Accurate and rapid detection of B-lines—vertical hyperechoic artifacts in lung ultrasound (LUS) images—is essential for diagnosing and monitoring pulmonary diseases. We propose a novel algorithm based on variational mode decomposition (VMD) for automatic segmentation and quantification of B-lines in LUS frames, marking the first application of VMD in this context. Each frame is treated as a 2-D signal and decomposed into intrinsic frequency modes, isolating characteristic vertical structures while suppressing noise and irrelevant anatomy. A blockwise projection strategy enhances structural consistency and robustness under low-contrast conditions. Evaluated on 116 annotated images from 42 patients using a portable convex probe, the method achieved a precision of 0.87, recall of 0.79, F1-score of 0.83, and Dice coefficient of 0.79. Although slightly below top deep learning models, its low computational complexity makes it suitable for real-time, point-of-care ultrasound systems. Future work may combine VMD with neural networks or beamforming to further improve performance.
Variational Mode Decomposition for B-Lines Segmentation in Lung Ultrasound Images: An Improved Approach
Bottino A.Primo
;Conversano F.
;Pisani P.;Lay-Ekuakille A.;
2025-01-01
Abstract
Accurate and rapid detection of B-lines—vertical hyperechoic artifacts in lung ultrasound (LUS) images—is essential for diagnosing and monitoring pulmonary diseases. We propose a novel algorithm based on variational mode decomposition (VMD) for automatic segmentation and quantification of B-lines in LUS frames, marking the first application of VMD in this context. Each frame is treated as a 2-D signal and decomposed into intrinsic frequency modes, isolating characteristic vertical structures while suppressing noise and irrelevant anatomy. A blockwise projection strategy enhances structural consistency and robustness under low-contrast conditions. Evaluated on 116 annotated images from 42 patients using a portable convex probe, the method achieved a precision of 0.87, recall of 0.79, F1-score of 0.83, and Dice coefficient of 0.79. Although slightly below top deep learning models, its low computational complexity makes it suitable for real-time, point-of-care ultrasound systems. Future work may combine VMD with neural networks or beamforming to further improve performance.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
