Difficult Airway Assessment Based on Multi-View Metric Learning

Jinze Wu, Yuan Yao, Guangchao Zhang, Xiaofan Li, Bo PengBioengineering 2024 · DOI

TL;DR

This paper introduces DMF-Net, a novel deep learning model that leverages multi-view facial images and metric learning to predict difficult airways with improved accuracy and reliability. By fusing information from frontal and lateral views and employing specialized loss functions, the model aims to provide a more robust preoperative assessment tool for anesthesiologists.

What Problem Does This Paper Address?

Anesthesiologists face a significant challenge in identifying patients with difficult airways, which can lead to serious complications during intubation. While clinical bedside screening tests exist, they are often subjective and lack comprehensive accuracy. Existing AI-based methods have shown promise but still struggle with low recognition accuracy and reliability, particularly in fully utilizing the rich, complementary information available from multiple patient views. This paper aims to bridge this gap by developing a more accurate and dependable AI system for difficult airway assessment using multi-view facial analysis.

Methods

The core of this work is the Dual-Path Multi-View Fusion Network (DMF-Net). This network employs a dual-path architecture to process frontal and lateral facial views separately, allowing for specialized feature extraction. Key components include:

  • Feature Extractor Module: Based on a modified ResNet18, this module extracts multi-scale features from both frontal and lateral views.
  • Multi-Scale Feature Fusion Module: This module synthesizes features from different scales to enhance representation.
  • Hybrid Co-Attention Module: This novel module combines channel and spatial attention mechanisms to help the model focus on relevant facial regions (mouth, neck) and important feature channels, while downplaying noise and irrelevant information.
  • Metric Learning with Custom Loss Functions: The paper introduces consistency loss and complementarity loss. Consistency loss ensures that information shared across different views of the same patient is aligned, while complementarity loss maximizes the unique information each view provides, thereby enriching the overall understanding of the airway status.
  • Focal Loss: Used to address the class imbalance between difficult and non-difficult airways in the dataset.

Datasets

The study utilized an observational dataset collected from 320 patients undergoing general anesthesia surgery at a hospital. After applying inclusion and exclusion criteria, 225 patients were included: 151 males and 74 females, with an average age of 52.4 years. The dataset comprised:

  • 200 patients with normal airways (88.9%)
  • 25 patients with difficult airways (11.1%)

Facial images were collected from five different views: frontal closed-mouth, frontal open-mouth, frontal tongue-protruding, lateral closed-mouth, and lateral closed-mouth chin-lift. The periocular area was mosaiced for privacy. The dataset was split into 80% for training and 20% for testing, with a five-fold cross-validation strategy employed for robust evaluation. Data augmentation techniques were applied to address data imbalance and improve generalization.

Key Insights and Ideas

  • Multi-view data is crucial: The paper strongly emphasizes that combining information from different facial views (frontal and lateral) is essential for a comprehensive and accurate difficult airway assessment.
  • Complementarity and Consistency are key: Simply fusing features isn't enough. Explicitly learning the complementary (unique) and consistent (shared) information across views through specialized loss functions significantly boosts performance.
  • Attention mechanisms enhance focus: The Hybrid Co-Attention Module effectively guides the model to concentrate on the most informative parts of the face, improving its ability to discern subtle differences indicative of difficult airways.
  • Addressing data challenges: The use of Focal Loss and targeted data augmentation strategies is vital for handling the inherent class imbalance and limited data size in medical imaging datasets.

Key Results

The proposed DMF-Net achieved promising results:

  • Accuracy: 77.92%
  • Specificity: 75.62%
  • Sensitivity: 82.50%
  • F1-score: 71.35%

These metrics represent a significant improvement over several state-of-the-art (SOTA) deep learning methods for difficult airway assessment, with absolute advantages of up to 10.00% in sensitivity and 7.96% in F1-score. The five-fold cross-validation showed stable performance with AUC values ranging from 0.73 to 0.82. Ablation studies confirmed the effectiveness of each proposed module (MFFM, HCAM) and loss function (complementarity, consistency) in improving the model's performance.

Limitations and Future Work

The authors acknowledge several limitations:

  • Sample bias: Potential bias due to irregular criteria used by physicians for capturing facial images.
  • Complex background noise: Images often contain distracting background elements.
  • Limited dataset size: The overall quantity of data, especially for difficult airways, is a constraint.

For future work, the researchers plan to:

  • Standardize data collection: Develop clearer criteria for capturing patient images.
  • Expand the dataset: Include a more diverse range of airway difficulty cases.
  • Incorporate multi-modal data: Integrate patient height, weight, medical history, and audio data to further enhance generalization and robustness.
  • Develop a mobile deployment system: Aim to create a difficult airway management system deployable on mobile terminals.

Summary updated

Abstract

The preoperative assessment of difficult airways is of great significance in the practice of anesthesia intubation. In recent years, although a large number of difficult airway recognition algorithms have been investigated, defects such as low recognition accuracy and poor recognition reliability still exist. In this paper, we propose a Dual-Path Multi-View Fusion Network (DMF-Net) based on multi-view metric learning, which aims to predict difficult airways through multi-view facial images of patients. DMF-Net adopts a dual-path structure to extract features by grouping the frontal and lateral images of the patients. Meanwhile, a Multi-Scale Feature Fusion Module and a Hybrid Co-Attention Module are designed to improve the feature representation ability of the model. Consistency loss and complementarity loss are utilized fully for the complementarity and consistency of information between multi-view data. Combined with Focal Loss, information bias is effectively avoided. Experimental validation illustrates the effectiveness of the proposed method, with the accuracy, specificity, sensitivity, and <i>F</i><sub>1</sub> score reaching 77.92%, 75.62%, 82.50%, and 71.35%, respectively. Compared with methods such as clinical bedside screening tests and existing artificial intelligence-based methods, our method is more accurate and reliable and can provide a reliable auxiliary tool for clinical healthcare personnel to effectively improve the accuracy and reliability of preoperative difficult airway assessments. The proposed network can help to identify and assess the risk of difficult airways in patients before surgery and reduce the incidence of postoperative complications.
BibTeX
@inproceedings{wu2024difficult,
  author = {Jinze Wu and Yuan Yao and Guangchao Zhang and Xiaofan Li and Bo Peng},
  title = {{Difficult Airway Assessment Based on Multi-View Metric Learning}},
  booktitle = {Bioengineering},
  year = {2024},
  doi = {10.3390/bioengineering11070703},
}
Figure