the curve (AUC) of 0.965 and overall accuracy of 96.4%, underscoring its potential to assist clinicians in identifying exudative macular disease efficiently.

Study Partnership and Context

This research was conducted collaboratively by retina specialists and machine learning engineers from the Instituto Nacional de Investigación en Oftalmología and Arkangel AI in Medellín, Colombia. The setting in Colombia is significant, as timely diagnosis and treatment of retinal diseases can be particularly challenging in middle-income countries with limited subspecialist coverage.

The patient population included a mix of common retinal pathologies relevant to clinical practice, encompassing normal eyes, DME, wet AMD, and nonexudative retinal diseases. This diverse dataset ensured the CNN was trained and tested across a realistic spectrum of cases encountered by retina specialists.

Study Design and Methodology

The study followed a retrospective secondary source analysis approach, including OCT images from 108 patients totaling 158 eyes. Images were collected and examined by two retina specialists who labeled those containing biological markers indicative of intraretinal fluid secondary to exudative disease.

• Image dataset: 158 OCT B-scans from 108 patients, captured and curated with strict exclusion of images with artifacts or poor signal quality (below 6/10).
• Training and validation split: 110 images used for training, 28 for testing, and 20 for validation.
• Data augmentation: To overcome the relatively small dataset, the team applied extensive augmentations, including rotations, brightness adjustments, shear, zoom, and flips, increasing training images to 2,000.
• AI architecture: A convolutional neural network was designed to classify OCT scans as positive or negative for intraretinal fluid and macular edema, mimicking clinical decision-making.

The CNN operated by analyzing pixel-level patterns consistent with fluid accumulation, learning the subtle features that human experts rely on, yet automating and standardizing this complex interpretation process.

Key Results

• Classification accuracy: The model accurately classified 27 out of 28 test images, corresponding to an accuracy of 96.4%.
• AUC: The area under the receiver operating characteristic curve was 0.965, indicating outstanding discriminative ability.
• Precision and Recall: Precision reached 97.9%, showing a high positive predictive value, and recall (sensitivity) was 90.0%, demonstrating robust detection of true positives.
• Specificity: At 90.0%, the algorithm effectively minimized false positives.
• F1 Score: The harmonic mean of precision and recall was 0.934, highlighting balanced performance.
• Statistical Significance: The odds ratio was 14.81 (P < 0.001), confirming that correct classifications were far beyond chance.

Additionally, validation on a separate subset confirmed the reliability of these metrics, with confidence intervals supporting consistent high performance.

Interpretation and Clinical Implications

This CNN-based tool offers a powerful adjunct to retina specialists by quickly flagging OCT scans that show intraretinal fluid and macular edema, both key indicators of exudative retinal disease. Early identification leads to prompt referral, treatment initiation—often anti-VEGF therapy—and potentially preserves vision.

Automating this critical diagnostic step can streamline workflows, reduce clinician burden, and promote more accessible screening, especially in resource-limited settings. However, AI outputs should support rather than replace clinical judgment; specialists remain essential for definitive diagnosis and management decisions.

Limitations such as dependency on high-quality imaging and small training datasets mean additional prospective validation in real-world, noisy clinical environments is needed before widespread adoption.

Deployment and Scalability

The CNN model is well-suited for integration into clinical OCT devices or image review software, enabling near real-time analysis as images are captured. Implementing this technology could enhance screening programs or teleophthalmology services, improving early detection in regions with limited retina specialist availability.

Barriers to deployment include ensuring consistent imaging quality and establishing regulatory approvals. Addressing these will require collaboration among healthcare providers, device manufacturers, and regulatory bodies.

Moreover, the methodological framework can be adapted to identify other retinal pathologies or fluid types, expanding the utility of AI in ophthalmology.

Conclusion and Next Steps

This study showcases a high-performing AI model capable of accurately detecting macular edema and intraretinal fluid on OCT scans, demonstrating the promise of deep learning to augment clinical retina care. Prospective, real-time evaluations are necessary to confirm effectiveness in clinical workflows.

Future research should focus on increasing dataset diversity, improving robustness to lower quality images, and exploring integration pathways to facilitate broad clinical adoption. Such steps will help translate this innovation from the research setting to tangible improvements in patient outcomes worldwide.

Reference: Acosta CP, Piedrahita MA, Sanchez JG, Muñoz-Ortiz J, Martinez J, Zea J, et al. Using AI to Identify Exudative Macular Disease. Retina Today. July/August 2024.