Pupil-DLC: an open-source deep learning pipeline for scalable, markerless tracking of pupil dynamics across conscious and unconscious states
Journal:
bioRxiv
Published Date:
Jan 21, 2026
Abstract
Pupil diameter provides a powerful, non-invasive biomarker of brain state, correlating with arousal, attention, cognitive processing, and level of consciousness. Despite its widespread use, pupillometry remains limited by software tools that lack scalability, robustness, and flexibility across experimental conditions. Here we introduce Pupil-DLC, an open-source, DeepLabCut-based pipeline for scalable, markerless tracking of pupil dynamics. Pupil-DLC is trained on 19,500 manually annotated frames selected from over 130 pupil videos of head-fixed mice spanning wakefulness to diverse drug-induced states of consciousness, including psychedelics, and anesthesia. The pipeline robustly captures state-dependent pupil dynamics, achieving high agreement with human-annotated ground truth data and outperforming an established automated method in both accuracy and detection reliability, while maintaining computational efficiency. Pupil-DLC incorporates a dual-model framework comprising a General Model (GM) for high-throughput analysis of infra-red recorded mice pupils and an Individual Model (IM) tailored for session-specific optimization, as well as interpretable confidence metrics that enable principled trade-offs between accuracy and data retention. Notably, the model generalizes, without retraining, to human infrared pupil recordings, supporting cross-species pupillometry. Pupil-DLC provides a flexible, reproducible platform for quantifying pupil-linked brain state dynamics across diverse experimental paradigms and species.
