Automated Detection of Retinal Signs of Degeneration Using Convolutional Neural Networks: A Retrospective Observational Study

Abstract:

Retinal degeneration, a major cause of vision impairment, demands early detection and intervention. Optical coherence tomography (OCT) facilitates the identification of retinal abnormalities, but manual interpretation is time-consuming and subjective. This study aimed to develop a deep learning-based system using convolutional neural networks (CNNs) for automated detection of retinal signs of degeneration. The system, trained and tested on OCT images, achieved high accuracy, sensitivity, and specificity in identifying various abnormalities. This highlights the potential of deep learning algorithms for early diagnosis and monitoring of retinal degeneration.

Introduction:

Retinal degeneration, encompassing conditions like age-related macular degeneration, diabetic retinopathy, and glaucoma, leads to irreversible vision loss. Early detection and intervention are crucial. OCT provides cross-sectional retinal images, enabling the identification of abnormalities. However, manual interpretation is challenging, especially for large populations. Deep learning, particularly CNNs, has shown promise in image recognition and classification tasks. This study aimed to develop a CNN-based system for automated detection of retinal signs of degeneration.

Methods:

Data Collection:
– Retrospective observational study at University Eye Clinic, Trieste, Italy.
– OCT scans acquired from healthy and pathological eyes.
– Inclusion criteria: presence of specific retinal signs (epiretinal membrane, inner foveal layer abnormalities, subretinal fluid, drusen, microaneurysms, vitreous macular adhesion, macular hole, breaks in Bruch’s membrane) in pathological group.

Image Labeling and Preprocessing:
– Images labeled by two experienced retinal specialists.
– Poor-quality images and images with disagreement excluded.
– Images cropped and resized to standard input size.

Datasets Population and Training Process:
– Nine binary predictive models created.
– Balanced datasets with 10% test set and 90% for fivefold cross-validation.

Modeling:
– Modified VGG-16 CNN architecture used.
– Transfer learning and fine-tuning techniques applied.
– Early stopping technique for optimal performance.
– Models trained for a maximum of 70 epochs.

Evaluation Metrics:
– Confusion matrices, accuracy, sensitivity, specificity, AUC, and Cohen’s Kappa indexes calculated.

Model Visualization (GRAD-CAM):
– Grad-CAM heatmaps used to understand CNN predictions.

Results:

Overall Performance:
– High accuracy, sensitivity, specificity, and AUC achieved by the models.
– Model 1: Healthy vs. Pathological (AUC: 0.999)
– Model 2: Epiretinal Membrane (AUC: 0.992)
– Model 3: Inner Foveal Layer Abnormalities (AUC: 0.987)
– Model 4: Subretinal Fluid (AUC: 0.991)
– Model 5: Drusen (AUC: 0.989)
– Model 6: Microaneurysms (AUC: 0.993)
– Model 7: Vitreous Macular Adhesion (AUC: 0.988)
– Model 8: Macular Hole (AUC: 0.995)
– Model 9: Breaks in Bruch’s Membrane (AUC: 0.990)

Model Visualization (GRAD-CAM):
– Heatmaps highlighted regions influencing the model’s decisions.

Conclusion:

The developed CNN-based system demonstrated high accuracy in detecting retinal signs of degeneration. This automated system has the potential to aid in early diagnosis and monitoring of retinal degeneration, facilitating timely intervention and preserving vision. Further studies with larger datasets and clinical implementation are warranted to confirm the findings and assess the system’s real-world applicability.