Gait-level analysis of mouse open field behavior using deep learning-based pose estimation

Keith Sheppard, Justin Gardin, Gautam Sabnis, Asaf Peer, Megan Darrell, Sean Deats, Brian Q. Geuther, Cathleen Lutz, Vivek KumarbioRxiv 2020 · DOI

TL;DR

This paper introduces a novel, deep learning-based approach for analyzing mouse gait and whole-body posture in an open field setting. By leveraging modern neural networks for pose estimation, the method provides a simple, accurate, and scalable way to extract detailed behavioral metrics, enabling new insights into neurogenetics and disease models.

What Problem Does This Paper Address?

Studying gait and posture in rodent models is crucial for understanding neurological, neuromuscular, and psychiatric disorders. However, existing methods for mouse gait analysis are often tedious, require specialized equipment, and are limited in their ability to capture natural, unforced behavior. This technological gap has hindered the translation of human gait analysis insights to rodent models and limited the exploration of gait as a phenotype in behavioral neurogenetics. The authors aim to bridge this gap by developing an accessible, high-throughput method that can analyze gait in free-moving mice within a standard open-field apparatus.

Methods

The core of the method is a deep convolutional neural network (CNN), specifically an HRNet-W32 architecture, adapted for mouse pose estimation. This network takes top-down video of mice in an open field and outputs twelve 2D keypoints representing the mouse's anatomical locations (e.g., nose, ears, paw joints, tail). This pose information is then used to:

  • Stride Inference: Identify individual strides by analyzing paw speed and transitions between stance and swing phases, focusing on hind paws due to better visibility.
  • Whole Body Posture Estimation: Extract metrics like lateral displacement and phase of the tail and nose during the gait cycle.
  • Statistical Analysis: Employ Linear Mixed Models (LMMs) to account for within-subject variation and analyze gait parameters, controlling for factors like body length, speed, and age. For circular gait metrics, a circular-linear regression model is used.
  • Strain Clustering: Utilize k-means clustering and Principal Component Analysis (PCA) to group strains based on their gait phenotypes and visualize the resulting clusters.
  • Genome-Wide Association Study (GWAS): Conducted to identify quantitative trait loci (QTLs) associated with gait and posture traits, using a linear mixed model approach.

Datasets

  • Training Data: 8,910 grayscale frames from diverse mouse strains, manually annotated with twelve keypoints.
  • Validation Data: 1,000 frames from the training set.
  • Strain Survey Data: Data from 44 classical inbred strains, 7 wild-derived inbred strains, and 11 F1 hybrid strains (totaling 1898 animals, 1740 hours of video). This dataset was used for exploring the range of gait phenotypes and for clustering strains.
  • Genetic Mutant Data: Data from mice with known gait deficits (Rett syndrome, ALS, Down syndrome models) and autism spectrum disorder (ASD) models.

Key Insights and Ideas

  • Open Field Gait Analysis is Feasible and Informative: The study demonstrates that detailed gait and posture analysis can be effectively performed in the ethologically relevant setting of an open field, overcoming previous limitations of specialized equipment or unnatural environments.
  • Deep Learning for Mouse Pose is Robust: The HRNet architecture, adapted for mouse pose, shows high accuracy and robustness across diverse strains, including variations in coat color, size, and fur.
  • Whole Body Coordination is a Key Phenotype: Beyond traditional stride metrics, measures of whole-body coordination (like lateral displacement of the tail and nose) reveal significant strain-specific differences and are highly heritable.
  • ASD Models Show Consistent Gait Deficits: The research suggests that gait and posture deficits are a general feature of autism spectrum disorder, observed across multiple ASD mouse models.
  • Gait Phenotypes are Highly Heritable: The large strain survey reveals that gait and posture measures exhibit significant heritability, providing a strong foundation for genetic studies.
  • C57BL/6J is Distinct: The reference C57BL/6J strain and its related strains form a distinct cluster, differing significantly in tail and nose movement compared to other laboratory and wild-derived strains.

Key Results

  • The developed method achieves high precision and accuracy in gait and posture analysis across 62 diverse mouse strains.
  • The approach successfully characterized gait deficits in four genetic mutants and identified gait and posture deficits as a common feature in multiple ASD models.
  • A large strain survey of 1898 mice revealed that gait and whole-body posture measures fall into three distinct classes and are highly heritable.
  • GWAS identified 346 significant genomic regions (QTLs) associated with gait and posture phenotypes, indicating a complex genetic architecture.
  • The study provides one of the largest laboratory mouse gait-level data resources to date.

Limitations and Future Work

  • 2D Pose Limitation: The current method relies on top-down video, providing a 2D representation of pose. This limits the measurement of kinetic properties of gait and may miss nuances captured by 3D pose estimation.
  • Forepaw Occlusion: Forepaw keypoints are often occluded in top-down views, leading to less accurate estimates and their exclusion from analysis. However, hind paw data proved sufficient for detecting robust differences.
  • High-Speed Behaviors: The 30 Hz frame rate may not be sufficient for accurately analyzing very high-speed behaviors like galloping.
  • Future Work: The authors suggest incorporating multiple camera angles and higher frame rates for future iterations to capture more detailed kinematic information. They also envision this method commoditizing gait analysis for behavioral neurogenetics, enabling broader research applications.

Summary updated

Abstract

1 Abstract Gait and whole body posture are sensitive measures of the proper functioning of numerous neural circuits, and are often perturbed in many neurological, neuromuscular, and neuropsychiatric illnesses. Rodents provide a tractable model for elucidating disease mechanisms and interventions, however, studying gait and whole body posture in rodent models requires specialized methods and remains challenging. Here, we develop a simple assay that allows adoption of the commonly used open field apparatus for gait and whole body posture analysis. We leverage modern neural networks to abstract a mouse into keypoints and extract gait and whole body coordination metrics of the animal. Gait-level analysis allows us to detect every step of the animal’s movement and provides high resolution information about the animal’s behavior. We quantitate gait and whole body posture with high precision and accuracy across 62 highly visually diverse strains of mice. We apply our approach to characterize four genetic mutants with known gait deficits. In extended analysis, we demonstrate that multiple autism spectrum disorder (ASD) models show gait and posture deficits, implying this is a general feature of ASD. We conduct a large strain survey of 1898 mice, and find that gait and whole body posture measures are highly heritable in the laboratory mouse, and fall into three classes. Furthermore, the reference mouse strain, C57BL/6J, has a distinctly different gait and posture compared to other standard laboratory and wild-derived strains. We conduct a genome wide association study (GWAS) to define the genetic architecture of mouse movement in the open field. In sum, we describe a simple, sensitive, accurate, scalable, and ethologically relevant method of mouse gait and whole body posture analysis for behavioral neurogenetics. These results provide one of the largest laboratory mouse gait-level data resources for the research community and show the utility of automated machine learning approaches for deriving biological insights.
BibTeX
@misc{sheppard2020gait,
  author = {Keith Sheppard and Justin Gardin and Gautam Sabnis and Asaf Peer and Megan Darrell and Sean Deats and Brian Q. Geuther and Cathleen Lutz and Vivek Kumar},
  title = {{Gait-level analysis of mouse open field behavior using deep learning-based pose estimation}},
  year = {2020},
  doi = {10.1101/2020.12.29.424780},
  howpublished = {bioRxiv},
}
Figure