AUTOMATED HOME-CAGE MOUSE SLEEP STAGING: MINIMIZING YOLOV5 MISSED DETECTIONS FOR LONG-TERM BEHAVIORAL ANALYSIS

Daily behavioral observation of rodents is a critical criterion for assessing phenotypes, particularly in transgenic models such as Down syndrome mice, where deviations in sleep cycles compared to wild-type mice serve as significant indicators. Addressing the limitations of traditional long-term quantification methods, this study leverages computer vision technology to propose an automated analysis framework based on YOLOv5s. The primary objective is to investigate the impact of data preprocessing and quality augmentation on detection performance. Localized overexposure artifacts in nocturnal infrared videos often lead to severe miss rates in YOLOv5 mouse detection. This research utilizes Contrast-Limited Adaptive Histogram Equalization (CLAHE) to mitigate overexposure and enhance local contrast, successfully reducing the miss rate in extreme lighting by around 69% (from 80.7% to 11.7%). Based on the stabilized localization, we extract spatiotemporal features to calculate 3-second movement distributions as the primary criteria for sleep-wake staging. The proposed framework achieves a 96.3% accuracy relative to manual annotation. In conclusion, this study underscores the critical role of data quality enhancement in boosting deep learning model performance and provides a high-efficiency, non-invasive automated tool for long-term behavioral analysis.