Edge feature enhancement: Generating adversarial edge perturbations for preterm infant movement recognition

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Infant pose estimation in Neonatal Intensive Care Units (NICUs) is crucial for premature infant care but is hindered by poor lighting and limited data diversity.
• Existing models struggle with edge information extraction and irregular pixel distributions due to illumination and skin tone variations in preterm infants.

Purpose of the Study:

• To develop an advanced method for accurate infant pose estimation and limb movement recognition in challenging NICU environments.
• To address the limitations of current models in handling low-light conditions and diverse infant skin pigmentation.

Main Methods:

• Proposed the Generative Edge Guidance Network (GEGN) incorporating an autoencoder for edge reconstruction and an adversarial edge perturbation branch.
• Integrated a dual-loss framework combining cross-entropy and signal similarity loss to enhance feature representation and learning.
• Guided adversarial perturbations to focus on infant skin regions for improved feature extraction.

Main Results:

• Achieved a mean Average Precision (mAP) of 95.3% on the Skeleton-V1 dataset.
• Demonstrated superior performance and robustness in challenging NICU conditions through extensive experiments and visualizations.
• Successfully enhanced low-level features and increased source sample diversity for improved pose estimation.

Conclusions:

• The GEGN method significantly improves infant pose estimation accuracy and robustness in NICU settings.
• The proposed approach effectively handles challenges posed by poor illumination and varying skin tones in preterm infants.
• This advancement supports critical medical interventions and real-time nursing care for premature infants.