Quantifying lung ultrasound comets with a convolutional neural network: Initial clinical results

Area of Science:

• Medical imaging
• Artificial intelligence in medicine
• Pulmonology

Background:

• Lung ultrasound comets are crucial artifacts for diagnosing lung pathologies and estimating extravascular lung water.
• Human counting of these artifacts suffers from poor definition and observer variability.
• Automated quantification methods are needed to improve diagnostic accuracy and consistency.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) for automated quantification of lung ultrasound comets.
• To compare the CNN's performance against human observers and assess its correlation with clinical parameters.
• To evaluate the CNN's potential as a diagnostic tool for comet-positive lung images.

Main Methods:

• A CNN was trained on a dataset of 4864 labeled clinical lung ultrasound images.
• The CNN's comet counting accuracy and intraclass correlation (ICC) with human counts were evaluated on a test set.
• The validated CNN was applied to a separate clinical dataset of 6272 images to analyze correlations with clinical parameters (diastolic blood pressure, ejection fraction, BMI).

Main Results:

• The CNN achieved correct comet counts in 43.4% of test images, with an ICC of 0.791 compared to human observers, surpassing inter-observer reliability.
• Automated comet counts showed positive correlation with diastolic blood pressure (r=0.448) and negative correlations with ejection fraction (r=-0.513) and BMI (r=-0.566).
• The CNN demonstrated 80.8% accuracy when formulated as a diagnostic test for comet-positive images.

Conclusions:

• The developed CNN provides a reliable and consistent method for quantifying lung ultrasound comets, outperforming human observer variability.
• Automated comet analysis may serve as a valuable non-invasive tool for assessing lung fluid status and correlating with key clinical indicators.
• Further improvements are anticipated with larger datasets and refined neural network architectures, enhancing its clinical utility.