Environment-sensitive motion modelling in healthcare with synthetic retargeting

Area of Science:

• Computer Vision
• Medical Imaging
• Clinical Research

Background:

• Video-based motor behavior assessment is crucial in clinical settings.
• Data scarcity and privacy concerns limit the use of real-world clinical data for training AI models.
• Existing methods struggle with occlusions, varied views, and environmental factors.

Purpose of the Study:

• To develop a synthetic data generation framework to address data scarcity and privacy constraints in clinical video analysis.
• To enable robust human detection with high fidelity in challenging clinical scenarios.
• To improve video-based motor behavior assessment through advanced synthetic data techniques.

Main Methods:

• Employed synthetic data generation tailored to specific clinical environments.
• Implemented a novel synthetic retargeting approach using procedural image synthesis.
• Addressed limitations of training data due to privacy, constrained views, occlusions, and environmental variability.

Main Results:

• The synthetic retargeting approach significantly improved human detection performance on real-world clinical data.
• Performance gains of up to 19.4% were observed in challenging scenarios (p < 0.001).
• Demonstrated high fidelity and robustness of the method in diverse clinical environments.

Conclusions:

• Synthetic retargeting offers an efficient solution for adapting AI models to clinical settings.
• Generates scenario-tailored synthetic data, bypassing privacy and logistical hurdles.
• Enables robust video-based motor behavior quantification for clinical management and research.