Using deep learning to measure movement impairments
Our PLOS Biology paper explores how video models can score motor impairments and identify movement features that conventional assessments might miss.
Last month, a paper we provided technical assistance on Data-driven analyses of motor impairments in animal models of neurological disorders, was published in PLOS Biology. I’m the third author, for contributing the software.
Assessing movement impairments in laboratory animals takes time and trained observers. Researchers watch videos and score individual movements, such as a rat reaching for a food pellet. Those assessments help them study the effects of neurological injury and track recovery.
The authors explored whether a DNN could learn those same features from video. That’s where our Five video classification methods… contributions were used. The architecture is Inception V3 as feature extractor, with RNNs for temporal movement detections.
In rodent stroke experiments, the network’s disagreement with an expert was within the range observed among trained human scorers. More interestingly, its scores were more closely associated with stroke lesion volume than the expert’s scores in this dataset. Examining what the network used to make its predictions helped identify informative movement patterns beyond the standard scoring categories.
The paper is open access, and the team has shared code and example data.