Contactless Estimation of Respiratory Frequency Using 3D-CNN on Thermal Images.

Area of Science:

• Physiological monitoring
• Medical imaging
• Artificial intelligence in healthcare

Background:

• Contactless monitoring of physiological parameters like respiratory rate (f$_{R}$) is crucial for diagnosing conditions.
• Traditional methods often involve invasive sensors or obtrusive wearables.
• Existing thermal imaging techniques for f$_{R}$ estimation require extensive pre-processing and manual region-of-interest (ROI) tracking, limiting practical use.

Purpose of the Study:

• To develop a deep learning-based method for direct f$_{R}$ estimation from raw thermal videos.
• To eliminate the need for complex pre-processing and ROI tracking in thermal imaging analysis.
• To enhance the feasibility of contactless respiratory rate monitoring in real-world applications.

Main Methods:

• A 3D Convolutional Neural Network (3D-CNN) was developed to process raw thermal video data.
• Data augmentation and transfer learning from synthetic datasets were employed to address small dataset challenges.
• The model was trained and validated on thermal video data for f$_{R}$ estimation.

Main Results:

• The proposed 3D-CNN method achieved a validation R$^{2}$ score of approximately 0.61.
• The model demonstrated effectiveness on both pre-processed and raw thermal videos.
• The deep learning approach simplified the workflow for f$_{R}$ estimation.

Conclusions:

• The developed deep learning method enables direct f$_{R}$ estimation from thermal videos, bypassing traditional pre-processing and ROI tracking.
• This contactless approach shows promise for improving the practicality of thermal imaging in remote healthcare and automotive driver monitoring.
• The study highlights the potential of AI in advancing non-invasive physiological monitoring techniques.