Complexity-Pursuit Structural Descriptors for Lung Ultrasound Video Analysis

Lung ultrasound (LUS) videos contain temporal structure beyond static image appearance. This work introduces Complexity Pursuit (CP) descriptors for compact video-level analysis of LUS structural dynamics. Forty-six public videos (COVID-19, healthy, bacterial or viral pneumonia) were converted into pixel-time matrices, PCA-whitened subspaces, and CP spatial-temporal maps. Per-video descriptors combined CP spatial-gradient and dominant-frequency summaries. These motion-structural descriptors outperformed appearance-only baselines, gaining 0.161 in binary ROC-AUC and 0.157 in multiclass macro-F1. A logistic regression model reached 0.924 ROC-AUC in leave-one-video-out COVID-19 classification. Results demonstrate that CP-derived descriptors capture vital video-level information missed by static appearance alone.