Estimating Uncertainty of Geographic Atrophy Segmentations with Bayesian Deep Learning

Purpose: To apply methods for quantifying uncertainty of deep learning segmentation of geographic atrophy (GA). Design: Retrospective analysis of OCT images and model comparison. Participants: One hundred twenty-six eyes from 87 participants with GA in the SWAGGER cohort of the Nonexudative Age-Related Macular Degeneration Imaged with Swept-Source OCT (SS-OCT) study. Methods: The manual segmentations of GA lesions were conducted on structural subretinal pigment epithelium en face images from the SS-OCT images. Models were developed for 2 approximate Bayesian deep learning techniques, Monte Carlo dropout and ensemble, to assess the uncertainty of GA semantic segmentation and compared to a traditional deep learning model. Main Outcome Measures: Model performance (Dice score) was compared. Uncertainty was calculated using the formula for Shannon Entropy. Results: The output of both Bayesian technique models showed a greater number of pixels with high entropy than the standard model. Dice scores for the Monte Carlo dropout method (0.90, 95% confidence interval 0.87-0.93) and the ensemble method (0.88, 95% confidence interval 0.85-0.91) were significantly higher (P < 0.001) than for the traditional model (0.82, 95% confidence interval 0.78-0.86). Conclusions: Quantifying the uncertainty in a prediction of GA may improve trustworthiness of the models and aid clinicians in decision-making. The Bayesian deep learning techniques generated pixel-wise estimates of model uncertainty for segmentation, while also improving model performance compared with traditionally trained deep learning models.