Uncertainty Estimation on Semantic Segmentation for Microscopy Data

In microscopy, many common data analysis tasks rely on an initial semantic segmentation step. Microscopy data are very diverse, and thus this segmentation might fail due to being out-of-distribution (OOD). For users to know whether the downstream tasks are possible or accurate, it is necessary to assess the accuracy of the semantic segmentation step. This can be done through uncertainty estimation of the predictions, either at the image or pixel level. To address this, we are conducting detailed research focusing on uncertainty estimation methods across four key categories: Deterministic, Bayesian Neural Networks (BNN), Ensemble, and Test Time Augmentation (TTA). This work aims to explore both well-established and emerging methods for uncertainty estimation in semantic segmentation applied to microscopy data.