Development of a Novel, Low-Power, Ultrasound Algorithm for the Detection of Pneumothorax Using a Large Animal Model
Steven G Schauer1,2,3, Kendall S Hunter4, Joseph K Maddry5
1US Army Medical Center of Excellence, JBSA Fort, Sam Houston, TX 78234, USA.
Introduction:
Tension pneumothorax is the third leading cause of potentially survivable death in the prehospital, combat setting. Identification of the presence of a pneumothorax before tension physiology develops remains challenging in this setting. We conducted an early developmental pilot study to determine if unprocessed raw radio frequency (RF) data from a single-crystal ultrasound array could fill this gap.
Materials And Methods:
We prospectively enrolled sus scrofa models as part of a medical education training program with intentional induction of pneumothorax. We obtained thoracic imaging using Clarius (clinical) and Verisonics (research) devices, only the latter of which could provide RF data from the entire probe array. We assembled RF time histories into a feature vector and used principal components analysis to extract features with the greatest variance. We examined linear discriminant analysis (LDA) and logistic regression as classifiers.
Results:
Six sus scofa were included in the final analysis. The Clarius system yielded single image-based RF-traces per acquisition, which did not prove useful for further analysis. From the Verisonics system, we obtained 49 acquisitions pre-pneumothorax and 41 acquisitions pneumothorax, each of which contained 20 image frames and raw RF data for all scanlines. A vast majority of the RF signal variance was contained in the first PC, although all but the last PC contained at least >0.3% of the total variance. Only PC0 mean is statistically significant between pre- and post- groups (P = .0472). A bivariate logistic model using PC0 and PC8 (P = .184) correctly predicted 5 of 6 animals in each condition (83.3%), with 1 animal misclassified from each condition. The LDA analysis yielded only 1 linear discriminator feature, which showed a difference in the means between groups (P = .0161). This single LD used as input to a univariate logistic model yielded equal prediction accuracy to the previous classifier (83%, 1 misclassified per group), with animal 3 pre and animal 1 post misclassified by this reduced feature, and animal 2 post being nearly misclassified.
Conclusions:
In this pilot study, we were able to determine a potential signal for the diagnosis of pneumothorax using RF data. Our findings will aid in the development of low-power devices to detect pneumothorax.
