Medical application of deep-learning-based head pose estimation from RGB image sequence.

Area of Science:

• Computer Vision
• Medical Technology
• Artificial Intelligence

Background:

• Telemedicine addresses challenges in healthcare access, including remote consultations and patient mobility limitations.
• Measuring head movement is crucial for daily activities but is often impaired by aging, injury, or disease.
• Existing vision-based methods for cervical range of motion (CROM) lack accuracy and require specialized equipment.

Purpose of the Study:

• To develop and evaluate a novel deep neural network for precise head pose estimation (HPE) for telemedicine applications.
• To apply the developed HPE technique for accurate CROM measurements in a clinical setting.
• To offer a computationally efficient and cost-effective solution for remote CROM assessment.

Main Methods:

• A deep neural network incorporating multi-level pyramidal feature extraction and a Pyramidal Feature Aggregation Structure (PFAS).
• A modified Atrous Spatial Pyramid Pooling (ASPP) module for enhanced feature representation.
• A multi-bin classification and regression module to derive Euler angles for head pose parameters.

Main Results:

• The model achieved comparable performance on public HPE datasets (mean MAE: 2.16°-3.50°).
• On a private medical dataset, the method yielded the lowest mean absolute error (MAE) of 3.73° for CROM measurement.
• The model demonstrated a fast inference speed of 2.27 ms per image.

Conclusions:

• The proposed deep learning approach offers an accurate and efficient method for head pose estimation.
• This technique is suitable for CROM measurement in telemedicine, overcoming limitations of current methods.
• The solution provides accuracy, convenience, and low operational costs for remote healthcare applications.