An accurately supervised motion-aware deep network for non-contact pain assessment of trigeminal neuralgia mouse model

Zhiheng Feng, Mingcai Chen, Jue Zhang, Xin PengJournal of Oral & Facial Pain and Headache 2024 · DOI

TL;DR

This paper introduces TNPAN, a novel deep learning model for accurately assessing pain in trigeminal neuralgia (TN) mouse models. By fusing static facial texture with dynamic behavioral cues and using an objective, physiologically-derived ground truth, TNPAN achieves high accuracy and generalization, addressing key limitations in current pain assessment methods.

What Problem Does This Paper Address?

Trigeminal neuralgia (TN) is a debilitating condition, and understanding its mechanisms and developing treatments relies heavily on animal models, particularly TN mouse models. However, reliably assessing pain in these models is a significant challenge. Existing methods often suffer from:

  • Subjective Supervision: Pain grading frequently relies on human annotators' subjective interpretations, leading to biased training data and models.
  • Ignoring Dynamics: Many approaches focus only on static facial features, neglecting crucial dynamic behavioral changes that indicate pain.
  • Limited Generalization: Models trained on specific mouse strains or conditions often fail to perform well on others.

This gap hinders progress in TN research, necessitating more objective, comprehensive, and generalizable pain assessment tools.

Methods

The researchers developed a novel deep neural network called the Trigeminal Neuralgia Pain Assessment Network (TNPAN). TNPAN employs a dual-channel feature fusion strategy, combining two key components:

  1. Motion Awareness Network (MAN): This network focuses on extracting dynamic behavioral characteristics. It uses a pre-trained ResNet-50 as a feature detector to identify key points of the mouse's pupils and forelimbs. These are crucial indicators, as pupil size changes and forelimb movements (e.g., rubbing the face) are known pain-related behaviors. MAN utilizes deconvolutional layers to output score maps for these key points.
  2. 3D Convolutional Neural Network (3D-CNN): This component extracts static texture characteristics from the mouse's facial expressions. It uses 3D convolutional kernels to capture spatial and temporal features that change with pain. The researchers conducted experiments to optimize the 3D-CNN's architecture, finding that a spatial kernel size of 3x3 and a temporal depth of 16 yielded the best results.

These two networks are cascaded, with MAN's output (key point coordinates) being fused with the 3D-CNN's features to produce the final pain assessment. A fixed model reuse (FMR) strategy is employed, where MAN's weights are frozen during TNPAN training, allowing the model to focus on learning the fusion and the 3D-CNN's parameters.

Datasets

The study introduced a novel, objective pain grading dataset as the ground truth for model training. This dataset was derived from physiological indicators rather than subjective human annotations. Specifically:

  • Infraorbital Nerve (IoN) Activity Assay: Mice were subjected to varying concentrations of NaCl solution to induce pain, and IoN activity was measured using fluorescence microscopy.
  • Facial Pain Threshold Assay: Mice were tested using von Frey filaments to determine their mechanical pain threshold.

These assays provided objective measures of pain intensity. For training the MAN component, a dataset of 9000 images with manually annotated pupil and forelimb key points was used. For the overall TNPAN training, 3000 images were selected from each of three pain levels (No Pain, Moderate Pain, High Pain) to balance the dataset.

Key Insights and Ideas

  • Objective Ground Truth is Crucial: Moving away from subjective human annotations to physiologically derived data (IoN activity, pain threshold) significantly improves model reliability and objectivity.
  • Dynamics Matter: Incorporating dynamic behavioral cues (pupil, forelimb movements) alongside static facial features is essential for accurate pain assessment, leading to a substantial accuracy improvement.
  • Fusion is Key: Neither static texture nor dynamic behavior alone is sufficient. The synergistic fusion of both through TNPAN's dual-channel architecture unlocks superior performance.
  • 3D-CNN Optimization is Important: The performance of the 3D-CNN component is sensitive to its hyperparameters, with specific spatial kernel sizes and temporal depths proving optimal for capturing facial pain features.
  • Generalization Potential: The model's ability to perform well on a different TN model (IoN-CCI) suggests its potential for broader application in pain research.

Key Results

  • TNPAN Achieves High Accuracy: TNPAN achieved an accuracy of 92.27%, significantly outperforming baseline models.
  • Superior Precision and Recall: TNPAN also demonstrated the highest precision (86.45%) and recall (92.48%) among the tested models.
  • MAN and 3D-CNN Contributions: Ablation studies showed that TNPAN's accuracy was 22.79% higher than using only MAN and 9.43% higher than using only 3D-CNN, highlighting the benefit of their integration.
  • MAN Accuracy: The MAN component demonstrated high accuracy in predicting key points, comparable to manual annotation, with errors mainly within a 3-pixel range.
  • 3D-CNN Performance: The optimal 3D-CNN configuration achieved 82.84% accuracy with a 3x3 spatial kernel and a temporal depth of 16.
  • Generalization to IoN-CCI Model: TNPAN successfully predicted pain progression in the IoN-CCI mouse model, showing consistency with facial pain threshold measurements.

Limitations and Future Work

  • Limitations:
    • While the study aimed for objectivity, the initial MAN training still relied on human-annotated key points.
  • The study focused on acute TN models; chronic pain assessment might require further validation.
  • The current dataset is specific to mouse models.
  • Future Work:
    • The authors suggest exploring the feasibility of applying this dual-channel feature fusion strategy to other species, such as primates (monkeys), for pain assessment.
  • Further investigation into chronic pain models and different pain etiologies could expand the model's applicability.
  • Developing fully automated annotation for MAN could further enhance objectivity.

Summary updated

Abstract

Pain assessment in trigeminal neuralgia (TN) mouse models is essential for exploring its pathophysiology and developing effective analgesics. However, pain assessment methods for TN mouse models have not been widely studied, resulting in a critical gap in our understanding of TN. With the rapid advancement of deep learning, numerous pain assessment methods based on deep learning have emerged. Nonetheless, these methods have some limitations: (1) insufficiently objective supervision signals for training, (2) failure to account for the dynamic behavioral characteristics of mouse models in the constructed models and (3) inadequate generalization ability of the models. In this study, we initially constructed an objective pain grading dataset as the ground truth for model training, which remedy the limitations of prior studies that relied on subjective evaluation as supervisory signals. Then we proposed a novel deep neural network, named trigeminal neuralgia pain assessment network (TNPAN), which fuses the static texture characteristics and dynamic behavioral characteristics of mouse facial expressions. The promising experimental results demonstrate that TNPAN exhibits exceptional accuracy and generalization capability in pain assessment.
BibTeX
@article{feng2024accurately,
  author = {Zhiheng Feng and Mingcai Chen and Jue Zhang and Xin Peng},
  title = {{An accurately supervised motion-aware deep network for non-contact pain assessment of trigeminal neuralgia mouse model}},
  journal = {Journal of Oral \& Facial Pain and Headache},
  year = {2024},
  doi = {10.22514/jofph.2024.008},
}
Figure