Multi-Task Hybrid Conv-Transformer With Emotional Localized Ambiguity Exploration for Facial Pain Assessment.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Medical Informatics

Background:

• Automatic pain assessment using facial expressions shows promise but faces limitations.
• Current methods struggle with analyzing local pain-related facial action units and resolving emotional ambiguity.
• Accurate pain intensity estimation is complicated by ambiguous facial expressions.

Purpose of the Study:

• To propose a novel multi-task hybrid Conv-Transformer method for enhanced facial pain assessment.
• To improve the analysis of local facial regions crucial for pain intensity estimation.
• To mitigate emotional ambiguity in pain expression analysis.

Main Methods:

• A hybrid Convolutional Neural Network (CNN) and Vision Transformer (ViT) architecture was developed.
• The self-attention mechanism was employed to focus on pain-related local facial features.
• A multi-task joint optimization module was designed to address classification and regression tasks, mitigating ambiguity.

Main Results:

• The proposed method demonstrated superior performance in pain assessment compared to existing state-of-the-art techniques.
• Experimental validation was conducted on the UNBC Pain dataset.
• The multi-task approach effectively regularized features and improved prediction accuracy.

Conclusions:

• The developed multi-task hybrid Conv-Transformer method offers a significant advancement in automatic facial pain assessment.
• Addressing local facial features and emotional ambiguity leads to more precise pain intensity estimation.
• This approach holds potential for improving objective pain monitoring in clinical and research settings.