Istudor Andrei, Schatz Alexej, Winter York
Humboldt-Universität zu Berlin, Berlin 10099, Germany
Humboldt-Universität zu Berlin, Berlin 10099, Germany.
eNeuro. 2025 Jul 15;12(7). doi: 10.1523/ENEURO.0369-24.2025. Print 2025 Jul.
Animal behavior is crucial for understanding both normal brain function and dysfunction. To facilitate behavior analysis of mice within their home environments, we developed DeepEthoProfile, an open-source software powered by a deep convolutional neural network for efficient behavior classification. DeepEthoProfile requires no spatial cues for either training or processing and is designed to perform reliably under real laboratory conditions, tolerating variations in lighting and cage bedding. For data collection, we introduce EthoProfiler, a mobile cage rack system capable of simultaneously recording up to 10 singly housed mice. We used 36 h of manually annotated video data sampled in 5 min clips from a 48 h video database of 10 mice. This published dataset provides a reference that can facilitate further research. DeepEthoProfile achieved an overall classification accuracy of over 83%, comparable with human-level accuracy. The model also performed on par with other state-of-the-art solutions on another published dataset ( Jhuang et al., 2010). Designed for long-term experiments, DeepEthoProfile is highly efficient-capable of annotating nearly 2,000 frames per second and can be customized for various research needs.
动物行为对于理解正常脑功能和功能障碍都至关重要。为了便于在小鼠的自然环境中进行行为分析，我们开发了DeepEthoProfile，这是一款由深度卷积神经网络驱动的开源软件，用于高效的行为分类。DeepEthoProfile在训练或处理过程中都不需要空间线索，并且设计为在真实实验室条件下可靠运行，能够容忍光照和笼内垫料的变化。为了进行数据收集，我们引入了EthoProfiler，这是一种移动笼架系统，能够同时记录多达10只单独饲养的小鼠。我们使用了从10只小鼠的48小时视频数据库中以5分钟片段采样的36小时手动标注视频数据。这个已发布的数据集提供了一个参考，有助于进一步的研究。DeepEthoProfile实现了超过83%的总体分类准确率，与人类水平的准确率相当。该模型在另一个已发布的数据集上（Jhuang等人，2010年）的表现也与其他最先进的解决方案相当。专为长期实验设计，DeepEthoProfile效率极高，能够每秒标注近2000帧，并且可以根据各种研究需求进行定制。
