How to Box Your Cells: An Introduction to Box Supervision for 2.5D Cell Instance Segmentation and a Study of Applications

Fabian Schmeisser, Maria Caroprese, Gillian Lovell, Andreas Dengel, Sheraz Ahmed

Abstract: Cell segmentation in volumetric microscopic images is a fundamental step towards automating the analysis of life-like representations of complex specimens. As the performance of current Deep Learning algorithms is held back by the lack of accurately annotated ground truth, a pipeline is proposed that produces accurate 3D cell instance segmentation masks solely from slice-wise bounding box annotations. In an effort to further reduce the time requirements for the annotation process, a study is conducted on how to effectively reduce the size of the training set. To this end, three slice-reduction strategies are suggested and evaluated in combination with bounding box supervision. We find that as low as 1% of weakly labeled training data suffices to produce accurate results, and that predictions produced by a 10 times smaller dataset are of equal quality to when the full dataset is exploited for training.