Salient projections over maximum projections: Improving deep-learning detection and segmentation during invasion of cell-dense spheroids.

Recently, cell shape analysis is boosted by techniques from computer vision. As a result, it becomes possible to obtain high content information of cellular states from morphological data alone. In this work, we present on a highly adaptable deep-learning module which accepts 3D inputs to propogate 3D spatial latent embeddings as inputs to 2D instance segmentation neural networks. We apply the modified networks to 3D time-lapse images of MDA-MB-231 cell-dense spheroids to show improved continuous detection and quality of segmentations of individual cells during invasion. We show that the segmentations made by the modified 2D networks are markedly improved in comparison to maximum projection input images, making vital improvements for confocal imaging experiments with low axial resolution.