Related Experiment Video
Updated: Jul 26, 2026
08:25
Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment
Published on: May 7, 2019
Towards generalizable contactless oximetry from multispectral video via deep domain-adaptive learning strategy
Wang Liao1, Fengyuan Liang1, Chen Zhang1
1Department of Mechanical Engineering, Ilmenau University of Technology, Gustav-Kirchhoff-Platz 2, 98693 Ilmenau, Germany.
Biomedical Optics Express
|July 16, 2026
Summary
This study introduces a deep learning framework to improve camera-based oxygen saturation monitoring. The method enhances accuracy and reliability across diverse conditions, paving the way for practical clinical use.
Area of Science:
- Biomedical Engineering
- Computer Vision
- Machine Learning
Background:
- Camera-based contactless oximetry offers a comfortable and hygienic alternative for long-term oxygen saturation monitoring.
- Current deep learning models struggle with domain shifts, limiting generalization across different subjects, skin tones, spectral data, and environments.
Purpose of the Study:
- To develop a robust deep domain adaptation framework for contactless oximetry using multispectral facial video.
- To enhance the generalization and accuracy of camera-based oxygen saturation estimation across various domain shifts.
Main Methods:
- A multi-level target unsupervised domain adaptation framework was proposed, building on a 3D Convolutional Neural Network (CNN) baseline.
- The method aligns feature distributions across multiple network levels, incorporating a dynamic training scheduler and adaptive batch normalization.
- Systematic evaluation was performed on three custom datasets, including laboratory and clinical recordings with two different sensor systems.
Main Results:
- In leave-one-participant-out validation, the framework reduced mean absolute error (MAE) from 2.31% to 1.98% and improved Pearson's correlation coefficient from 0.64 to 0.71.
- Significant performance gains were observed across cross-skin-type, cross-spectral, and cross-sensor-system domain shifts.
- Clinical validation showed that estimations remained within 2% of polysomnography ground truth for 90% of recorded time in sleep apnea patients.
Conclusions:
- The proposed deep domain adaptation framework significantly improves the robustness and generalization of contactless oximetry.
- This advancement demonstrates the potential for practical deployment of camera-based oxygen saturation monitoring in clinical and real-world settings.
