1659: A Deep Learning Approach for Ultrasound-Based Identification of Pneumothorax in Traumatic Injury

Introduction: Point of care ultrasound (POCUS) is a diagnostic tool used in a range of emergency care and intensive care unit (ICU) settings that offers the ability to rapidly assess internal injuries after trauma. However, its efficacy is limited by manual operator dependency, particularly in identification of subtle yet critical life-threatening injuries including traumatic pneumothorax, with operator accuracy ranging from 60% to 90%. With the emergence of artificial intelligence (AI) technology, particularly deep learning techniques, we sought to develop a model using convolution neural network (CNN) methods to improve POCUS accuracy in diagnosing pneumothorax. We sought to develop and evaluate a CNN model for the automated detection of pneumothorax from publicly available POCUS lung exams. We hypothesized that we would demonstrate improvement in comparison to a human operator with a pilot CNN-based model. Methods: A curated dataset of 110 annotated videos was generated from 81 original publicly available POCUS exams from two types of ultrasound probes (linear, convex) across ultrasound brands that were annotated for lack and existence of sliding sign. Ultrasound clips were standardized, with spatial augmentations applied to expand the dataset. An R(2 + 1)D-18 spatiotemporal (VideoResNet) architecture was trained using a 85/15 split with a Google Colab A100 GPU over multiple epochs with regularization techniques applied. Model performance was assessed via statistical analysis, p< 0.05 as statistically significant. Results: The CNN model achieved a mean area under the receiver operating characteristic curve (AUROC) of 0.713 (95% CI: 0.673–0.753) and a mean area under the precision-recall curve (AUPRC) of 0.635 (95% CI: 0.589–0.681), demonstrating moderate diagnostic capability. Conclusions: A deep-learning model in the context of the POCUS lung exam may accurately detect the presence of lung sliding sign, improving manual operator accuracy and potentially augmenting manual operator diagnosis in an emergency trauma or ICU setting. However, external validation of this model is needed in a clinical setting.