Bayesian techniques outperform traditional models in geographic atrophy segmentation

Bayesian deep learning techniques, such as Monte Carlo dropout and ensemble methods, not only improve the accuracy of geographic atrophy (GA) segmentation compared to traditional models but also provide valuable pixel-wise estimates of uncertainty, according to a study.

The retrospective analysis included 126 eyes from 87 participants in the SWAGGER cohort of the Nonexudative Age-Related Macular Degeneration Imaged with Swept-Source OCT (SS-OCT) study, using OCT images to compare model performance.

The Bayesian deep learning techniques—Monte Carlo dropout and ensemble—were applied to assess uncertainty in segmenting GA lesions. These techniques were compared to a traditional deep learning model. The Monte Carlo dropout achieved a Dice score of 0.90, while the ensemble method scored 0.88, both significantly higher than the traditional model’s 0.82.

In addition to improved accuracy, the Bayesian models also provided better estimates of uncertainty.

Reference
Spaide T, Rajesh AE, Gim N, et al. Estimating Uncertainty of Geographic Atrophy Segmentations with Bayesian Deep Learning. Ophthalmol Sci. 2024;5(1):100587. doi: 10.1016/j.xops.2024.100587. PMID: 39380882; PMCID: PMC11459066.

This content is independent editorial sponsored by Astellas. Astellas had no input in the development of this content.