A New Open Source Toolkit for Segmenting 3D Intracellular Structures in Microscopy Images

Abstract

The Allen Institute for Cell Science is developing a state space of stem cell structural signatures to understand the principles by which cells reorganize during the cell cycle and differentiation. We have developed a pipeline that generates high-replicate, dynamic image data on cell organization and activities using endogenous fluorescently tagged human induced pluripotent stem cell (hiPSC) lines (www.allencell.org). Each line expresses a monoallelic EGFP-tagged protein that represents a particular cellular structure. For each structure, we take advantage of thousands of replicate high-resolution 3D images to develop quantitative image-based assays, analyses, and computational models. An important step in extracting information from the image data is to accurately segment each cellular structure in 3D. However, there remains a significant gap between state-of-the-art 3D computer vision algorithms and the openly accessible tools that biologists without computer vision expertise can easily use to accurately segment their 3D image data. To address this, we developed a Python-based open source toolkit for 3D segmentation that combines traditional segmentation algorithms with an iterative deep learning workflow. The general workflow is based on a minimal number of selectable algorithms and tunable parameters. We successfully applied this workflow to over 30 different intracellular structures, and our results can be used like a “lookup table” starting point for new applications. We implemented an iterative deep learning workflow that uses the results of the traditional segmentation together with minor manual inspection to generate a ground truth for training deep learning models. This approach improves the accuracy and robustness of those segmentation challenges that are not satisfactorily solved with standard algorithms. This open source segmentation toolkit aims to facilitate quantitative cell biology, conveniently leveraging state-of-the-art algorithms in computer vision for specific image segmentation needs and challenges.