Towards human-level performance on automatic pose estimation of infant spontaneous movements

Area of Science:

• Medical imaging and biomechanics
• Pediatric neurology and developmental science
• Artificial intelligence in healthcare

Background:

• Spontaneous infant movements are crucial indicators for predicting long-term developmental disorders in high-risk infants.
• Accurate infant pose estimation is essential for developing automated prediction algorithms.
• Existing methods lack the precision required for reliable clinical application.

Purpose of the Study:

• To develop and evaluate convolutional neural networks for precise infant pose estimation.
• To assess the localization performance and computational efficiency of these networks.
• To determine the feasibility of automated infant pose estimation in clinical practice.

Main Methods:

• Trained and evaluated four types of convolutional neural networks on a novel dataset of 1424 infant videos.
• Assessed localization performance by comparing estimated keypoint positions with human expert annotations.
• Evaluated computational efficiency for clinical practice feasibility.

Main Results:

• The best-performing neural network achieved localization error comparable to human expert inter-rater variability.
• The network demonstrated efficient computational performance, suitable for clinical settings.
• Pose estimation successfully quantified infant spontaneous movements with human-level accuracy.

Conclusions:

• Infant pose estimation using convolutional neural networks shows significant potential for early detection of developmental disorders.
• This technology can support research on children with perinatal brain injuries by quantifying movements from video.
• Automated pose estimation offers a promising tool for objective assessment of infant development.