About his postdoctoral project

Deep-learning techniques have significantly improved the robustness and speed of cell segmentation in fluorescence microscopy. Yet their application to deep tissue imaging, such as organoïd imaging, remains limited. Indeed, these neural networks are trained with manual annotations, a process dependent on intuitive visual inspection, which becomes impractical in dense, heterogeneous and three-dimensional cellular aggregates in 3D. As a result, developing unsupervised approaches remains a critical bottleneck for the analysis of physiologically relevant tissue models.

To overcome this limitation, our project explores unsupervised learning strategies that exploit the temporal consistency of cellular dynamics. By integrating a measure of dynamic coherence directly into the training process, we aim to refine neural networks using biological priors rather than manual labels, enabling more robust and generalizable segmentation in complex 3D tissues.