Introduction
Age-related Macular Degeneration (AMD) remains a primary driver of vision impairment globally. It is characterized by the development of abnormal retinal vasculature, known as neovascularization. Optical Coherence Tomography Angiography (OCTA) has emerged as a critical diagnostic tool, providing non-invasive, three-dimensional visualization of these vascular structures. However, the standard clinical workflow often simplifies this complex 3D data into 2D projections. This reductionist approach introduces significant vulnerabilities, particularly when automated segmentation tools fail to accurately delineate retinal layers in diseased eyes.
The Problem: The Fragility of 2D Projections
In clinical practice, 2D projections are generated by segmenting specific retinal layers. When these layers are distorted by AMD, segmentation software frequently fails. Research presented at ICCV 2023 highlights that nearly 75% of active AMD samples contain such segmentation errors. These artifacts lead to missing vascular information or the introduction of noise, which compromises both human expert diagnosis and traditional 2D-based AI models.
Methodology: Direct 3D Volume Analysis
To address these limitations, researchers from UC San Diego developed a system termed Ours-3D. This approach bypasses the error-prone 2D projection step by analyzing raw 3D OCTA volumes directly.
Architecture and Training
The core of the system is a 2D EfficientNet backbone adapted for 3D data. Instead of utilizing a computationally intensive 3D CNN, the researchers treated the y-axis (the stack of B-scans) as the "channel" dimension. This design choice leverages the high redundancy between consecutive B-scans, allowing a standard convolution layer to compress the data into 64 channels without losing essential diagnostic features.
To guide the feature extraction process, the team implemented projection supervision. During the training phase, a secondary decoder branch attempts to reconstruct 2D OCTA projections from the 3D features. This injects domain knowledge into the model regarding vascular morphology. Crucially, this branch is discarded during inference, ensuring the final model remains independent of flawed 2D projections during actual clinical use.
A two-stage warmup strategy was also employed. By freezing the pretrained backbone initially and training only the randomly initialized layers, the researchers prevented the corruption of useful features learned from ImageNet.
Dataset and Evaluation
The study utilized a curated dataset from the Jacobs Retina Center at the Shiley Eye Institute, consisting of 889 raw OCTA volumes. The data was categorized into four stages: Active, Remission, Dry, and Normal. To rigorously test the model's resilience, the researchers created specific "error-free" and "error-prone" test sets, the latter containing the segmentation artifacts common in clinical settings.
Results and Discussion
The Ours-3D model demonstrated superior performance and robustness compared to both human experts and 2D-based AI models.
- Accuracy: The model achieved an accuracy of 83.93% on error-free data and maintained a high performance of 80% on error-prone data.
- Human Benchmarking: Human experts averaged approximately 60% accuracy, frequently struggling to distinguish between "active" (leaking) and "remission" (treated) stages.
- Robustness: While 2D models suffered significant performance degradation when faced with segmentation errors, the 3D approach remained stable. This suggests that raw 3D volumes contain subtle vascular patterns—invisible to the human eye or lost in 2D projections—that are critical for accurate staging.
Conclusion and Future Directions
The Ours-3D approach represents a significant step toward more reliable automated AMD diagnosis. By bypassing the "human bottleneck" of 2D interpretation and the technical failure point of layer segmentation, the model provides a more robust diagnostic pathway. Future work will focus on expanding the dataset, improving model explainability to identify new biomarkers, and integrating structural OCT B-scans for a more comprehensive multimodal diagnostic tool.
Summary updated