Fast CT-CT Fluoroscopy Registration with Respiratory Motion Compensation for Image-Guided Lung Intervention

Transcription

1 Fast CT-CT Fluoroscopy Registration with Respiratory Motion Compensation for Image-Guided Lung Intervention Po Su a,b, Zhong Xue b*, Kongkuo Lu c, Jianhua Yang a, Stephen T. Wong b a School of Automation, Northwestern Polytechnical University, Xi an China; b The Methodist Hospital Research Institute, Houston, TX; c Philips Research North America, Briarcliff, NY. ABSTRACT CT-fluoroscopy (CTF) is an efficient imaging method for guiding percutaneous lung interventions such as biopsy. During CTF-guided biopsy procedure, four to ten axial sectional images are captured in a very short time period to provide nearly real-time feedback to physicians, so that they can adjust the needle as it is advanced towards the target lesion. Although popularly used in clinics, this traditional CTF-guided intervention procedure may require frequent scans and cause unnecessary radiation exposure to clinicians and patients. In addition, CTF only generates limited slices of images and provides limited anatomical information. It also has limited response to respiratory movements and has narrow local anatomical dynamics. To better utilize CTF guidance, we propose a fast CT-CTF registration algorithm with respiratory motion estimation for image-guided lung intervention using electromagnetic (EM) guidance. With the pre-procedural exhale and inhale CT scans, it would be possible to estimate a series of CT images of the same patient at different respiratory phases. Then, once a CTF image is captured during the intervention, our algorithm can pick the best respiratory phase-matched 3D CT image and performs a fast deformable registration to warp the 3D CT toward the CTF. The new 3D CT image can be used to guide the intervention by superimposing the EM-guided needle location on it. Compared to the traditional repetitive CTF guidance, the registered CT integrates both 3D volumetric patient data and nearly real-time local anatomy for more effective and efficient guidance. In this new system, CTF is used as a nearly real-time sensor to overcome the discrepancies between static pre-procedural CT and the patient s anatomy, so as to provide global guidance that may be supplemented with electromagnetic (EM) tracking and to reduce the number of CTF scans needed. In the experiments, the comparative results showed that our fast CT-CTF algorithm can achieve better registration accuracy. Keywords: CT fluoroscopy, image-guided lung intervention, respiratory motion compensation, deformable image registration, parallel computation. 1. INTRODUCTION Lung cancer is considered to be as the leading cancer-related cause of death worldwide, with 1.35 million new cases and 1.18 million deaths per year 1. According to histological type, lung cancers can be mainly classified into non-small-cell lung carcinoma (NSCLC) and small-cell lung carcinoma (SCLC) 2-4. NSCLC can also be further classified as adenocarcinoma, squamous cell carcinoma, large-cell carcinoma, and others. It is estimated that about half of the lung cancer cases are peripheral. In addition to detect abnormal nodules from CT imaging, biopsy through percutaneous intervention of peripheral lung cancer is currently the standard procedure for diagnosis and confirmation of lung cancer. In clinical practice, the major difficulties in percutaneous lung intervention are targeting accuracy, efficiency of the procedure, and complications 5. When dealing with small and relatively deep lesions, accurately targeting the lesion becomes challenging due to patient s respiratory motion. Currently, multimodality imaging techniques have been adopted to alleviate the burden. Fluoroscopy takes real time X-ray images but only provides 2D image information. Ultrasound is an alternative real-time imaging modality and can provide instant visualization of the chest structures. It is, however, not suggested for lungs and only applicable to lesions close to the diaphragmatic surface due to its imaging * Correspondence: Zhong Xue; zxue@tmhs.org; Tel:

