Using Machine Learning to Automate the Analysis of an Olfactory Habituation-Dishabituation Task in Mice

Improving the efficiency and accuracy of annotation and extraction of performance data from mouse behavioral tasks will improve both the throughput and scientific value of preclinical research.

Here, we present and validate an automated pipeline for the annotation and quantification of performance in a mouse olfactory habituation-dishabituation task, using a single side-view camera, resulting in occluded body parts. We created a pipeline for task analysis, combining DeepLabCut (DLC) for pose-estimation and SimBA, for behavioral classification to automatically quantify odor interaction (sniffing time) in a three-odor (water, familiar mouse social odor, novel mouse social odor) variant of the task. We used a subset of previously published, fully manually annotated datasets to train the models and unseen videos from the same study to validate the utility of our machine learning (ML) pipeline.

Our analysis pipeline estimated behavioral performance in the task with high accuracy, and the data produces similar technical and biological results to manual methods when analyzed by linear mixed modeling. Thus, we validated the utility of our customized pipeline for the automated scoring of this mouse sensory task, and we discuss its strengths and limitations.