Inferior vena cava tracking in ultrasound videos

The inferior vena cava (IVC) is the largest vein in the human body that carries oxygen-poor blood from the lower part of the body to the right atrium of the heart. The objective of this study is to assess and compare tracking algorithms of OpenCV library to determine the most effective approach for tracking IVC in ultrasound (US) videos. Tracking is used to follow the movement of the vein, providing a complete view of the IVC's position, leading to the identification of respiratory movements and the assessment of diameter variations induced by them. The algorithms chosen for tracking are Lucas-Kanade, KCF, MIL, TLD, MOSSE, Median Flow, Mean Shift, Block Matching, Boosting, ORB and CSRT. These algorithms are widely used for pedestrian and vehicle tracking in traffic surveillance and people recognition in crowded environments. Here, they were applied to US videos for the specific task of IVC tracking. The algorithms were tested on 13 US videos showing a longitudinal view of the IVC, with different length and resolution to increase variability. The experiments were conducted on a MacBook Air equipped with an Apple M1 chip featuring an 8-core CPU and 8 GB of RAM. The 11 algorithms selected were implemented using Python, and the tracking results for each frame were compared with the ground-truth regions (GT-ROI). The GT-ROIs for each frame were created by manually selecting points on the upper and lower borders of IVC. Based on the extreme points, ROI was defined as a rectangle, which is used to evaluate each tracker's performance. The algorithms were tested on videos by initiating tracking from the location of the GT-ROI in the first frame of each video analyzed. To assess the trackers’ ability to correctly locate the IVC, the Intersection over Union (IoU) and the Euclidean distance between the center of the algorithm’s resulting ROI and the GT-ROI were calculated. Two other parameters used to evaluate performance are the percentage of failed frames and the percentage of false positives, which, using a 0.6 IoU threshold, indicate frames where the IVC position is inaccurately identified. For each pair of successive frames, the distance between the centers of the tracked ROIs was compared with the corresponding distance in the GT-ROIs. If the discrepancy exceeds 0.5 cm, the tracking is considered unreliable, as it indicates that the tracker fails to accurately follow the movement of the IVC. The best performing trackers are KCF, Block Matching and Median Flow which demonstrate good execution speed (KCF: FPS = 32.84, Block Matching: FPS = 47.05, Median Flow: FPS= 55.0). Median Flow achieved an IoU value of 0.82 ± 0.08, KCF obtained an IoU of 0.79 ± 0.09 while Block Matching has an IoU of 0.80 ± 0.08. The false positive rate is 1.84 ± 0.06% for Median Flow, 6.03 ± 0.11% for KCF and 6.50 ± 0.10% for Block Matching. The Euclidean distance is 0.44 ± 0.21 cm for Median Flow, KCF and Block Matching show a value of 0.57 ± 0.30 cm and 0.49 ± 0.23 cm respectively. All three trackers show good ability to track the movement of the IVC in the various frames of the video, with the percentage of unreliable distances ranging from 0.61% for KCF to 1.02% for Block Matching. The difference between the best trackers lies in the percentage of failed frames: Median Flow in some video is unable to detect the position of IVC in any frame. In contrast, KCF and Block Matching show no errors in detection. In conclusion, to ensure reliable tracking of IVC and follow its movements, KCF turns out to be the best tracker.