An annotation dataset facilitates automatic annotation of whole-brain activity imaging of C. elegans

Yu Toyoshima, Stephen Wu, Manami Kanamori, Hirofumi Sato, Moon‐Sun Jang, Suzu Oe, Yuko Murakami, Takayuki Teramoto, ChanHyun Park, Yuishi Iwasaki, Takeshi Ishihara, Ryo Yoshida, Yuichi IinobioRxiv 2019 · DOI

TL;DR

This paper introduces a novel method for automatically annotating neurons in whole-brain imaging of C. elegans. By creating a comprehensive dataset of neuron positions and expression patterns from 311 worms and developing an atlas-generation approach, they significantly improve the accuracy of cell identification, paving the way for more efficient and robust neuroscience research.

What Problem Does This Paper Address?

Accurately identifying individual neurons is crucial for understanding neural activity and comparing findings across different animals. While C. elegans has a well-defined connectome, annotating neurons in live imaging data is challenging due to significant variations in cell positions between individual worms. Existing methods struggle with this variability, especially in dense regions like the head. This paper aims to overcome these limitations by developing a systematic approach to create a rich annotation dataset and an automated annotation tool that accounts for positional variations.

Methods

The authors developed a multi-pronged approach:

  • Dataset Creation: They systematically imaged the head neurons of 311 adult C. elegans worms using 35 cell-specific fluorescent promoters. This generated a dataset containing the 3D positions and expression patterns of identified neurons.
  • Atlas Generation: To account for positional variations, they created a large set of synthetic atlases (around 3000) by combining positional information from multiple partially annotated animals. This process involved registering animals to each other and interpolating displacement fields to maintain relative cell positions.
  • Automatic Annotation: They proposed an annotation method combining bipartite matching and majority voting. Bipartite matching assigns cells from a target animal to cells in an atlas based on features like position and promoter expression. Majority voting then uses the assignments from multiple atlases to determine the most likely cell identity, leveraging the collective "wisdom" of the atlases.
  • Human-in-the-Loop: The system is integrated into a GUI (RoiEdit3D) allowing for manual correction of annotations, which can then be used to improve the automatic estimation for other neurons.

Datasets

  • Newly Introduced Benchmark Dataset: The core contribution is a newly created annotation dataset comprising volumetric images of 311 adult C. elegans worms. This dataset includes the 3D positions and expression patterns of identified neurons, derived using 35 cell-specific promoters. This dataset is intended to serve as a benchmark for future cell identification methods.

Key Insights and Ideas

  • Positional Variation is a Major Hurdle: The study highlights the significant positional variability of neurons in adult C. elegans, demonstrating that a single reference atlas is insufficient for accurate annotation.
  • Synthetic Atlases Capture Variation: Generating a large ensemble of synthetic atlases, rather than relying on a single one, is an effective strategy to capture and account for the observed positional variations.
  • Majority Voting Enhances Accuracy: Combining bipartite matching with majority voting across multiple atlases significantly improves annotation accuracy compared to methods relying solely on mean/covariance of positions or a single atlas.
  • Human Feedback is Crucial: Integrating human corrections into the annotation pipeline allows the system to learn and improve, making the process more efficient and accurate.

Key Results

  • The developed annotation dataset contains positional and expression information for 175 identified cells in the head region of adult worms.
  • The automatic annotation method, using bipartite matching and majority voting, achieved an average top-rank estimation error rate of 41.4% on the cross-validation dataset.
  • Considering the top 5 voted estimations reduced the error rate to 7.3%.
  • When using fluorescent landmarks, the error rate for top-rank estimation was 37.7%, and for the top 5 voted estimations, it was 8.1%.
  • The system was applied to whole-brain imaging in a microfluidic chip, achieving error rates of 52.1% (top rank) and 12.2% (top 5), though distortion in the chip increased errors.

Limitations and Future Work

  • Annotation Accuracy: While improved, the fully automatic annotation error rate is still relatively high, particularly for fully automatic annotation. The authors suggest that increasing the number of animals and promoters used in the dataset, as well as the number of fluorescent channels, could further improve accuracy.
  • Information Utilization: The current automatic annotation algorithm does not fully utilize information like nuclear size and fluorescence intensity, which were used in manual annotation. The authors suggest that advances in artificial neural networks could leverage this information.
  • Dataset Size: While 311 animals is substantial, the authors note that a larger dataset would be ideal for training more robust models.
  • Future Work: The authors propose that their annotation dataset can serve as a benchmark for new cell identification methods. They also envision the application of their framework to other species and the integration of artificial neural networks for more advanced annotation. The interactive feedback system is highlighted as a key component for future improvements.

Summary updated

Abstract

Abstract Annotation of cell identity is an essential process in neuroscience that allows for comparing neural activities across different animals. In C. elegans , although unique identities have been assigned to all neurons, the number of annotatable neurons in an intact animal is limited in practice and comprehensive methods for cell annotation are required. Here we propose an efficient annotation method that can be integrated with the whole-brain imaging technique. We systematically identified neurons in the head region of 311 adult worms using 35 cell-specific promoters and created a dataset of the expression patterns and the positions of the neurons. The large positional variations illustrated the difficulty of the annotation task. We investigated multiple combinations of cell-specific promoters to tackle this problem. We also developed an automatic annotation method with human interaction functionality that facilitates annotation for whole-brain imaging.
BibTeX
@misc{toyoshima2019annotation,
  author = {Yu Toyoshima and Stephen Wu and Manami Kanamori and Hirofumi Sato and Moon‐Sun Jang and Suzu Oe and Yuko Murakami and Takayuki Teramoto and ChanHyun Park and Yuishi Iwasaki and Takeshi Ishihara and Ryo Yoshida and Yuichi Iino},
  title = {{An annotation dataset facilitates automatic annotation of whole-brain activity imaging of <i>C. elegans</i>}},
  year = {2019},
  doi = {10.1101/698241},
  howpublished = {bioRxiv},
}
Figure