Poster K2 - Automated Pose Estimation of Optically Stimulated Rat Hindlimbs Using DeepLabCut

Thursday, October 24, 2024

10:00 AM - 11:00 AM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Logan Read (he/him/his)

Undergraduate Student
Marquette University
Milwaukee, Wisconsin, United States

Co-Author(s)

Emma Moravec

Ph.D Student
Marquette University and Medical College of Wisconsin, United States

Last Author(s)

Jordan Williams, MD, PhD

Principle Investigator
Marquette University and Medical College of Wisconsin
Milwaukee, Wisconsin, United States

Introduction: Compared to electrical stimulation, optogenetic nerve stimulation produces a more natural recruitment order of muscle fibers, a resulting decrease in fatigue, and target specificity as previously shown in rodent and non-human primate models. However, a tradeoff results from the need to first biologically label target nerves with light sensitive opsins via viral methods, a process which may be highly variable in its time course. Electromyography (EMG) or force measurements may be used to track this time course and characterize muscle responses and nerve sensitivity to optical stimulation, but these measures may be invasive or noisy from session to session due to variability in sensor placement. As an alternative, DeepLabCut’s markerless pose estimation toolbox can be used to automatically extract stimulated joint movements from video data using deep neural network models trained on manually labeled video frames. Using this framework, short- and long-term characterization of optically stimulated joint kinematics can be quantified accurately and efficiently with less experimenter bias and variability. Additionally, the models can be expanded to track more sophisticated movements, at the cost of a larger amount of more detailed training data. This study aimed to use joint angles extracted from video of optically stimulated hindlimb movements by DeepLabCut in conjunction with EMG and force data to more completely characterize the movements, which may then be used to inform more complicated stimulation patterns. Furthermore, the analysis can potentially act as a noninvasive, qualitative assessment of movements.

Materials and

Methods: DeepLabCut pose estimation models were trained on videos of Fischer344 rats previously injected in the tibialis anterior muscle with various adeno-associated viral vectors (AAVs) targeting channelrhodopsin expression to the peroneal nerve. The body parts chosen to be labeled were the toe tips, toe base, ankle, knee, and calf. Testing of anesthetized rats’ muscle responses to optical stimulation with varying light pulse train frequency, pulse width, and light intensity parameters were conducted on a weekly basis after virus injection. For testing including pose estimation, video was captured while EMG data was recorded using two needle electrodes placed in the subject’s tibialis anterior muscle. Force data was collected with identical testing parameters during a different session without video recording, since the transducer prevented any visible joint movement. The model was initially trained for 500,000 iterations on 120 manually labeled frames taken from 6 videos with slightly varying orientation and lighting qualities. As more videos were collected, any instances of incorrect tracking/labeling had a few frames of the erroneous video extracted, corrected, and then retrained to account for any additional changes in environment/video quality to improve the model’s robustness and overall performance. To monitor the model’s performance, a combination of manual labeling review, interpretation of label likelihood metrics, and correlation with EMG and force recordings were examined.

Results, Conclusions, and Discussions: The model developed to track leg movements during optical stimulation of the peroneal nerve accurately quantified the angle of the ankle when compared to manually measured angles. Likelihoods of label placement were all greater than 90%, indicating consistency despite variations in camera orientation and lighting. Estimated deflections in ankle angle with stimulation correlated strongly with EMG and force measurements within a session, suggesting this information can similarly be used to characterize muscle recruitment properties to optical stimulation. Furthermore, longitudinal data collected across the virus incubation period demonstrated considerable variability in EMG responses from week to week, likely due to variability in electrode placement. However, limb kinematics taken from videos recorded during each experiment were more consistent from week to week, providing smoother trends in nerve optical sensitivity and presumed opsin expression over weeks.
These results demonstrate that a DeepLabCut model can be used to characterize and track optically stimulated lower hindlimb responses. Rather than implementing motion tracking as an alternative to EMG, this method can be used in a complementary fashion to further characterize movements. In the future, this method may be expanded to characterize multi-joint kinematic responses to stimulation of multiple muscles or leveraged to assess responses to optical stimulation during awake, behaving locomotion. Additionally, real-time extraction of limb kinetics could prove useful as feedback in closed-loop optical stimulation. Overall, these results provide strong motivation for the continued use of automatic pose estimation in experiments examining peripheral stimulation of motor function.

Acknowledgements (Optional):