Abstract Body: Introduction: Despite efforts to improve stroke outcomes in patients, a translational gap exists between preclinical and clinical studies. Due to this gap, the Stroke Preclinical Assessment Network has incorporated the corner test (CoT) for behavioral outcome as a primary measure for evaluating whether a treatment is successful or not. Standard behavioral analysis for CoT uses the laterality index to detect if there is mouse turning preference on a scale from -1 to 1. We sought to determine if a deep learning approach using “DeepLabCut” could be applied towards enriching our CoT data to better evaluate aspects of mouse locomotion.
Methods: Six C57/Bl6 mice were subjected to an 1 hr transient middle cerebral artery occlusion in the right hemisphere of the brain. CoT were recorded with an isometric view and performed at both baseline (BL) prior to the stroke and one day post stroke (D1). The same set of six mice performed 10 turns per CoT, totalling to 60 turns for BL and 60 turns for D1. The pose estimation model was made using a ResNet-101 neural network trained on 1064 manually-labeled frames, with the assistance of “DeepLabCut” software packages. Videos were analyzed by the pose estimation model and sequentially processed through a newly developed R script and DLC Analyzer R script. Turns were defined as in SPAN with a 90 degree head turn upon vibrissae contact on both sides of the corner boards. Turn latency was defined as the time lapsed during a turn, and head turn speed as the average speed during a turning event.
Results: The average laterality index showed clear preference towards ipsilateral turning in all D1 mice (-1.0 ± 0.04, N = 6). Furthermore, a comparison between BL data and respective D1 data showed significantly longer turn latencies and slower head turn speeds (p < 0.05) for D1 mice. The average turn latency for BL mice was 3.58 ± 0.57 s, which was 4.6 times shorter than that of D1 mice (16.45 ± 3.11 s). The average speed for BL mice was 2.01 cm/s ±0.21, which was 2.3 times faster than that of D1 mice (0.86 ± 0.14 cm/s).
Conclusion: This deep learning approach enriches current stroke behavioral analysis methods by offering additional quantitative information upon which behavior can be assessed. Future studies can use these behavioral metrics for stratification or correlation with variables of interest (e.g. infarct size) to provide a more refined assessment of preclinical stroke behavior.