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

This repository is for the experimental notebooks for the paper "Explainable Deep Learning for Glaucomatous Visual Field Prediction: Artifact Correction Enhances Transformer Models". This research aims to develop and evaluate the accuracy and interpretability of deep learning models trained on optical coherence tomography (OCT) scans with artifact-removal preprocessing in predicting visual field (VF).

Methods

• Cross-sectional, retrospective datasets including reliable VF and OCT measurements obtained within six months.
• Training set: 1,674 VF-OCT pairs from 951 eyes
• Testing set: 429 VF-OCT pairs from 345 eyes
• Models trained: CNNs, Vision Transformer (ViT), and DINO-ViT
• Task: Estimate HFA 24-2 VF thresholds
• Input: Peripapillary retinal nerve fiber layer (RNFL) thickness maps
• Comparison: Models trained on original vs artifact-corrected datasets

Results

• Evaluation metrics: Pointwise root mean square error (RMSE) and mean absolute error (MAE)
• Explainability techniques: GradCAM, GradCAM++, attention maps, uniform manifold approximation and projection, and principal component analysis
• Best performing model: DINO-ViT trained on artifact-corrected datasets
  • RMSE = 4.44 dB
  • MAE = 3.46 dB
• Improvement: Global RMSE and MAE reductions of 0.15 dB compared to performance on original maps
• Findings: Artifacts compromise DINO-ViT's predictive ability but improve with artifact correction

Conclusions

Transformer-based models enhance the accuracy and interpretability of visual function estimations from OCT scans, with RNFL artifact correction further refining these improvements.

Notebooks

You can see the experimental notebooks in

• resnet34.ipynb
• vgg16.ipynb
• vit.ipynb
• dino.ipynb

Requirements

Please see requirements.txt for dependencies. You can install the dependencies using pip via

pip install -r requirements.txt