Creation of an Open-Access Lung POCUS Image Database for Deep Learning and Neural Network Applications.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Lung point of care ultrasound (POCUS) offers advantages over traditional imaging for diagnosing pulmonary conditions.
• Widespread adoption of POCUS is limited by operator dependency, interrater reliability, and training needs.
• Deep learning (DL) development is hindered by a lack of comprehensive, well-annotated lung POCUS image repositories.

Purpose of the Study:

• To create and share a large, open-source dataset of lung POCUS images.
• To provide expert-annotated data for developing and validating DL algorithms.
• To facilitate advancements in automated lung POCUS interpretation and acquisition.

Main Methods:

• A multi-center study collected lung POCUS videos from 226 adult patients with respiratory symptoms.
• Images were acquired using standardized protocols and various POCUS devices.
• Three blinded researchers independently analyzed images, with disagreements adjudicated for definitive interpretations. Videos were preprocessed, and frames were extracted and standardized.

Main Results:

• The dataset comprises 1,871 video clips (324,027 frames) with 50% of patients having COVID-19 pneumonia.
• Abnormalities included B-lines (18%), consolidations (4.5%), and combined findings (6.4%).
• Pathological findings showed significant variation by lung zone, with anterior zones being more frequently normal.

Conclusions:

• This annotated lung POCUS repository is a valuable resource for DL applications in respiratory diagnostics.
• The dataset includes patients with and without COVID-19, enhancing its utility for AI tool development.
• This resource can help overcome challenges in lung POCUS acquisition and interpretation through AI.