Data_Sheet_1_Automated Grooming Detection of Mouse by Three-Dimensional Convolutional Neural Network.PDF

Grooming is a common behavior for animals to care for their fur, maintain hygiene, and regulate body temperature. Since various factors, including stressors and genetic mutations, affect grooming quantitatively and qualitatively, the assessment of grooming is important to understand the status of experimental animals. However, current grooming detection methods are time-consuming, laborious, and require specialized equipment. In addition, they generally cannot discriminate grooming microstructures such as face washing and body licking. In this study, we aimed to develop an automated grooming detection method that can distinguish facial grooming from body grooming by image analysis using artificial intelligence. Mouse behavior was recorded using a standard hand camera. We carefully observed videos and labeled each time point as facial grooming, body grooming, and not grooming. We constructed a three-dimensional convolutional neural network (3D-CNN) and trained it using the labeled images. Since the output of the trained 3D-CNN included unlikely short grooming bouts and interruptions, we set posterior filters to remove them. The performance of the trained 3D-CNN and filters was evaluated using a first-look dataset that was not used for training. The sensitivity of facial and body grooming detection reached 81.3% and 91.9%, respectively. The positive predictive rates of facial and body grooming detection were 83.5% and 88.5%, respectively. The number of grooming bouts predicted by our method was highly correlated with human observations (face: r = 0.93, body: r = 0.98). These results highlight that our method has sufficient ability to distinguish facial grooming and body grooming in mice.

梳理行为（grooming）是动物照料毛发、维持卫生状态与调节体温的常见行为。由于应激源、基因突变等多种因素会从定量与定性层面影响梳理行为，因此对梳理行为进行评估，对于了解实验动物的生理状态至关重要。然而，现有梳理行为检测方法不仅耗时耗力，还需配备专用设备；此外，这类方法通常无法区分诸如洗脸、舔舐身体这类精细的梳理行为亚型。本研究旨在借助人工智能开展图像分析，开发一种可区分面部梳理与身体梳理的自动化梳理行为检测方法。实验中使用标准手持摄像机记录小鼠的行为，随后我们逐帧细致观察视频，并将每一时间点标注为面部梳理、身体梳理或非梳理行为。我们构建了三维卷积神经网络（3D-CNN），并使用标注后的图像对其进行训练。由于训练完成的3D-CNN输出结果中会包含不合理的短时长梳理片段与行为中断片段，因此我们设置了后置滤波流程以剔除此类干扰。我们使用未参与训练的首次检视数据集（first-look dataset）对训练后的3D-CNN与滤波流程的整体性能进行了评估。结果显示，面部梳理与身体梳理检测的灵敏度分别达到81.3%与91.9%，两类检测的阳性预测值分别为83.5%与88.5%。本方法预测的梳理片段数量与人工观测结果具有极高的相关性（面部梳理：r=0.93，身体梳理：r=0.98）。上述结果表明，本方法具备足够的能力区分小鼠的面部梳理与身体梳理行为。