A Line Is All You Need: Weak Supervision for 2.5D Cell Segmentation

We propose a novel 2.5D weak annotation scheme called Linely Supervision, alongside a 2.5D Cell Segmentation pipeline that produces full and accurate 3D segmentation masks from weak labels. Even though the manual annotation of microscopic images is up to 10.5 times faster using line annotations, we show that even a single weak label per cell suffices to retain \(98.4\%\) of segmentation performance. We additionally benchmark several multi-slice input strategies, as well as 3D reconstruction strategies, and show that our weakly supervised approach is capable of outperforming one of the top contenders in the ISBI Cell Tracking Challenge. Since an ever-growing amount of 3D microscopic data is publicly available but not sufficiently annotated to be used in deep learning tasks, Linely annotation can make these valuable resources more accessible to researchers in AI and biology alike.