LUV-Net: Multi-Pattern Lung Ultrasound Video Classification Through Pattern-Specific Attention with Efficient Temporal Feature Extraction

Abstract

Lung ultrasound (LUS) has emerged as a crucial bedside imaging tool for critical care, yet its interpretation remains challenging due to its artifact-based nature and high operator dependency. While deep learning approaches offer promising solutions for LUS pattern analysis, existing methods are limited by their focus on single-pattern recognition or disease-specific classification, and inadequate handling of temporal dynamics in video-based models. We propose LUV-Net (Lung Ultrasound Video Network), a novel deep learning model for multi-label classification of LUS patterns, combining pattern-specific attention mechanisms with temporal feature extraction. Our approach consists of two key modules: a spatial feature extraction module utilizing independent pattern-specific attention mechanisms, and a temporal feature extraction module designed to capture sequential relationships between adjacent frames. The model was evaluated using two distinct datasets: a development set of 341 LUS videos and a temporally separated validation set of 56 videos. Through 5-fold cross-validation, LUV-Net demonstrated superior performance in identifying all four LUS patterns (A-lines, B-lines, consolidation, and pleural effusion) compared to conventional video models, achieving higher AUC scores across patterns. The model's interpretability was validated through visualization of pattern-specific attention regions, providing insights into its decision-making process. The code is publicly available at https://github.com/iamhxxn2/LungUS_Video.