Explainable Deep Learning for Glaucomatous Visual Field Prediction: Artifact Correction Enhances Transformer Models.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) is crucial for diagnosing glaucoma.
• Artifacts in OCT scans can impair diagnostic accuracy.
• Predicting visual field (VF) loss from OCT is essential for glaucoma management.

Purpose of the Study:

• To develop a deep learning approach for restoring artifact-laden OCT scans.
• To predict functional loss on the Humphrey Visual Field (HVF) test using corrected OCT data.
• To enhance the correlation between glaucomatous structures and functions.

Main Methods:

• A generative diffusion model was used to correct peripapillary retinal nerve fiber layer (RNFL) thickness map artifacts.
• Three convolutional neural networks and two transformer-based models were trained on original and artifact-corrected datasets.
• Distillation with No Labels (DINO) Vision Transformers (ViT) were employed for prediction and explainability analysis.

Main Results:

• The DINO-ViT model trained on artifact-corrected OCT data achieved the highest predictive accuracy (RMSE: 4.44 dB, MAE: 3.46 dB).
• Artifact correction led to significant improvements in global RMSE and MAE (P < 0.05).
• Explainability tools confirmed that artifacts compromise predictive ability, which improves after correction.

Conclusions:

• Combining self-supervised ViTs with generative artifact correction enhances structure-function correlation in glaucoma.
• The developed approach provides a tool for glaucoma management and research.
• Addressing artifacts is critical for accurate clinical interpretation of OCT scans.