Abstract
Experimental animals including mice interacts with others and exhibits variety of behaviors. However, conventional behavioral tests mostly focused on single mouse behavior since visual tracking for multiple mice is practically impossible. Here, we aimed to develop the tracking tool for multiple mice using deep learning methods for image recognition. Behaviors of two to four C57BL/6 mice were recorded with handy camera in an open field arena. First, we manually labeled the mouse contours for hundreds of frame images and trained a deep learning model with labeled images. Next, mouse counters in all frames were predicted by the trained model, and assigned IDs by calculating similarities of every pair of contours between frames. Finally, we tracked the geometric center of contours that has the same IDs and semi-automatically corrected predictive errors to improve the performance. The established system could accurately track two to four mice under light conditions. In addition, we confirmed that this system accurately predicted the videos with bedding in the arena and could evaluate the videos recorded with infrared lights. This technology provides a new approach to evaluate mouse behaviors in pharmacological research.
