PMotion: an advanced markerless pose estimation approach based on novel deep learning framework used to reveal neurobehavior

Abstract

Abstract Objective. The evaluation of animals’ motion behavior has played a vital role in neuromuscular biomedical research and clinical diagnostics, which reflects the changes caused by neuromodulation or neurodamage. Currently, the existing animal pose estimation methods are unreliable, unpractical, and inaccurate. Approach. Data augmentation (random scaling, random standard deviation Gaussian blur, random contrast, and random uniform color quantization) is adopted to augment image dataset. For the key points recognition, we present a novel efficient convolutional deep learning framework (PMotion), which combines modified ConvNext using multi-kernel feature fusion and self-defined stacked Hourglass block with SiLU activation function. Main results. PMotion is useful to predict the key points of dynamics of unmarked animal body joints in real time with high spatial precision. Gait quantification (step length, step height, and joint angle) was performed for the study of lateral lower limb movements with rats on a treadmill. Significance. The performance accuracy of PMotion on rat joint dataset was improved by 1.98, 1.46, and 0.55 pixels compared with deepposekit, deeplabcut, and stacked hourglass, respectively. This approach also may be applied for neurobehavioral studies of freely moving animals’ behavior in challenging environments (e.g. Drosophila melanogaster and openfield-Pranav) with a high accuracy.