Semi-supervised Ensemble Learning for Automatic Interpretation of Lung Ultrasound Videos.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Point-of-care ultrasound (POCUS) is a valuable bedside tool for patient assessment.
• Lung ultrasonography (LUS) effectively evaluates acute and chronic pulmonary conditions.
• Automatic interpretation of LUS, especially in multi-label scenarios, is underdeveloped.

Purpose of the Study:

• To develop a novel deep learning (DL) framework for automated interpretation of lung POCUS videos.
• To identify multiple pulmonary findings (e.g., A-lines, B-lines, consolidations) within LUS videos.
• To address the challenge of multi-label interpretation in LUS analysis.

Main Methods:

• A residual (2+1)D deep learning architecture was employed for video interpretation.
• Video masking and standardization were utilized in the pre-processing stage.
• A semi-supervised learning approach leveraged unlabeled data, complemented by a hierarchy-aware ensemble strategy for distinct label sets.

Main Results:

• The categorical DL model achieved a 92.4% average F1-score for expedited triage.
• The multi-label DL model attained a 70.5% average F1-score across five LUS findings for patient management.
• Semi-supervised learning significantly enhanced performance, with ensemble modeling offering moderate additional gains.

Conclusions:

• The proposed DL framework demonstrates versatility and high performance in interpreting lung POCUS videos.
• Automated LUS interpretation, particularly with multi-label capabilities, holds significant potential for clinical applications.
• The study highlights the effectiveness of semi-supervised learning and hierarchical ensemble methods in advancing LUS analysis.