Fully automated image-based estimation of postural point-features in children with cerebral palsy using deep learning

Area of Science:

• Biomedical Engineering
• Clinical Biomechanics
• Rehabilitation Technology

Background:

• The Segmental Assessment of Trunk Control (SATCo) test is crucial for evaluating motor control in children with cerebral palsy.
• Manual annotation of postural point-features in SATCo videos is time-consuming and labor-intensive.
• Accurate, automated analysis of trunk and limb posture is needed for objective clinical assessment.

Purpose of the Study:

• To develop an automated method for identifying postural point-features from SATCo videos.
• To estimate the location and orientation of the head, multi-segmented trunk, and arms.
• To enable objective and efficient assessment of trunk control in pediatric cerebral palsy.

Main Methods:

• Utilized convolutional neural networks (CNNs) trained on manually annotated video data.
• Employed linear interpolation to generate a large dataset of annotated images (30,825).
• Validated the model using cross-validation, providing test results for all participants.

Main Results:

• Automated point-feature estimation achieved an error of 4.4 ± 3.8 pixels at ~100 frames per second.
• Truncal segment angles were estimated with an error of 6.4 ± 2.8°, enabling accurate posture deviation classification (>80% F1 score).
• Upper limb contact with the support surface was classified with an F1 score of 80.5%.

Conclusions:

• Demonstrated the technical feasibility of automating the identification of segmental sitting posture.
• Showcased the ability to automatically detect deviations from reference posture and upper limb support.
• This automated approach offers a promising tool for objective and efficient clinical assessment in SATCo testing.