Enhancing Pre-Clinical Behavioral Pain Assessment Through the Implementation of Deep Learning Models

The DeepView application employs supervised learning to train a model for the quantification of rodent behaviors observed in open-field chambers. Researchers scored rodent videos, providing precise examples for each behavior. Utilizing an optical flow technique, the models were trained with static JPG images extracted from videos, and preliminary investigations indicated improved accuracy by representing motion on images.
For the Gait.AI application, instance segmentation models were developed by labeling rodent images obtained from a transparent treadmill system. Masks were created for each body part and background. Considering the pixel size disparity among different body parts, an ensemble of individual models was trained by fine-tuning pre-existing models. This ensemble was then deployed to detect body parts on a frame-by-frame basis from high-speed videos, utilizing positional and pixel data to calculate various temporal, spatial, and relational gait features with high accuracy.