A learning-based respiratory motion estimation with consecutive ultrasound images

One of the challenges in ultrasound image-guided liver puncture intervention is the patient’s physiological movement. The respiratory motion modeling is the basis for the accurate positioning of the target and precise puncture control. The existing method for estimating the motion among ultrasound images is speckle decorrelation. However, the traditional methods require complicated operations and are time-consuming, which cannot meet the real-time requirement of clinical use. In this study, we proposed a learning based method with deep neural network to calculate the out-of-plane distance between two consecutive ultrasound images and its direction directly from a pair of sequential ultrasound images, and the respiratory motion modeling is conducted furtherly. Preliminary experiments show that the accuracy of respiratory motion modeling is 0.98mm with our method, and the process speed can be as high as 45 frames per second on a 3080 GPU, which can meet the real-time requirement in clinical practice.