A Method Combining Kalman's Algorithm and the YOLOv7 for Estimating the Motion of Carotid Plaques in B-Mode Ultrasound Image Sequences

The accurate extraction of carotid plaque motion trajectories facilitates early diagnosis of cardiovascular diseases. While traditional block matching-based plaque motion estimation provides high accuracy, it demands extensive computing time and resources, impeding clinical utility. Hence, this study proposes a YOLOv7-based carotid plaque motion extraction approach. First, we introduced the YOLOv7 object detection algorithm into ultrasound image vascular motion estimation, employing multi-feature annotation to enhance detection accuracy. Next, a global attention mechanism was incorporated into YOLOv7 to improve model accuracy in tracking carotid plaque features. Finally, an optimized Kalman filter further refined motion trajectory precision. Experiments demonstrated that after training, YOLOv7 estimated three cardiac cycles across 150 consecutive frames of a simulated carotid artery, yielding a total RMSE of 0.0648px for block matching-based plaque motion versus 0.1941 for the proposed technique. Although it has marginally less precise, our approach processed single frames in just 0.028 seconds, compared to 0.0350 seconds for block matching. Clinical data further evidenced the speed of our approach processed single frames in just 0.0273 seconds, compared to 0.6645 seconds for block matching. It's robustness to reference block and search area size, superior consistency, and resilience to out-of-plane effects, with improved adaptability to non-rigid tissue motions are also evidenced.