This paper addresses the development of image segmentation, which plays an important role in deep learning-based computer vision, and its reliability-related challenges. In particular, the rapid adoption of CNN-based segmentation models in high-risk applications has led to active research on uncertainty quantification, which is rapidly expanding into a separate research field. This paper provides a comprehensive overview of probabilistic segmentation, discussing the basic concepts of uncertainty quantification, the developments in the field, and applications to various tasks. The literature on the two types of uncertainty (epistemic uncertainty, aleatoric uncertainty) traces four major applications: (1) quantifying statistical inconsistencies in annotation due to ambiguous images, (2) correlation between prediction errors and uncertainty, (3) expanding the model hypothesis space for better generalization, and (4) active learning. A broad discussion follows, including an overview of the datasets used in each application and an evaluation of available methods. We highlight challenges related to architecture, uncertainty quantification methods, standardization, and benchmarking, and conclude with recommendations for future research, such as single-forward-pass based methods and models that appropriately leverage volumetric data.
