Related Experiment Video
Updated: Jun 4, 2026
05:50
International Expert Consensus and Recommendations for Neonatal Pneumothorax Ultrasound Diagnosis and Ultrasound-guided Thoracentesis Procedure
Published on: March 12, 2020
Explainable transfer learning ensemble AI model for lung ultrasound pneumothorax detection with expert benchmark
Gábor Orosz1,2,3, Róbert Zsolt Szabó4, Marcell Szabó5,6
1Department of Military, Disaster and Law Enforcement Medicine, Semmelweis University, P.O.B. 2, Budapest, 1428, Hungary. orosz.gabor@semmelweis.hu.
Summary
An explainable artificial intelligence (AI) model achieved perfect sensitivity and specificity for diagnosing pneumothorax using lung ultrasound. This AI tool reduces diagnostic variability and false positives, enhancing patient safety at the bedside.
Area of Science:
- Medical imaging and diagnostics
- Artificial intelligence in healthcare
- Critical care medicine
Background:
- Lung ultrasound is vital for rapid, radiation-free pneumothorax diagnosis.
- Human interpretation of lung ultrasound is limited by variability and data gaps.
- Existing AI models lack interpretability and rigorous validation.
Purpose of the Study:
- Develop and validate a robust, explainable AI ensemble model for pneumothorax diagnosis.
- Address limitations in dataset diversity, image acquisition, and clinical interpretability.
- Improve diagnostic accuracy and reduce variability in lung ultrasound interpretation.
Main Methods:
- Developed an explainable soft-voting ensemble AI model.
- Trained on 1,856 diverse ultrasound clips from various sources.
- Validated model interpretability using clinician-validated heatmaps.
- Benchmarked AI performance against 11 expert clinicians on an independent test set.
Main Results:
- The AI ensemble achieved 100% sensitivity and 100% specificity.
- AI performance surpassed that of expert clinicians.
- Experts showed variable performance across ultrasound modes (e.g., lower specificity in M-mode).
- AI demonstrated consistent accuracy, reduced false positives, and generalized well to clinical and cadaveric cases.
Conclusions:
- The explainable AI ensemble matches expert consensus performance.
- This AI tool significantly reduces diagnostic variability and false-positive diagnoses.
- It can serve as a critical second reader, standardizing bedside decisions and improving patient safety.
