NeuroPupil: A generalization-first framework for scalable and biologically informative cross-species pupillometry

Abstract

Quantitative pupillometry provides a noninvasive window into brain state and neurological function, but its broader use across experimental and clinical settings is limited by challenges in achieving accurate, scalable, and generalizable measurements. Here, we present NeuroPupil, a deep learning framework for high-throughput, cross-species pupillometry that emphasizes robust generalization across subjects, behavioral contexts, and imaging conditions. Through systematic benchmarking of training strategies and network architectures, we identify pooled multi-subject training combined with an optimized U-Net architecture as a key determinant of reliable and transferable pupil tracking performance. Across diverse mouse and human datasets, NeuroPupil achieves improved accuracy and substantial gains in computational efficiency compared to existing approaches, enabling practical analysis of large-scale datasets. We further demonstrate that improved pupil tracking fidelity enhances downstream biological inference: NeuroPupil-derived pupil features significantly improve prediction of distributed cortical activity in behaving mice and preserve diagnostically relevant temporal structure in human clinical recordings. These findings highlight the importance of precise and scalable measurement for linking pupil dynamics to brain activity and clinical phenotypes. By integrating benchmarking, scalability, and accessible software tools, NeuroPupil provides a reproducible framework for large-scale pupillometry and facilitates its application in systems and translational neuroscience.