Deep Convolutional Neural Networks for Automated Convulsion Scoring using RGB-D Images

This paper presents a novel approach for mouse behavior recognition in convulsion experiments using deep neural networks with RGB-D images provided by a Microsoft Kinect sensor. Depth frames are used to track the animal in a computationally efficient way during preprocessing phase. Instead of hand-crafting certain features and developing image processing algorithm for calculating them, we transfer the knowledge learned by convolutional neural networks (CNN) trained on millions of labeled natural images to generate deep visual features of input mouse images and classify different behaviors for mouse convulsion scoring. The proposed method was tested with data collected from 14 mice in convulsion experiment. Our best networks, trained with real-time data augmentation, achieved an overall accuracy of 89.8% on the testing images and showed significant improvements over conventional SVM-based method.