Dubline: a Deep Unfolding Network for B-Line Detection in Lung Ultrasound Images

Tianqi Yang, Nantheera Anantrasirichai, Oktay Karakus, Marco Allinovi, Hatice Ceylan Koydemir, Alin Achim

Abstract: In the context of lung ultrasound, the identification of B-lines, which serve as indicators of interstitial lung disease and pulmonary edema, holds immense significance in clinical diagnosis. Presently, although vision-based automatic B-line detection techniques have emerged, their performance remains suboptimal. This paper introduces a novel approach, framing B-line detection as an inverse problem through the deep unfolding of the Alternating Direction Method of Multipliers. By leveraging the capabilities of deep neural networks and model-based methods, this methodology addresses the challenges associated with data labeling and model training in lung ultrasound image analysis. Our primary aim is to significantly augment diagnostic precision while maintaining efficient real-time capabilities. The experiment on 34 patients demonstrates that the proposed method outperforms traditional model-based approaches, achieving a 10.6% higher F1 score and running over 90 times faster, underscoring its potential for real-time clinical utility.