2 depth and the rib effects 6. Real-time MRI scanners are still too expensive to be used in the clinics for guiding interventions. Therefore, conventional CT is currently a common guidance tool and is widely used in percutaneous lung biopsy 7, 8. It, however, provides very limited guidance for interventions because of lack of real-time feedback of lung anatomy during the procedure. Recently, electromagnetic (EM) tracking is used for visualizing the tracked needle in a pre-procedural or diagnosis CT by fusing EM field and the 3D CT image. Since the guidance relies on a static CT, it might generate larger error due to the dynamic deformation of lung parenchyma during respiratory cycles. Therefore, a precise prediction of the movement of the lung from real-time sensors to better match patients anatomy would be highly desirable for more accurate localization. CT fluoroscopy (CTF) combines the advantages of both conventional CT and fluoroscopy and has been recognized as an efficient and convenient tool for image-guided lung intervention percutaneous During CTF-guided biopsy procedure, four to ten axial sectional images are captured in a very short time period by the traditional CT scanner to provide physicians nearly real-time feedback about patient s anatomy. These CTF images can provide guidance to adjustment of the needle, as it is advanced towards the target lesion. In current clinical practice, physicians normally advance the needle using CTF as guidance by first asking patients to hold breath, pressing the foot paddle of the CT scanner to capture the CTF images, and then advance the needle accordingly. There is normally a 15 to 30 second time window dependent on how well the patient can hold breath. Physicians can repeat such procedure until the needle hit the target lesion for biopsy. This procedure may require frequent scans and cause unnecessary radiation exposure to clinicians and patients. In addition, it also has limited response to respiratory movements and only provides narrow local anatomical dynamics. To better utilize CTF guidance, in this paper, we propose a fast CT-CTF registration algorithm with respiratory motion estimation for image-guided lung intervention using electromagnetic (EM) guidance. With the pre-procedural exhale and inhale CT scans, it would be possible to estimate a series of CT images of the same patient at different respiratory phases. Then, once a CTF image is captured during the intervention, our algorithm can pick the best respiratory phase-matched 3D CT image and performs a fast deformable registration to warp the 3D CT toward CTF. The new 3D CT image can then be used by the interventional system. Two sets of experiments were performed to evaluate the proposed fast CT-CTF registration algorithm. First, we tested the accuracy of the CT-CTF registration without motion compensation by using existing 4D CT datasets, where the inhale, exhale, and middle phase CT images are known, and CTF images are simulated from these 4D CT images using longitudinal deformations obtained from the serial CT images. In this case, the ground truth of the underlying deformation between CTF and CT (inhale) is known. Then, the accuracy of CT-CTF registration with motion compensation was tested in the same way. Finally, the two results were compared. The paper is organized as follows. Section 2 discusses the proposed algorithm in detail. The results are presented in Section 3. Section 4 is the conclusion of this study. 2. METHODS 2.1 Algorithm Framework Because the pre-procedural CT covers the entire or most of the lung region, and there are only 4 to 10 slices in the CTF images, the current 3D to 3D registration or 3D to 2D registration algorithms do not fit this kind of application. Because there is only limited number of slices for CTF, the deformation consideration in x and y direction should be different from the z direction. On the other hand, although the slices covered by CTF imaging can be precisely aligned, the deformation of the rest slices can only be interpolated or estimated. To solve these problems, we propose a CT-CTF registration algorithm to deal with limited number of slices. More over such CT-CTF registration can be incorporated in the framework of motion estimation so that the information of patient s respiratory motion can be embedded in the algorithm. Fig.1 illustrates the workflow of the image-guided intervention using CTF as the real-time patient feedbacks. The algorithm is performed in the following steps. First, the pre-procedural lung CT scans are captured in both exhale and inhale phases. Then, longitudinal deformation between the exhale and inhale phases can be calculated by performing a 3D/3D image registration 12. A series of CT images can then be estimated by scaling the longitudinal deformation to simulate the CT images between exhale and inhale phases. This process can be performed prior to the interventional procedure.

3 During the interventional procedure, when an intra-procedural CTF is captured, the breathing phase can be determined by matching the CTF with the above estimated serial CT images using least squares differences. Next, the CT-CTF registration is performed to warp the inhale CT image onto the CTF space by considering a combined deformation, i.e., first deforming the inhale CT image onto the intermediate CT that best matches the respiratory phase, and then deforming onto the CTF space. Notice that the longitudinal deformation of the intermediate CT caused by respiratory motion is known, and the goal is to estimate the elastic motion between such motion-compensated CT and the CTF. This process requires fast computation to satisfy intra-procedural guidance while the patient is holding breath. With the EM-guided system, the tracked needle can be displayed in the deformed 3D CT image for guiding the intervention. Currently we need to move the table out of the gantry so that EM tracking is not affected. The procedure is repeated till the tumor is targeted. In our work, we achieved the registration within one second, and hence interventional radiologists can use the breath holding time window to guide the intervention in 3D space. Registration Pre-procedural Scan (exhale) Pre-procedural Scan (inhale) Pre-procedural process Estimation of 4D CT images from exhaleinhale data by scaling the longitudinal deformations Series CT image for the patient Intra-procedural process CT-Fluoro Scan Pick up the best CT Fast image registration CT image for interventional guidance Figure 1. The framework of CT-CTF registration with motion compensation. 2.2 A Fast CT-CTF Registration Algorithm Many methods have been developed for deformable image registration They warp one image onto another by determining a deformation field that minimizes the image difference measure. Different image similarity measures, such as intensity-based, mutual information-based, and feature-based similarity measures, and different deformation models, such as thin plate spline, cubic B-Spline, diffeomorphism, optical flow, finite element method (FEM), as well as different optimization strategies can be used in various deformable registration algorithms 21. The basic registration framework can be generally summarized as solving a deformation field f that minimizes the energy function, f = arg min f E S I F, I C, f + λe r f, (1) where E S I F, I C, f represents the difference measure between the reference image (CTF) I F and the moving (CT) image I C using the current deformation field f, and E r f is the regularization energy term for the deformation field. In our algorithm, cubic B-Spline is used to model the deformation, and the deformation f is calculated by Eq. (2),

