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

Abstract

Infant pose estimation performance is degraded by domain gaps induced by constrained lighting and limited annotation diversity, particularly in private medical datasets in Neonatal Intensive Care Units (NICUs). Accurate diagnostic decisions rely heavily on Premature Infant Limb Movement Recognition (PI-LMR) to support critical medical interventions and real-time nursing care for premature infants. However, due to challenging lighting and the darker skin tones of preterm infants, existing models struggle to extract edge information and manage irregular pixel distributions. These issues are primarily driven by poor illumination and varying skin pigmentation. To address these challenges, the Generative Edge Guidance Network (GEGN) is proposed to enhance low-level features and increase source sample diversity. Using edge reconstruction via an autoencoder, a single-channel feature map is constructed to approximate the Laplacian edge with high fidelity. This edge perturbation feature is integrated into the infant pose estimation backbone through an adversarial edge perturbation branch, which injects informative perturbations into the learning process. Furthermore, a combination of cross-entropy loss and signal similarity loss is introduced to guide the supervised learning of global feature characteristics. The signal comparison and enhanced feature representation are improved by this dual-loss framework. Our method is validated on the Skeleton-V1 dataset, collected from Jiaxing Maternity and Child Health Care Hospital in China, achieving a mean Average Precision (mAP) of 95.3 %. In addition, adversarial perturbations are strategically guided to better focus on infant skin regions at the feature level. Extensive experiments and visualizations demonstrate that superior performance and robustness are achieved by our approach under the challenging conditions present in NICUs.