Artifact-Aware Self-supervised Temporal Learning for Pediatric Lung Ultrasound Scoring

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Pediatric Medicine

Background:

• Assessing pediatric pneumonia severity via AI-based lung ultrasound (LUS) video scoring faces challenges.
• Limitations include scarce expert annotations, underutilization of temporal dynamics, and inadequate capture of artifact evolution.

Purpose of the Study:

• To introduce PedLUS, a self-supervised video learning framework for Pediatric Lung Ultrasound Scoring.
• To address annotation scarcity and improve the analysis of temporal dynamics and artifact characteristics in LUS videos.

Main Methods:

• Developed PedLUS, a self-supervised framework utilizing a pretraining dataset of 1,646 unlabeled LUS clips.
• Employed a Semantic-Aware Clustering Reconstruction (SACR) module for improved reconstruction fidelity.
• Integrated Spatial Directional Attention (SDA) and Temporal-Aware Attention (TAA) modules to capture artifact directionality and temporal evolution.

Main Results:

• PedLUS achieved high performance in severity scoring on a test set of 156 labeled clips.
• Demonstrated accuracy of 82.37 ± 1.36%, F1-score of 81.56 ± 1.47%, and AUC of 92.03 ± 0.94% over ten runs.
• The framework effectively learns motion-aware representations and emphasizes clinically meaningful artifact cues.

Conclusions:

• PedLUS offers a robust solution for AI-based pediatric pneumonia severity assessment using LUS videos.
• The self-supervised approach mitigates annotation scarcity and enhances the analysis of complex video dynamics.
• The framework shows significant potential for improving diagnostic accuracy in pediatric respiratory conditions.