The NanoZoomer Connectomics Pipeline for Tracer Injection Studies of the Marmoset Brain

Alexander Woodward, Rui Gong, Hiroshi Abe, Ken Nakae, Junichi Hata, Henrik Skibbe, Yoko Yamaguchi, Shin Ishii, Hideyuki Okano, Tetsuo Yamamori, Noritaka IchinohebioRxiv 2019 · DOI

TL;DR

This paper introduces a novel computational pipeline for mapping neural connections in the marmoset brain using tracer injections. By combining advanced AI for image segmentation with a sophisticated 3D reconstruction and registration process, they've created a robust system for large-scale connectomics studies, overcoming limitations of previous methods.

What Problem Does This Paper Address?

Mapping the intricate wiring of the brain, known as connectomics, is crucial for understanding its function. Tracer injection studies are a powerful tool for this, but processing the resulting high-resolution fluorescence images is challenging. Traditional methods struggle with image artifacts, varying illumination, and the sheer volume of data. Furthermore, accurately aligning individual brains to a common reference space is difficult due to anatomical variations. This paper tackles these issues by developing an automated pipeline that can precisely segment tracer signals and normalize individual brain data into a standardized 3D space.

Methods

The core of this work is a multi-stage pipeline:

  • High-Resolution Imaging: Brain sections are imaged using a NanoZoomer 2.0-HT batch slide scanner.
  • AI-Powered Tracer Segmentation: A U-Net deep learning architecture is employed to precisely segment anterograde tracer signals from background noise, even in images with low contrast or artifacts. This is a key novelty, as previous methods struggled with these challenges.
  • 3D Brain Reconstruction: Block-face images are used to reconstruct the 3D shape of each individual brain.
  • Individual Brain Atlasing: Each brain is cytoarchitectonically annotated to create an individual 3D atlas, addressing the variability inherent in using a single template atlas.
  • Common Brain Space Registration: Individual brain data is non-linearly mapped into a common 3D brain space, which is based on a population-average MRI and the Brain/MINDS atlas.
  • 2D Cortical Flat Map Alignment: To further account for individual variations, the cortical boundaries in the 2D flat map space are aligned between the individual and common brain spaces.

Datasets

  • Marmoset Brain Data: The study utilizes experimental data from common marmoset brains, including:
  • NanoZoomer fluorescence images (F, T, and C channels).
  • Block-face images.
  • Myelin and Nissl staining images.
  • BZ-X700 microscopy images for injection sites.
  • Ex-vivo MRI T2-weighted and diffusion-weighted images.
  • Brain/MINDS Atlas: A pre-existing 3D digital marmoset brain atlas was used as a reference.
  • Training Data: Manually created segmentation masks for fluorescence images were used to train the U-Net model.

Key Insights and Ideas

  • AI for Tracer Segmentation: Using a U-Net architecture is a significant advancement for segmenting tracer signals, overcoming limitations of traditional thresholding methods that fail with non-uniform illumination and artifacts.
  • Individual Atlasing Reduces Bias: Creating individual 3D atlases for each brain, rather than relying solely on a template, is crucial for accurately mapping data and accounting for anatomical variability.
  • Multi-Modal Integration: The pipeline is designed to integrate data from various sources (fluorescence, MRI, histology), paving the way for comprehensive connectomics.
  • 2D Flat Map Alignment for Precision: Aligning cortical boundaries in the 2D flat map space is a clever way to absorb individual brain variations and precisely map tracer data.

Key Results

  • Successful Tracer Segmentation: The U-Net model demonstrated superior performance in tracer segmentation compared to traditional methods like Ridge Detection, Otsu thresholding, and Max Entropy thresholding, achieving statistically significant improvements in Jaccard index and Dice coefficient (p < 0.01).
  • Accurate 3D Reconstruction and Registration: The pipeline successfully reconstructed individual brain shapes and mapped them into a common brain space with an average error of approximately 0.39mm (around 1% of the brain size) for landmarks.
  • Pipeline Efficiency: The processing of an individual brain, including reconstruction and mapping, took approximately five hours per brain on a high-performance cluster.

Limitations and Future Work

  • Manual Masking: While the U-Net is powerful, some manual masking was still required for background and injection sites, especially for the injection site due to saturation in NanoZoomer images. The authors plan to automate this in the future.
  • Connectomics Matrix Calculation: The paper presents preliminary connectomics calculations but acknowledges that a systematic approach to integrating and normalizing data from multiple injections into a single connection matrix is still future work.
  • Subcortical Structures: Currently, only cortical regions are delineated for individual atlases. While the Brain/MINDS atlas includes subcortical structures, future work could involve delineating these for individual brains as well.
  • Improving Registration: The authors plan to explore manual adjustments and further AI-driven methods to improve 3D registration accuracy.
  • Expanding Atlas: Future work includes developing a population-based brain atlas that incorporates more detailed histological annotations and MRI/DWI information to better evaluate variations in cortical boundaries.

Summary updated

Abstract

Abstract We describe our connectomics pipeline for processing tracer injection data for the brain of the common marmoset ( Callithrix jacchus ). Brain sections were imaged using a batch slide scanner (NanoZoomer 2.0-HT) and we used artificial intelligence to precisely segment the anterograde tracer signal from the background in the fluorescence images. The shape of each brain was reconstructed by reference to a block-face and all data was mapped into a common 3D brain space with atlas and 2D cortical flat map. To overcome the effect of using a single template atlas to specify cortical boundaries, each brain was cytoarchitectonically annotated and used for making an individual 3D atlas. Registration between the individual and common brain cortical boundaries in the flat map space was done to absorb the variation of each brain and precisely map all tracer injection data into one cortical brain space. We describe the methodology of our pipeline and analyze tracer segmentation and brain registration accuracy. Results show our pipeline can successfully process and normalize tracer injection experiments into a common space, making it suitable for large-scale connectomics studies with a focus on the cerebral cortex.
BibTeX
@misc{woodward2019nanozoomer,
  author = {Alexander Woodward and Rui Gong and Hiroshi Abe and Ken Nakae and Junichi Hata and Henrik Skibbe and Yoko Yamaguchi and Shin Ishii and Hideyuki Okano and Tetsuo Yamamori and Noritaka Ichinohe},
  title = {{The NanoZoomer Connectomics Pipeline for Tracer Injection Studies of the Marmoset Brain}},
  year = {2019},
  doi = {10.1101/748376},
  howpublished = {bioRxiv},
}
Figure