IntegraPose: A unified framework for simultaneous pose estimation and behavior classification

Area of Science:

• Computational ethology
• Animal behavior analysis
• Machine learning in biology

Background:

• Fragmented computational ethology pipelines hinder synchronized pose and behavior analysis.
• Existing methods often separate pose estimation from behavior classification, creating workflow inefficiencies.

Purpose of the Study:

• To introduce IntegraPose, an integrated, GUI-driven toolkit for unified pose estimation and behavior classification.
• To benchmark IntegraPose against established models like DeepLabCut for accuracy, stability, and speed.
• To leverage high-throughput analysis for discovering novel behavioral phenotypes.

Main Methods:

• Developed IntegraPose using a single YOLOv11x-Pose model to combine pose estimation and behavior classification.
• Benchmarked IntegraPose against DeepLabCut, evaluating keypoint accuracy, prediction jitter, and processing speed (FPS).
• Applied the toolkit to analyze motor behaviors in neuropathy-prone C57BLKS/J (BLKS) mice.

Main Results:

• IntegraPose demonstrated comparable keypoint accuracy to DeepLabCut.
• IntegraPose significantly reduced prediction jitter, enhancing temporal stability.
• IntegraPose achieved substantially faster processing speeds (72-73 FPS vs. 32 FPS).
• A context-dependent motor deficit was identified in BLKS mice during wall-rearing but not locomotion.

Conclusions:

• IntegraPose offers a streamlined, efficient, and accurate solution for computational ethology.
• The toolkit's high-throughput capability facilitates the discovery of subtle, context-specific behavioral deficits.
• IntegraPose promotes hypothesis-driven research by providing synchronized pose and behavior data.