A CT-based 3D radiomic signatures combined with clinical airway examinations model for evaluation of mask ventilation in patients undergoing oral and maxillofacial surgeries

He Ren, Ling Qu, Weiwei Shi, Ping Li, Jiayi WangResearch Square 2024 · DOI

Abstract

Objective: The objective of this study is to develop a model that incorporates clinical measurements with 3D radiomic signatures extracted from CT images of oral and maxillofacial surgery patients to evaluate mask ventilation. Methods: A prospective cohort trial was conducted to enroll patients scheduled for oral and maxillofacial surgery. After obtaining informed consent, clinical measurements and head and neck CT images were collected. The anesthesiologist who managed the airway graded the mask ventilation, with difficult mask ventilation defined as mask ventilation provided without an oral airway or other adjuvant. For radiomics analysis, 3D airway segmentation was extracted and calculated 3D radiomic signatures and corresponding radiological features. Subsequently, features in the clinical measurements model and radiomic signatures model were determined using the least absolute shrinkage and selection operator (LASSO) classifier. A mixed model was developed that incorporated both radiomic signature features and clinical measurement features. Results: A total of 716 patients were enrolled in the study. The mixed model combined the five 3D radiomic signatures and six clinical measurements, and was found to have the highest predictive accuracy. In the validation group, the mixed group had an area under the curve (AUC) of 0.851, which was higher than the AUC of 0.812 in the clinical measurements model and 0.827 in the radiomic signatures model. Conclusions: This study developed a mixed model that combines 3D radiomic signatures and clinical measurements. Its application in clinical practice can assist in identifying patients at risk of experiencing difficult mask ventilation during oral and maxillofacial surgeries.

Introduction

In anesthesia, mask ventilation is the primary method for maintaining oxygenation before endotracheal intubation. Difficult mask ventilation (DMV) can lead to rapid desaturation and life-threatening complications. For patients undergoing oral and maxillofacial surgeries, the risk is particularly high, with an incidence of difficult airways reaching 10-15%—significantly higher than in general surgical populations.

Traditional preoperative assessments rely on bedside tests such as the Mallampati score, neck circumference, and jaw mobility. While these are standard practice, they are often subjective and lack high predictive precision. Although CT scans are routinely performed for these patients, their use in airway assessment has historically been qualitative. This study explores the use of radiomics—the high-throughput extraction of quantitative data from medical images—to identify structural patterns in the 3D airway that may predict ventilation challenges more accurately than traditional methods.

Methodology

This prospective cohort study enrolled 716 patients (622 after exclusions) scheduled for oral and maxillofacial surgery between 2018 and 2021. The research workflow integrated three distinct data streams to build a robust predictive tool.

Study workflow illustrating the integration of clinical measurements and 3D radiomic features.

First, 16 clinical measurements were recorded, including BMI, history of snoring, and thyromental distance. Second, researchers performed 3D airway segmentation from preoperative CT scans using 3D-Slicer software. From these segments, 1,321 quantitative radiomic features—encompassing shape, size, and voxel intensity textures—were extracted using the PyRadiomics library.

To manage this high-dimensional data and prevent overfitting, the team employed the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm. This method filtered the data down to the most predictive features, which were then used to construct three models for comparison: a clinical-only model, a radiomics-only model, and a combined "mixed model" presented as a clinical-radiomic nomogram.

Results

The study evaluated model performance using the Area Under the Curve (AUC) metric. The results demonstrated that the multimodal approach provided superior predictive power:

  • Mixed Model (Clinical + Radiomics): Achieved an AUC of 0.851 in the validation group.
  • Radiomics-only Model: Achieved an AUC of 0.827.
  • Clinical-only Model: Achieved an AUC of 0.812.

Beyond overall accuracy, the mixed model showed a high sensitivity of 92.3% in the validation cohort. This is a critical safety metric, as it indicates the model's ability to correctly identify nearly all patients who would present with difficult ventilation. Analysis of the model components revealed that radiomic signatures contributed approximately 45.5% of the predictive power, while clinical measurements contributed 54.5%, suggesting that AI-driven image analysis and traditional bedside exams are complementary partners in risk assessment.

Discussion and Conclusion

A key insight from the radiomic analysis was the importance of "median gray level intensity" (CT density values). This suggests that the density of tissues surrounding the airway—potentially reflecting fat deposition or tissue stiffness—is a vital structural factor that clinical exams cannot detect.

While the mixed model represents a significant advancement, the authors acknowledge limitations, including the single-center nature of the data and the current need for manual airway segmentation. Future research will focus on automating the segmentation process through deep learning and validating the model across multiple institutions and different CT hardware.

In conclusion, the integration of 3D radiomic signatures with clinical examinations provides a powerful, objective tool for anesthesiologists. By identifying high-risk patients before surgery, this mixed model can enhance patient safety and improve airway management strategies in oral and maxillofacial surgery.

Summary updated

Abstract

<title>Abstract</title> Objective The objective of this study is to develop a model that incorporates clinical measurements with 3D radiomic signatures extracted from CT images of oral and maxillofacial surgery patients to evaluate mask ventilation. Methods A prospective cohort trial was conducted to enroll patients scheduled for oral and maxillofacial surgery. After obtaining informed consent, clinical measurements and head and neck CT images were collected. The anesthesiologist who managed the airway graded the mask ventilation, with difficult mask ventilation defined as mask ventilation provided without an oral airway or other adjuvant. For radiomics analysis, 3D airway segmentation was extracted and calculated 3D radiomic signatures and corresponding radiological features. Subsequently, features in the clinical measurements model and radiomic signatures model were determined using the least absolute shrinkage and selection operator (LASSO) classifier. A mixed model was developed that incorporated both radiomic signature features and clinical measurement features. Results A total of 716 patients were enrolled in the study. The mixed model combined the five 3D radiomic signatures and six clinical measurements, and was found to have the highest predictive accuracy. In the validation group, the mixed group had an area under the curve (AUC) of 0.851, which was higher than the AUC of 0.812 in the clinical measurements model and 0.827 in the radiomic signatures model. Conclusions This study developed a mixed model that combines 3D radiomic signatures and clinical measurements. Its application in clinical practice can assist in identifying patients at risk of experiencing difficult mask ventilation during oral and maxillofacial surgeries.
BibTeX
@inproceedings{ren2024ct,
  author = {He Ren and Ling Qu and Weiwei Shi and Ping Li and Jiayi Wang},
  title = {{A CT-based 3D radiomic signatures combined with clinical airway examinations model for evaluation of mask ventilation in patients undergoing oral and maxillofacial surgeries}},
  booktitle = {Research Square},
  year = {2024},
  doi = {10.21203/rs.3.rs-4567996/v1},
}
Figure