Fast Assembling of Neuron Fragments in Serial 3D Sections

Hanbo Chen, Daniel Maxim Iascone, Nuno Maçarico da Costa, Ed S. Lein, Tianming Liu, Lili YuanbioRxiv 2016 · DOI

TL;DR

Reconstructing entire neurons from serial 3D microscopy sections is a major bottleneck in neuroscience. This paper introduces NeuronStitcher, a novel software suite that automatically stitches together neuron fragments from adjacent sections, significantly accelerating the process and enabling the reconstruction of complex neuronal structures.

What Problem Does This Paper Address?

Understanding the brain's intricate wiring requires reconstructing individual neurons in their entirety. This is particularly challenging when dealing with thick brain tissue, which must be physically sectioned for high-resolution imaging. While advanced techniques like brain clearing and 3D imaging exist, many researchers still rely on serial sectioning. The process of imaging these sections and then digitally reassembling them into a coherent 3D volume is laborious. A critical step is stitching together neuron fragments (neurites) that are severed at the boundaries of these sections. This "stitching" process is often manual, time-consuming, and prone to errors, especially in densely packed neuronal regions, thus hindering large-scale brain mapping efforts.

Methods

The core of NeuronStitcher is a two-stage process. First, it employs a triangle-matching algorithm to estimate the initial alignment of severed neurites. This algorithm leverages the relative positions and branch orientations of neurite fragments on the sectioning plane. Second, an iterative alignment process refines these matches until the neurites are smoothly stitched together. To handle variations in neurite quality, NeuronStitcher also incorporates three methods for noise removal during the identification of severed neurites. The software is implemented as a plugin for the Vaa3D platform, providing an interactive graphical user interface for visual evaluation, parameter adjustment, and manual correction of stitching results.

Datasets

The paper evaluates NeuronStitcher using two main types of datasets:

  • Mouse Brain Tissue: A labeled pyramidal neuron from the hippocampal CA1 region of a mouse brain was used to generate a "ground truth" reconstruction. This neuron was then digitally sectioned, imaged, and reconstructed from fragments.
  • Simulated Data: To further test robustness, reconstructions from mouse visual cortex were digitally "sectioned" into halves, then randomly rotated and shifted. Various gap sizes (1, 2, 4, and 8 µm) were introduced to simulate tissue loss during sectioning.
  • Biocytin-filled Neurons: Experiments also involved stitching images of a mouse V1 neuron from confocal microscopy and two biocytin-filled human neurons imaged by bright field microscopy.

Key Insights and Ideas

  • Bridging the Gap: The paper highlights that stitching across section boundaries is a distinct and more challenging problem than stitching overlapping tiles within a single section, due to potential tissue loss and distortions.
  • Algorithmic Efficiency: NeuronStitcher's triangle-matching algorithm provides an efficient way to establish initial matches, reducing the reliance on purely manual or brute-force approaches.
  • Robustness to Imperfections: The inclusion of noise reduction methods and the demonstrated robustness to simulated tissue loss (up to 8 µm gaps) are crucial for real-world applications where perfect data is rare.
  • User-Centric Design: Integrating NeuronStitcher into Vaa3D with an interactive GUI allows for expert oversight and correction, balancing automation with the need for accuracy in complex biological data.

Key Results

  • High Accuracy: In experiments with a mouse hippocampal neuron, NeuronStitcher achieved a 98% correspondence of bifurcations in the stitched reconstruction compared to the ground truth.
  • Robustness to Gaps: Even with an 8 µm gap simulating significant tissue loss, the alignment was close to the ground truth, with 75% precision and 62% sensitivity for matching severed neurites using default parameters. With a 1 µm gap, precision reached 90±6% and sensitivity 87±15%.
  • Speed and Resource Efficiency: NeuronStitcher typically completed computations within seconds, requiring less than 100MB of memory.
  • Expert Validation: Across various experiments, an expert accepted 86.6% of the automatically matched neurites.
  • Time Savings: The average time for a user to visually check and manually adjust results for an adjacent pair of serial sections was approximately 13 minutes (median 9'41"), a significant improvement over purely manual methods.

Limitations and Future Work

  • Parameter Sensitivity: While robust, the paper notes that the quality of neuron reconstructions and the selected parameters can influence NeuronStitcher's performance.
  • User Intervention: The time for manual adjustment still depends on the complexity of the reconstruction, indicating that fully automated reconstruction remains an ongoing challenge.
  • Future Directions: The authors aim to broaden NeuronStitcher's utility across different data acquisition processes by refining its interactive capabilities and potentially extending its application to other types of biological structures. The software's open-source nature on the Vaa3D platform facilitates further development and community contributions.

Summary updated

Abstract

Reconstructing neurons from 3D image-stacks of serial sections of thick brain tissue is very time-consuming and often becomes a bottleneck in high-throughput brain mapping projects. We developed NeuronStitcher, a software suite for stitching non-overlapping neuron fragments reconstructed in serial 3D image sections. With its efficient algorithm and user-friendly interface, NeuronStitcher has been used successfully to reconstruct very large and complex human and mouse neurons.
BibTeX
@misc{chen2016fast,
  author = {Hanbo Chen and Daniel Maxim Iascone and Nuno Maçarico da Costa and Ed S. Lein and Tianming Liu and Lili Yuan},
  title = {{Fast Assembling of Neuron Fragments in Serial 3D Sections}},
  year = {2016},
  doi = {10.1101/097253},
  howpublished = {bioRxiv},
}
Figure