Diffeomorphic respiratory motion estimation of thoracoabdominal organs for image‐guided interventions

Abstract

Purpose Percutaneous image‐guided interventions are commonly used for the diagnosis and treatment of cancer. In practice, physiological breathing‐induced motion increases the difficulty of accurately inserting needles into tumors without impairing the surrounding vital structures. In this work, we propose a data‐driven patient‐specific hierarchical respiratory motion estimation framework to accurately estimate the position of a tumor and surrounding vital tissues in real time. Methods The motion of optical markers attached to the chest or abdomen skin is used as a surrogate signal to estimate tumor motion based on ɛ‐support vector regression (ɛ‐SVR). With the estimated tumor motion as the input, a novel respiratory motion model is developed to estimate the diffeomorphic deformation field of the whole organ (liver or lung) without intraoperative, iterative optimization computations. The respiratory motion model of the whole organ is established in Lie algebra space based on the kriging algorithm to ensure that the estimated deformation field is diffeomorphic, optimal, and unbiased. Preoperative prior knowledge for modeling the motion of whole organs is obtained by deformation registration between four‐dimensional computed tomography (4D CT) images using a hybrid diffeomorphic registration method. Results and Conclusions Experimental results on an in vivo beagle dog show that the minimum value of the determinant of the Jacobian of the estimated deformation field is greater than zero, so the estimated deformation field of the whole liver with our method is diffeomorphic. The mean position error of the tumor is 1.2 mm corresponding to a mean accuracy improvement of 76.5%, and the mean position error of the whole liver is 2.1 mm, corresponding to a mean accuracy improvement of 37.9%. The experimental results based on public human subject data show that the mean position error of the tumor is 1.1 mm, corresponding to a mean accuracy improvement of 83.1%, and the mean position error of the whole lung is 2.1 mm, corresponding to a mean accuracy improvement of 41.4%. The positioning errors for the tumor and whole organ are hierarchical and consistent with clinical demand.

Automatic registration and precise tumour localization method for robot‐assisted puncture procedure under inconsistent breath‐holding conditions

“…The tumour breath‐holding position estimation model is built based on the LibSVM library 29 . The radial basis function kernel function is adopted for the SVR algorithm to address the nonlinearity of respiration motion 30 . The feature vector for extracting the principal component of the marker's position is .…”

“…29 The radial basis function kernel function is adopted for the SVR algorithm to address the nonlinearity of respiration motion. 30 The feature vector for extracting the principal component of the marker's position is E 3�1 ¼ ½0:03; 0:3; 0:95�. The model's parameters C and γ are obtained with the particle swarm optimization method and set as C ¼ 13 and γ ¼ 0:1.…”