Using Transfer Learning Method to Develop an Artificial Intelligence Assisted Triaging for Endotracheal Tube Position...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Critical Care Medicine

Background:

• Endotracheal tube (ETT) placement is crucial for mechanical ventilation, but confirmation via chest X-ray (CXR) requires expert interpretation and can cause delays.
• Improper ETT placement risks ventilation inefficiency and patient injury.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI)-based triage system for automated assessment of endotracheal tube (ETT) placement in chest X-rays (CXRs).

Main Methods:

• Trained six AI models, including four pre-trained (VGG16, INCEPTION_V3, RESNET, DENSENET169) and two custom models, using transfer learning on 4293 CXRs from 2568 ICU patients.
• Models were trained on regions of interest (ROIs) containing the clavicles, carina, and ETT tip, with CXRs labeled as "CORRECT" or "INCORRECT" ETT placement.

Main Results:

• The VGG16 model, trained on ROIs, achieved acceptable performance with an AUROC of 92% and an F1 score of 0.87.
• Other models did not achieve comparable performance levels.

Conclusions:

• The study demonstrates the feasibility of using transfer learning for developing AI models to assess endotracheal tube placement in chest X-rays.
• AI-powered assessment of ETT placement in CXRs holds potential for clinical application.