Comparative Evaluation of DeepLabCut Convolutional Neural Network Architectures for High-Precision Markerless...

Area of Science:

• Neuroscience
• Behavioral Science
• Machine Learning

Background:

• Markerless pose estimation using DeepLabCut (DLC) is crucial for analyzing fine motor behavior.
• Convolutional Neural Network (CNN) backbone choice impacts DLC tracking performance, especially with small joints and occlusions.

Purpose of the Study:

• Systematically compare nine CNN architectures within DLC for analyzing rodent reaching movements.
• Evaluate model performance based on accuracy, robustness, speed, and computational resource usage.

Main Methods:

• Utilized the Montoya Staircase test for skilled forelimb reaching in control and M1-lesioned mice.
• Implemented and compared nine different CNN architectures in DLC.
• Assessed performance using metrics like RMSE, PCK@5 px, mAP, occlusion robustness, inference speed, and GPU memory.

Main Results:

• Multi-scale DLCRNet architectures significantly outperformed conventional backbones.
• DLCRNet_ms5 demonstrated the highest overall spatial accuracy.
• DLCRNet_stride16_ms5 offered the best trade-off between precision and computational efficiency.

Conclusions:

• CNN architecture selection is critical for reliable fine motor behavior quantification in preclinical research.
• DLCRNet architectures provide enhanced performance for behavioral phenotyping.
• Findings offer practical guidance for neuroscience labs using DLC.