Comparative Evaluation of DeepLabCut Convolutional Neural Network Architectures for High-Precision Markerless Tracking in the Mouse Staircase Test

Precise quantification of fine motor behaviour is essential for understanding neural circuit function and for evaluating therapeutic interventions in neurological disorders. While markerless pose estimation frameworks such as DeepLabCut (DLC) have transformed behavioural phenotyping, the choice of convolutional neural network (CNN) backbone has a critical impact on tracking performance, particularly in tasks involving small distal joints and frequent occlusions. In this study, we present the first systematic comparison of nine CNN architectures implemented in DLC for lateral-view analysis of skilled reaching movements in the Montoya Staircase test, a gold-standard assay for forelimb dexterity in rodent models of stroke and neurodegenerative disease. Using a dataset comprising both control and primary motor cortex (M1)-lesioned mice, we evaluated model performance across six key dimensions: spatial accuracy (RMSE, PCK@5 px), mean average precision (mAP), robustness to occlusions, inference speed, and GPU memory usage. Our results demonstrate that multi-scale DLCRNet architectures substantially outperform conventional backbones. DLCRNet_ms5 achieved the highest overall accuracy, while DLCRNet_stride16_ms5 provided the most favourable balance between precision and computational efficiency. These findings provide practical methodological guidance for neuroscience laboratories and highlight the importance of CNN architecture selection for the reliable quantification of fine motor behaviour in preclinical research.