Classifying rodent behaviors one frame at a time: A CNN-based method

Abstract

The classification of animal behavior is essential in disciplines such as pharmacology, neuroscience, and ecology. Traditional methods are labor-intensive, time-consuming, and dependent on human observation, while many existing automated approaches rely on pose estimation and temporal models that demand extensive data, annotations, and computational resources. This study proposes a non-temporal, frame-by-frame deep learning approach for classifying rodent behaviors directly from annotated video frames, without relying on pose estimation, body-part segmentation, or temporal context, and requiring only frame-level annotations. We constructed a large dataset of images representing distinct rodent behaviors in operant conditioning chambers to evaluate multiple convolutional neural network (CNN) architectures of various sizes. Our results show that deep CNNs achieve high classification performance in terms of accuracy and F1-score. The models maintain strong performance under challenging conditions, including restricted training data, noisy labels, and low subject diversity, demonstrating their data efficiency and resilience. Performance remains stable even when trained on data from a limited number of animals, supporting the method’s practicality for population-level behavior classification. Together, these results suggest that robust behavior classification can be achieved without temporal context, even under realistic experimental constraints such as limited data availability and label noise. The proposed framework also offers flexibility across different data formats, including static images and partial video recordings. Overall, our approach provides an accessible and scalable alternative for behavioral researchers, lowering barriers to automated animal behavior analysis while preserving strong classification performance.