Lung ultrasound, the most precise diagnostic tool for pleural effusions, is underutilized due to healthcare providers' limited proficiency. To address this, deep learning models can be trained to recognize pleural effusions. However, current models lack the ability to diagnose effusions in diverse clinical contexts, which presents significant challenges. To develop and validate a deep learning model for detecting pleural effusions in lung ultrasound images, with adaptable performance characteristics tailored to specific clinical scenarios. A retrospective study was conducted at two Canadian tertiary hospitals to evaluate the detection of pleural effusions of varying sizes and complexities using lung ultrasound. A deep learning model incorporating a frame-level convolutional neural network and a clip-level prediction algorithm was developed and validated against expert annotations. The model was evaluated using a holdout dataset of 103 lung ultrasound clips from 46 patients with pleural effusion and 136 clips from 83 patients without effusion. The general model achieved a sensitivity of 0.90 for small-to-large effusions, with a specificity of 0.89. The large effusion model demonstrated a sensitivity of 0.97 for large effusions while maintaining a specificity of 0.90. The trauma model showed high sensitivity to all effusions, including trace (0.91) and small (0.97) effusions. Our research highlights the development of a deep learning model that effectively detects pleural effusions of varying sizes and complexities on lung ultrasound in different clinical settings. This tool has the potential to enhance emergency physicians' ability to quickly and accurately diagnose effusions, particularly in time-sensitive situations.
Read full abstract