Abstract
The prediction model reveals the potential to predict persistent disease activity (PDA) in neovascular age-related macular degeneration (nAMD) eyes. The integrity of the ellipsoid zone (EZ) and external limiting membrane (ELM) constituted the strongest predictive factor for PDA in nAMD eyes in real-world clinical practice. The results of this study are a significant step toward image-guided prediction of long-term disease activity in the management of nAMD and highlight the importance of the automatic identification of photoreceptor layers.
Introduction
Neovascular age-related macular degeneration (nAMD) is a leading cause of vision loss globally. While anti-VEGF injections have revolutionized treatment, patient response is highly variable. Many patients suffer from persistent disease activity (PDA)—characterized by ongoing fluid leakage and hemorrhage despite regular treatment. Currently, clinicians rely on manual inspection of Spectral Domain Optical Coherence Tomography (SD-OCT) scans to make treatment decisions. However, manual segmentation is time-consuming, prone to error, and often overlooks subtle structural changes in the photoreceptor layers. There is a critical need for automated, objective tools that can analyze initial scans to predict long-term disease trajectories.
Researchers have developed a deep learning pipeline that automatically quantifies retinal biomarkers from OCT scans to predict one-year treatment outcomes. The study reveals that the integrity of the outer retinal layers—specifically the EZ and ELM—is a far more potent predictor of PDA than the presence of retinal fluid alone.
Methodology
The authors proposed a two-stage "segmentation-then-prediction" framework to automate the identification of biomarkers and the prediction of clinical outcomes.
Automated Feature Extraction
The algorithm targets two types of retinal structures:
- Linear Structures: A classification Convolutional Neural Network (CNN) identifies the Ellipsoid Zone (EZ) and External Limiting Membrane (ELM). A "clipping-training" approach was used, slicing B-scans vertically to detect these thin, often disrupted layers.
- Blocky Structures: A Fully Convolutional Network (FCN) with an EfficientNet backbone segments four other biomarkers: intraretinal fluid (IRF), subretinal fluid (SRF), pigment epithelium detachment (PED), and subretinal hyperreflective material (SHRM).
Outcome Prediction
Quantitative features such as area, diameter, length, and intensity from these six biomarkers were fed into a Random Forest model. The model predicts 1-year disease activity, classifying eyes as Stable (Type 0), PDA (Type 1), or Cured (Type 2).
Datasets
The study utilized a longitudinal dataset of 671 SD-OCT B-scans from 186 patients at the Qilu Hospital of Shandong University. Data was collected at baseline, 1 month, 3 months, and 12 months. Manual annotations by expert ophthalmologists provided the ground truth. The dataset was split into 148 patients for training and 38 for testing, with no patient overlap.
Results and Discussion
The deep learning algorithm achieved high performance in both segmentation and prediction tasks.
Segmentation and Prediction Performance
The algorithm achieved a mean accuracy of 0.930 and a Dice coefficient of 0.873 across all six features. For the 3-month prediction model, the results were highly accurate:
- AUC: 0.980
- Accuracy: 0.930
- Sensitivity: 0.920
- Specificity: 0.962
Key Insights
- Photoreceptors over Fluid: The integrity of the EZ and ELM at baseline and month 1 showed the strongest correlation with 1-year disease activity. While clinicians traditionally focus on fluid (IRF/SRF), this study suggests that the structural health of the retina at the start of treatment is a superior indicator of long-term prognosis.
- Longitudinal Data Synergy: Predictive power improved significantly as more data points were added. A model using only baseline data achieved an AUC of 0.948, while incorporating data through the 3-month "loading phase" reached an AUC of 0.980.
- Biomarker Synergy: The "six-metric" model (including EZ/ELM) significantly outperformed models that only considered fluid and PED, highlighting the biological importance of specific, targeted feature extraction.
Conclusion
This work moves toward precision medicine in ophthalmology, where AI predicts long-term disease trajectories rather than just describing current states. However, limitations remain. The study was limited to a relatively small sample of Chinese patients, requiring further validation in multiracial populations. Additionally, the model does not yet account for other factors like hyperreflective foci or choroidal thickness. Future work will focus on clinical integration and validation in prospective trials to transition this research tool into a real-world clinical decision support system.
Summary updated