Estimating Uncertainty of Geographic Atrophy Segmentations with Bayesian Deep Learning.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Geographic atrophy (GA) is a leading cause of vision loss in age-related macular degeneration (AMD).
• Accurate segmentation of GA lesions is crucial for monitoring disease progression and evaluating treatment efficacy.
• Deep learning models show promise for automated GA segmentation, but quantifying prediction uncertainty remains a challenge.

Purpose of the Study:

• To apply and evaluate Bayesian deep learning techniques for quantifying uncertainty in semantic segmentation of geographic atrophy (GA).
• To compare the performance of Bayesian methods (Monte Carlo dropout, ensemble) against a traditional deep learning model for GA segmentation.

Main Methods:

• Retrospective analysis of spectral-domain optical coherence tomography (SD-OCT) images from the SWAGGER cohort.
• Development of two approximate Bayesian deep learning models: Monte Carlo dropout and ensemble methods.
• Comparison of model performance using Dice scores and calculation of pixel-wise uncertainty using Shannon Entropy.

Main Results:

• Bayesian deep learning models demonstrated significantly higher Dice scores (MC dropout: 0.90, ensemble: 0.88) compared to the traditional model (0.82).
• Both Bayesian methods produced a greater number of pixels with high entropy, indicating higher uncertainty estimates.
• The models provided pixel-wise estimates of uncertainty for GA segmentation.

Conclusions:

• Quantifying prediction uncertainty in GA segmentation enhances model trustworthiness for clinical applications.
• Bayesian deep learning techniques improve segmentation performance and provide valuable uncertainty measures, aiding clinician decision-making.
• These advanced methods offer a more robust approach to analyzing GA in OCT imaging.