Densely Populated Cell and Organelles Segmentation With Multi-Plane Deep Learning Pipeline

Abstract

The complex and highly intertwined morphology of activated platelets within blood clot thrombi poses significant challenges for segmentation. Here we develop a robust machine vision pipeline for cell and organelle segmentation within volume electron microscope (vEM) datasets and validate that against both model CREMI neuron segmentation challenge data and focused isotropic FIB-SEM imaged platelets and their organelles within a series of ferric chloride indued murine occlusive clots Neural network predictions were collected along multiple planes, capturing 3D correlations using only 2D neural networks. Using a workstation class computer, we segmented and analyzed hundreds of platelets and report quantitative morphological measurements of platelets and their thousands of included organelles. CREMI neuron segmentation challenge analysis produced state-of-the-art outcomes as did platelet data in comparison to manual segmentation. Our work paves the way for large-scale, single cell, 3D studies across multiple examples and lends initial single cell level insights into occlusive thrombus structure and platelet heterogeneity therein. We include a lightweight Jupyter notebook for initializing and running the neural network (Data and Materials Availability). We also provide Amira Avizo protocols for converting neural network outputs to cell and organelle instances and subsequent morphological measurements.