Implicit Neural Representations for Generative Modeling of Living Cell Shapes

Methods allowing the synthesis of realistic cell shapes could help generate training data sets to improve tracking and segmentation in biomedical images. Deep generative models for shape require a light-weight flexible representation shape. However, commonly used voxel-based representations are unsuitable high-resolution synthesis, polygon meshes have limitations when modeling topology changes such as growth or mitosis. In this work, we propose use level signed distance functions (SDFs) represent shapes. We optimize neural network an implicit SDF value at any point 3D+time domain. The model is conditioned on latent code, thus new unseen sequences. validate our approach quantitatively qualitatively C. elegans cells that grow divide, lung cancer with growing complex filopodial protrusions. Our results show descriptors synthetic resemble those real cells, able topologically plausible sequences 3D+time.