Real-time medical image registration for compensating respiratory motion

Abstract

Real-time imaging such as US and fluoroscopy is widely used as a guidance tool for image-guided intervention. To supplement the low quality of target image, clinicians additionally use high quality preoperative image such as CT or MR image. In those applications, accurate alignment via registration among the multi-modal images is imperative. This dissertation covers two kinds of medical image registration frameworks to compensate respiratory motion. Most existing techniques for registration between 2D US and 3D preoperative images have some limits, either in the registration speed or the performance. We develop a real-time and automatic registration system between the two images of the liver, for image-guided interventions. We first generate a 4D preoperative image along the respiratory phase based on the transformations obtained from non-rigid registration of several 3D preoperative images acquired at different respiratory phases. Via rigid registration of acquired two 3D US to the 4D image, pose information of a fixed-pose US transducer is then determined with respect to the preoperative image coordinates. Finally, we acquire 2D US images in real-time from the transducer, and we select the correct corresponding image among 2D preoperative image candidates from the 4D image via real-time 2D registration. In addition, if needed, we obtain the position information of the target lesion using the 3D preoperative image to which the registered 2D preoperative slice belongs. Meanwhile, we also propose a real-time registration algorithm between 2D fluoroscopic image and 3D CT images to improve guiding accuracy of lung biopsy by compensating respiratory motion. The proposed registration algorithm is robust to small movement and deformation due to cardiac motion. Based on the transformations obtained from non-rigid registration between 3D CT images acquired at expiratory and inspiratory phases, we generate sequential 3D CT images and 2D digitally reconstructed radiographs (DRRs) of vessels. We then determine a 3D CT images corresponding to each real-time 2D fluoroscopic image, by matching the fluoroscopic image to the 2D DRRs using the proposed measure less vulnerable to cardiac motion. The both proposed system is evaluated on various clinical datasets and the evaluated results show that the registration error is less than 3mm on average.