Towards an Unsupervised Spatiotemporal Representation of Cilia Video Using A Modular Generative Pipeline


Motile cilia are a highly conserved organelle found on the exterior of many human cells. Cilia beat in rhythmic patterns to transport substances or generate signaling gradients. Disruption of these patterns is often indicative of diseases known as ciliopathies, whose consequences can include dysfunction of macroscopic structures within the lungs, kidneys, brain, and other organs. Characterizing ciliary motion phenotypes as healthy or diseased is an essential step towards diagnosing and differentiating ciliopathies. We propose a modular generative pipeline for the analysis of cilia video data so that expert labor may be supplemented for this task. Our proposed model is divided into three modules: preprocessing, appearance, and dynamics. The preprocessing module augments the initial data, and its output is fed frame-by-frame into the generative appearance model which learns a compressed latent representation of the cilia. The frames are then embedded into the latent space as a low-dimensional path. This path is fed into the generative dynamics module, which focuses only on the motion of the cilia. Since both the appearance and dynamics modules are generative, the pipeline itself serves as an end-to-end generative model. This thorough and versatile model allows experts to spend less time caught in the minutiae of cilia biopsy analysis, while also enabling new insights by quantifying subtle patterns that would be otherwise difficult to categorize.

Keywords:Machine LearningData ScienceVideo AnalysisGenerative ModelingVariational AutoencoderModularPipeline