Reconstructing axons from sectioned brains through alignment and stitching

Abstract

Since the turn of this century, new and powerful cell labeling techniques have allowed the consistent and complete labeling of axonal arborizations of individual long range projection neurons (LRPNs) in mouse (Furuta et al, 2001). The prospect of creating a connectome at the cellular level, together with new developments in the automated scanning of whole (rodent) brain volumes, has resulted in a the release of thousands of fully reconstructed LRPNs (Peng et al. 2021, Winnubst et al. 2019). In parallel with this number-driven ‘industrial’ approach, there are smaller scale, hypothesis-driven projects that focus on particular populations of neurons, perform targeted single-cell labeling experiments and explore the resulting data to the maximum possible extent. In this setting, axons are traced from a stack of brain slices, stained to highlight particular features and mounted on a glass slide. Tracing axons from this data is a two step process: 1. Find all pieces of axon in each individual section, carried out with microscope-attached software NeuroLucida (MBF Bioscience) 2. Connect pieces of axon across adjacent sections to create a full axonal reconstruction, a process that we refer to as alignment and stitching. This work takes care of the second step, having sections with traced axonal fragments as input, and complete axons as output. Our approach consists of: Preparation of a structured dataset that contains for each section all pieces of axon and a section image in a shared coordinate system. Implemented in Python. Alignment of sections using two complementary methods: For sections that contain only a few pieces of axon: use a tissue-based alignment tool that presents the user with the overlay of (a) a tissue section, (b) the contours of the adjacent section, and (c) all pieces of axon that are traced in the two sections. The user manually rotates and shifts the overlay to obtain the best possible match. Implemented as a web-based tool. For sections with many pieces of axon: align sections to create a maximum number of matching pieces, using a modified version of the Dercksen et al. (2005) algorithm. Implemented in Python. Stitching pieces of neuron using a greedy approach that starts with an initial segment (the soma), and at each iteration connects a piece from the set of unconnected pieces to one of the endings of the growing neuron, in such a way that the added axonal length is minimal, and permitting only stitches between adjacent sections. Implemented in Python. Registration: combine slices into a volume by using the POSSUM pipeline of Majka et al. (2016) and use 3d warping to register to a reference space.