Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Biomedical Imaging and Instrumentation

Biomedical Imaging and Instrumentation - Poster Session A

Poster G4 - Characterizing Spontaneous Movements During Early Development of Mice

Thursday, October 24, 2024

10:00 AM – 11:00 AM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Logan Cook

Summer Intern
National Institutes of Health
Madison, Alabama, United States

Co-Author(s)

Jonathan Krynitsky

Engineer
National Institutes of Health
Bethesda, Maryland, United States
LZ

Lu Zhang

Scientist
National Institute of Mental Health, United States
AL

Ana Lourenco Inacio

Postdoctoral Fellow
National Institute of Mental Health, United States
TP

Thomas J. Pohida

Lab Chief
National Institute of Biomedical Imaging and Bioengineering, United States

Primary Investigator(s)

SL

Soohyun Lee

Principle Investigator
National Institute of Mental Health, United States

Introduction: Cortical circuit formation emerges during early development, yet little is known about the developmental mechanisms driving the specificity of these circuits. Understanding when and how the distinct properties of cortical neurons emerges during this cortical development is one of the research focuses of the Unit on Functional Neural Circuits at the National Institute of Mental Health.
To determine the developmental emergence of the functional heterogeneity of individual cortical neurons that are behavior state-related, we conducted longitudinal chronic calcium imaging in the primary somatosensory cortex to follow the same population of neurons throughout postnatal stage from early development to adulthood. Together with longitudinal neuronal recording, we monitored the spontaneous movements of developing mice. To correlate individual neuronal activity with these spontaneous movements, comprehensive tracking analysis of spontaneous movement was required. The goal of this project is to explore image processing and computer vision methods for the characterization of movement during imaging. To this end, we systematically compared Social LEAP Estimates Animal Poses (SLEAP) and DeepLabCut (DLC) for their performance in body part pose estimation performance. In addition, we also developed motion energy methods to analyze the facial regions of the mouse where pose estimation methods provide insufficient data.

Materials and

Methods: Experiments involved using two-photon imaging to record and track the same population of pyramidal neurons over time in mouse pups. Pups were imaged across postnatal ages day 8 to 30, while four cameras with near-infrared illumination recorded the pups' behavior at 50 Hz. The datasets used here were acquired using the same mouse at postnatal days 9, 12, and 14 (two 154 second sessions on each day). An additional session from a different adolescent mouse was also added to enrich the dataset.
We compared the performance of body part pose estimation between two popular open-source deep-learning packages, DeepLabCut (DLC) and SLEAP. Twenty frames evenly sampled from nine videos, totaling 180 frames, were redundantly annotated by two independent researchers using the annotation tool provided by SLEAP. The keypoint constant used to calculate object keypoint similarity (OKS) was refined by determining the standard deviation of each keypoint across these labels. DLC was configured to use ResNet50, and SLEAP was configured to use the U-Net network. Default parameters were applied to both networks. Training involved one of the redundantly labeled sets, with testing and analysis conducted on a separate annotated dataset comprising labels from 180 randomly sampled frames.
In addition to pose tracking, we analyzed changes in motion energy in the face by developing methods to calculate the average pixel intensity change between two frames. Phase correction was necessary to mitigate artifacts caused by the differences in frequency between video capture and calcium fluorescent imaging.

Results, Conclusions, and Discussions: We found that the DeepLabCut and SLEAP packages resulted in comparable similarity scores across all videos; however, SLEAP’s integrated U-net architecture required less time to train and run inferences. Additionally, SLEAP’s labeling user interface was more intuitive and less error-prone than DLC’s.
While evaluating the performance of DLC and SLEAP, we observed that pose estimation software mostly detects large movements but cannot reliably capture smaller facial features that also indicate behavioral states. Therefore, we described the movement of facial features by computing the motion energy within five regions (eyes, nose, mouth, whisker pad, and overall face) across video frames. PCC analysis of face motion energy and keypoint velocities demonstrates that facial motion energy may provide additional information that cannot be extracted from body part tracking alone, thereby offering more comprehensive data for behavioral description.

Acknowledgements (Optional):