Deep Pain: Exploiting Long Short-Term Memory Networks for Facial Expression Classification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomedical Engineering

Background:

• Objective pain assessment is crucial for patient recovery but remains challenging due to reliance on subjective human evaluation.
• Current pain assessment methods primarily focus on facial features, limiting accuracy and objectivity.
• Automatic systems are needed to overcome the limitations of manual pain assessment.

Purpose of the Study:

• To develop an advanced deep learning model for objective pain assessment using raw video frames.
• To enhance pain detection accuracy by integrating facial feature analysis with temporal dynamics.
• To address the challenge of imbalanced data in pain expression datasets.

Main Methods:

• Utilized Convolutional Neural Networks (CNNs) pre-trained on VGG_Faces for facial feature extraction.
• Integrated Long Short-Term Memory (LSTM) networks to capture temporal relationships between video frames.
• Compared performance using canonical normalization versus whole image analysis for pain assessment.

Main Results:

• Achieved state-of-the-art performance in area under the curve on the UNBC-McMaster Shoulder Pain Expression Archive Database.
• Demonstrated superior results compared to existing methods relying solely on facial features.
• Reported competitive performance in facial motion recognition on the Cohn Kanade+ database.

Conclusions:

• Deep learning models fed with raw video frames significantly enhance objective pain assessment accuracy.
• The proposed CNN-LSTM approach effectively captures both spatial facial features and temporal dynamics for pain detection.
• The methodology shows strong generalization capabilities for facial motion recognition tasks.