Automatic quantification of pleural effusion volume in lung ultrasound: an AI-based approach

Pleural effusion is a clinical condition characterized by abnormal fluid collection within the pleural space. One of the most common treatments, especially in critical patients, is thoracentesis consisting of removing the excess fluid through a needle not only to restore a normal lung expansion, but also to analyze the pleural fluid to determine the underlying cause. Thanks to its high temporal resolution, Lung Ultrasound (LUS) not only offers a crucial real-time guidance for needle insertion during thoracentesis, but it allows a rapid and noninvasive evaluation of pleural effusion volume. However, while LUS enables for rapid assessment at the bedside, accurate and reproducible quantification of the volume of pleural effusion remains an unmet clinical need. In this work, effusion volumes were reconstructed by integrating segmented areas across longitudinal and transverse scans and compared with thoracentesis output. A modified ResNet-based architecture was trained on manually annotated 2D ultrasound images to segment pleural effusion areas. Automated segmentations were then blindly compared with expert manual delineations. The segmentation algorithm achieved a mean Dice Similarity Coefficient (DSC) of 0.86 ± 0.04, outperforming inter-operator agreement (DSC 0.82 ± 0.05). Starting from two longitudinal and one transverse 2D segmentations, a volume reconstruction and quantification was finally performed. Automated volume estimates showed good agreement with thoracentesis, with a MAE of 31.6 ± 14.2 cm3 and an RMSE of 35.9 cm3. These preliminary results demonstrated that deep learning-based analysis of LUS can provide fast, non-invasive and accurate quantification of pleural effusion volume. This can represent a crucial support in clinical decision-making.