Related Experiment Video
Updated: Nov 2, 2025
05:22
Author Spotlight: Demonstrating Systematic Endobronchial Ultrasound to New Endoscopists
Published on: August 11, 2023
2.5K
Automatic Image Selection Model Based on Machine Learning for Endobronchial Ultrasound Strain Elastography Videos.
Xinxin Zhi1,2,3, Jin Li4, Junxiang Chen1,2,3
1Department of Respiratory Endoscopy, Shanghai Chest Hospital, Shanghai Jiao Tong University, Shanghai, China.
Frontiers in Oncology
|June 17, 2021
Summary
Machine learning can automatically select high-quality images from endoscopic ultrasound (EBUS) strain elastography videos for diagnosing lymph nodes (LNs). This automated method demonstrates stability and diagnostic accuracy comparable to expert selections, improving efficiency in clinical practice.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Gastroenterology
Background:
- Endoscopic ultrasound (EBUS) strain elastography assesses tissue stiffness for diagnosing intrathoracic lymph nodes (LNs).
- Subjectivity in manual image selection limits the diagnostic efficiency of EBUS strain elastography.
- This study addresses the need for automated, objective image selection from EBUS strain elastography videos.
Purpose of the Study:
- To develop and evaluate a machine learning model for automatic selection of high-quality, stable representative images from EBUS strain elastography videos.
- To compare the performance of the automated image selection model against expert and trainee selections.
- To enhance the diagnostic utility of EBUS strain elastography through objective image selection.
Main Methods:
- A machine learning model was trained on EBUS strain elastography videos from 415 LNs (training/validation sets).
- The model automatically selected three representative images per LN from a test set of 91 LNs.
- Image quality and stability were evaluated using qualitative grading and four quantitative methods, compared against expert and trainee selections.
Main Results:
- The machine learning model produced stable representative images with no statistical difference in quality.
- Quantitative analysis showed lower coefficient of variation values for the machine learning model compared to expert and trainee groups.
- Diagnostic accuracies for the machine learning group (78.02%–83.52%) were comparable to the expert group (80.22%–82.42%).
Conclusions:
- An automated machine learning model effectively selects stable, high-quality representative images from EBUS strain elastography videos.
- This automated approach shows significant potential for improving the diagnosis of intrathoracic LNs.
- The model offers a more objective and consistent method for image selection, reducing inter-observer variability.