4 3 3 f x, y, z = B l u B m v c i+l,j +m,z (2) l=0 m =0 where c is the mesh of control points, i = x n x 1, j = y n y 1, u = x n x x n x, v = v n y y n y. n x and n y are the grid spacing. B l represents the lth basis function of cubic B-Spline. Smaller values for grid spacing will result in a finer grid but the computation time will increase as the number of control points increases. The basis functions are listed as follows: B 0 = 1 r 3 /6 B 1 = 3r 3 6r /6 B 2 = 3r 3 + 3r 2 (3) + 3r + 1 /6 B 3 = r 3 /6 Notice that in our work the B-Spline interpolation is not used in z-direction because CTF has a limited number of slices. However, c and f are be defined in 3D space to model 3D deformation. Because B-splines are used in the transverse plane, we can guarantee smoothness in x-y plane. In z-direction, an additional smoothness constraint is necessary, thus E r f in Eq. (1) can be defined as, E r f = 1 2 v Ω f z 2, (4) where v is a voxel in the CTF image space Ω. In this paper, we use the sum of squared differences of the CT and CTF as E S. The minimization of the energy function is achieved through the finite derivatives of the whole energy with respect to the deformations defined on each control point numerically. Notice that global rigid transformation is first performed before deformable registration. Block1 Block2 Block1 Block2 Block3 Block4 Overlapping Figure 2. The image domain can be divided into overlapping sub-blocks for parallel registration (left), and the size of overlapping region is set to cover two B-Spline control points (right). In order to allow more time for intervention, the multiple thread programming is used so that the registration can be performed in parallel. In this way, the images are first partitioned into overlapping sub-blocks in x-y plane, and the registration of each block is calculated in parallel. Fig. 2 illustrates that we only divided x-y plane because the number of slices in z-direction is limited. Fig. 2 (right) shows that in the cubic B-Spline implementation, the size of overlapping region needs to cover an image area defined by at least two control points. In this work, our program can finish the registration between a 3D CT image (with size 512x512x35) and a CTF image (with size 512x512x4) within 1 second by dividing the image into 4 blocks. We suggest the number of blocks used equal to the number of processor cores of the

5 computer. Fig.3 gives some examples of the registration results, showing that the CT images aligned with CTF well after registration. (a) (b) (c) (d) Figure 3. Examples of the CT-CTF registration results by overlapping the CTF (red) onto the CT (gray scale) images. (a) and (c): before registration; (b) and (d): after registration. 2.3 CT-CTF Registration with Respiration Motion Compensation The CT-CTF algorithm can rapidly register CT with CTF images. Since only the slices covered by CTF images can be precisely aligned, the deformation of the rest slices can only be estimated by gradually decreasing and expanding the deformation field outside the boundary CTF slices along z-direction. Therefore, if there is a motion discrepancy between the pre-procedural images and CTF, a gradual transition or gap might be notable. Ideally one would have the 4D CT ready, however it is not a standard clinical practice to capture 4D CT for biopsy patients due to radiation and cost concerns. In order to better estimate such intra-procedural CT images, we embedded a linear interpolation motion compensation module with the CT-CTF algorithm in this work. Given the exhale and inhale CT images, I 0 and I T, the longitudinal deformation h T 0 of the patient can first be estimated by registering these two images. Then, we linearly scale the deformation, and the deformation from an intermediate time point 0 t T can be calculated as,

6 h T t = T t T h T 0. (5) Thus the deformation from t to T h t T can be estimated as the inverse of h T t. Given a CTF image, once a proper time point t is determined, we can embed h t T into the registration procedure. Then the deformation field f can be described as, f = h t T h CTF t. (6) We can see that the deformation procedure is that we first register the CTF image onto the motion estimated CT image at time point t and then further register it onto the inhale data at time T. Notice the intermediate image at time t may not necessarily be used explicitly in the registration to prevent accumulation of image re-sampling errors. In this case, we only need to solve h CTF t in Eq. (6) because h t T can be considered as known. Therefore, we will use the cubic B-Spline in Eq. (2) to model h CTF t, and the new energy function now becomes, h CTF t = arg min h E S I F, I C, h t T h CTF t + λe r h CTF t. (7) (a) (b) (c) Figure 4. An example of the CT-CTF registration with motion compensation. (a) CTF; (b) CT; (c) estimated intermediate image; (d) overlapping deformed CT in (b) (gray scale) onto CTF. (d)

