Label Uncertainty Suppression Based on Heterogeneous Siamese Network for Neonatal Pain Assessment in Uncontrolled Conditions

Songyang Lin, Huaiyu Zhu, Kai Tong, Yisheng Zhao, Shuohui Chen, YUN PAN

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Neonate pain, label uncertainty, facial expression, heterogeneous Siamese network, uncontrolled conditions

TL;DR: A heterogeneous Siamese network architecture mitigates neonatal pain image label uncertainty through integrated diverse network perspectives, facilitating robust pain assessment in uncontrolled, real-world settings.

Abstract: Reliable assessment of neonatal pain is essential for timely clinical intervention. However, the ambiguity and variability of facial expressions in real-world clinical environments pose significant challenges. Although deep learning-based approaches have shown promise in automated neonatal pain assessment (NPA), their performance is often hindered by label noise, annotation ambiguity, and limited labeled data. To address these issues, we propose an uncertainty-aware dual-branch heterogeneous Siamese network that suppresses label noise and enhances the robustness of NPA. The framework incorporates an uncertainty-guided rank regularization module and a pseudo-label correction strategy, enabling dynamic label refinement during training. Moreover, visual analysis based on Gradient-weighted Class Activation Mapping (Grad-CAM) demonstrates that the model learns pain-intensity-sensitive attention patterns, providing interpretability and practical guidance for clinical annotation, particularly under occlusion conditions. Extensive experiments on our neonatal pain dataset show that the proposed method achieves superior accuracy compared to existing state-of-the-art label correction approaches, indicating the effectiveness and reliability of our model for automated NPA in challenging real-world settings.

Track: 3. Imaging Informatics

Registration Id: MRNXHLKN9NB

Submission Number: 78