Luoma, Charlie

Excess extravascular lung water (EVLW) is a common consequence for patients with congestive heart failure, inflammatory conditions, or acute respiratory distress syndrome. Computed tomography (CT) is widely used to assess EVLW, but with a higher radiation dose and logistical complexity. Lung ultrasound (LUS) using B-line artifacts demonstrates a reasonable correlation with EVLW, but user experience with the acquisition technique, equipment-dependent factors and the subjective analysis of the B-line artifacts results in significant inter-observer variability. We propose a technique to automatically detect lung ultrasound B-lines over heart cycles. The method first extracts B-mode image features from videos in a region of interest (ROI). A ResNet50-UNet network is then applied to segment regions of B-lines. This technique was performed on 8 videos (i.e., 4 patients versus 4 controls with the duration of each video 2–3 s, resolution 1024 × 758 pixels, and frame rate 28 fps). The training accuracy and lossvalue of proposed models were 96.5% and 0.05, respectively. The paired t-test statical analysis shows that there is significant difference (p-value = 0.0035) of total number of B-line between normal and lungs with edema. The proposed method improves the accuracy for counting the B-lines and reduces the labor work and intra-observer variability from different users.