In the field of computer vision, probabilistic convolutional neural networks, which predict distributions of predictions instead of point estimates led to many recent advances. Besides state of the art benchmark results, these networks made it possible to quantify local uncertainties in the predictions. These were used in active learning frameworks to target the labeling efforts of specialist annotators or to assess the quality of a prediction in a safety-critical environment. However, for instance segmentation problems, which aim at separating different objects (e.g. cells) from one another, these methods are not frequently used so far. We seek to close this gap by proposing a generic method to obtain model-inherent uncertainty estimates within proposal-free instance segmentation models. Furthermore, the quality of the uncertainty estimates is analyzed with a metric adapted from semantic segmentation, which seeks to separate objects based on their class (e.g. epithelial cells or other cells). The method is evaluated on a dataset that contains C.elegans brightfield microscopy images, where it yields competitive performance while also predicting uncertainty estimates that carry information about object-level inaccuracies like false splits and false merges. These uncertainty estimates are then used in a simulation study to guide proofreading efforts.
The manuscript was accepted for the Bio Image Computing Workshop within ECCV2020 and is available at arXiv
Rumberger, J.L., Mais, L., Kainmüller, D. Probabilistic Deep Learning for Instance Segmentation