Preprint Highlight: Learning orientation-invariant representations enables accurate and robust morphologic profiling of cells and organelles

One application of deep learning in analysis of cell biological microscopy data is developing meaningful quantitative representations of cellular and/or molecular phenotypic signatures. Because image orientation has no relevance for shape and morphology, encoding orientation within such representations confounds downstream analyses.

This study presents O2-VAE, a neural method for learning orientation-invariant, image-based shape representations. The authors demonstrate that O2-VAE is not sensitive to image orientation for applications of cell/organelle shape phenotyping. Specifically, O2-VAE was verified on diverse experimental systems, ranging from simulations to human induced pluripotent stem cells, for downstream analyses that include clustering, dimensionality reduction, and/or outlier detection.

This article opens the door for design and evaluation of orientation-invariant representations that may enable more effective deep learning–driven phenotyping.