Deep Learning Tool for Automatic B-Line Detection and Localization in Lung Point-of-Care Ultrasound

ABSTRACT

Study Objectives: Point-of-care ultrasound (POCUS) offers real-time data to guide clinical decision-making and patient care. Despite having advantages over alternative imaging studies such as computed tomography or magnetic resonance imaging, performing POCUS requires technical expertise for image acquisition and interpretation, thereby limiting its use for many clinicians. Deep learning technologies can provide automated interpretation of POCUS images thus making POCUS accessible to even novice users. B-lines are sonographic artifacts seen on lung POCUS which are diagnostic for pulmonary diseases such as pneumonia, COVID-19, or decompensated heart failure. In this work we aim to develop a deep learning tool to automatically detect and localize B-lines on lung POCUS clips. Methods: Using a 12-point scanning protocol, we prospectively collected lung POCUS clips from 25 patients presenting to the emergency department with shortness of breath and/or flu-like symptoms. Sub-sampled frames from 500 POCUS clips were annotated for B-lines by 3 physicians with expertise in POCUS acquisition and interpretation. A 2D U-Net deep neural network was trained on frames annotated from 15 patients, with frames from the remaining 10 patients being set aside for validation studies. Transformations from polar to rectangular coordinates were performed as part of pre-processing the data. Frame-level predictions were aggregated to predict the presence or abscence of B-lines over an entire clip. Experiments are currently underway for determining the impact of weakly supervised vs. fully supervised training. Results: Initial results show an AUC score (95% CI) of 0.82 (0.74-0.89) for clip-level B-line detection based on a 5- fold cross-validation for the 15 patient subset. Additionally, by first segmenting B-lines, our approach for localization is substantially more specific than common alternatives, such as class-activation mapping. Conclusion: Here we generated a deep learning model that can detect the presence of B-lines on POCUS clips with significant accuracy. This model was developed from a limited training subset, thus we predict that with more integrated data, our model can be further refined to identify and ideally quantify B-lines on POCUS clips collected from an array of machines and from users with variable image acquisition experience. Ideally, this tool may enable clinicians with minimal prior training in POCUS to integrate this powerful imaging tool into patient care.