Inferring the location and orientation of cell divisions on time-lapse image sequences
We propose a two-stage supervised framework for characterizing cell divisions in 2D and 3D time-lapse microscopy. First, we recast division detection as a semantic segmentation task on image sequences. Second, a local regression model estimates the orientation and distance between daughter cells for each event. We validate this framework using image sequences of avian neuroepithelium and mouse gastruloids. Our results demonstrate that high performance is achieved with efficient architectures, namely a U-Net for segmentation and a CNN for regression, that are optimized through systematic hyperparameter exploration. We find that integrating temporal context via multiple consecutive frames significantly boosts segmentation accuracy. We achieve F1 scores exceeding 94% (2D+t) and 90% (3D+t), with orientation accuracy approaching the uncertainty limit of manual annotation. We provide the full codebase and training workflow, specifically designed for datasets where traditional tracking is challenging.