Artifact-Aware Self-supervised Temporal Learning for Pediatric Lung Ultrasound Scoring.
Authors
Abstract
Accurately assessing the severity of pediatric pneumonia using AI-based lung ultrasound (LUS) video scoring remains challenging due to three main limitations: (1) limited availability of expert annotations, (2) underutilization of temporal dynamics, and (3) lack of mechanisms to capture the directional patterns and temporal evolution of diagnostic artifacts. To address these challenges, we introduce PedLUS, a self-supervised video learning framework for Pediatric Lung Ultrasound Scoring, together with a dedicated dataset comprising 1,646 unlabeled clips for pretraining and 464/156 labeled clips for training/test severity scoring. PedLUS mitigates annotation scarcity by using a self-supervised pretraining strategy that masks the central 8 frames of a 24-frame sequence and reconstructs them from bidirectional temporal context, thereby learning motion-aware representations. To improve reconstruction fidelity, we design a Semantic-Aware Clustering Reconstruction (SACR) module that groups latent features into learnable prototypes and uses them to regenerate masked regions, reducing redundancy while emphasizing clinically meaningful artifact cues. To capture the directional and evolving characteristics of diagnostic artifacts, PedLUS further integrates two specialized modules: Spatial Directional Attention (SDA) for orientation-specific feature refinement, and Temporal-Aware Attention (TAA) for modeling progression across respiratory phases. Experimental results over ten repeated runs demonstrate that the proposed model achieves outstanding performance, with an accuracy, F1-score and AUC of 82.37 ± 1.36%, 81.56 ± 1.47%, and 92.03 ± 0.94%, respectively. Code and dataset will be released at:PedLUS.