7 It can be seen that comparing to Eq. (1), if the motion is big, estimation of h CTF t is relatively easy and more accurate than estimation of f because the deformations of h CTF t may be smaller than f, and many larger respiratory motion related deformation has been handled by h t T, which is considered as known. Fig. 4 shows the procedure of CT-CTF registration with motion compensation. 3. RESULTS Two sets of experiments were carried out to evaluate the proposed fast CT-CTF registration algorithm using 4D CT datasets from 30 subjects. We chose to use 4D CT data for evaluation because in real clinical cases the ground truth for the registered 3D CT images is not available. Twelve 3-D images were acquired for each patient, and the images were aligned so that the first and the last images were exhale data and the 7th data was the inhale data. The data were provided by Dr. Daniel Low and Dr. Wei Lu from Washington University in ST. Louis (now University of California at Los Angeles). All the images have an in-plane resolution of 0.98x0.98mm and a slice thickness of 1.5mm. For each subject, three images: exhale (I 1 ), inhale (I 2 ), one in between (I 3 ) were selected in the evaluation. Because for the 4D CT data we do not have CTF, for each subject, we simulated the CTF image. By registering images I 2 and I 3 using the FSL fnirt 22 algorithm, we obtained the deformation from I 3 to I 2, denoted as f 3 2. Then, image I 2 can be deformed using f 3 2 to get the warped image of I 2 that matches I 3. Finally, ten slices from the deformed image were cut to simulate the CTF image of the subject. By using these simulated CTFs, the underlying deformation from them to corresponding inhale images I 2 are known, and it will be possible to calculate the accuracy of the registration. In the first experiment, we evaluated the performance of the proposed CT-CTF registration without motion compensation, i.e., we directly registered the inhale CT image (I 2 ) of each subject onto the corresponding simulated CTF image. Finally, the accuracy of the proposed CT-CTF registration algorithm was calculated by using the deformation errors between the results and the ground truth. In the second experiment, we evaluated the performance of CT-CTF registration with motion compensation. To do so, we first used the FSL fnirt registration algorithm to calculate the longitudinal deformation between the exhale and inhale images I 1 and I 2, and the resultant deformation can be denoted as f 2 1. Then, the deformation was scaled to obtain a number of intermediate deformations, f t = (T t)/t f 2 1, and corresponding intermediate images I t, t=1,2,,t. Then, the image intensity-based similarity between the CTF image simulated above and intermediate images I t can be calculated on the available CTF slices. The best intermediate image was then selected as the one that matches the respiratory phase of the input CTF. Thus, the reverse of the deformation f t, f t 1, can now be regarded as h t T in Eq. (6), and the task is to solve h CTF t by minimizing Eq. (7). In this way, the motion compensation was embedded in the CT-CTF registration procedure, and the final registration result will be calculated according to Eq. (6). Similarly, we calculated the deformation errors between the result f, and the ground truth f 3 2. After processing all the images from the 4D CT datasets, CT-CTF registration without MC has an average deformation error 1.96mm. By contrast, CT-CTF registration with MC has 1.68mm. Compared with the result of CT- CTF registration without MC, CT-CTF registration with MC is more accurate. 4. CONCLUSION In this paper, we proposed a fast CT-CTF registration algorithm with motion estimation for image-guided lung intervention. With pre-procedural exhale and inhale CT scans, it is possible to estimate a series of CT images at different respiratory phases of the same patient. Then, once the real-time captured intra-procedural CTF images are captured during the intervention, our algorithm can pick the best respiratory phase and performs a fast deformable registration. Within one second a 3D CT image can be generated to match the CTF. This allows a > 15 seconds time window for interventional radiologists to advance the needle with 3D image guidance during the breath holding. Different from traditional 3D-3D or 2D-3D algorithms, our CT-CTF algorithm models the deformation in the axial plane using 2D B- Spline and constrains the deformation along z-direction using general smoothness constrains because B-Spline is limited

8 to model the deformation in this direction, giving limited number of slices from CTF. But the deformation of each voxel is still modeled in 3D. The comparative results showed that the proposed CT-CTF approach overcomes discrepancies between static pre-procedural/diagnosis CT and the patient s anatomy and corrects the respiratory motion and image deformations, so as to provide global guidance supplemented with EM tracking and possibly reduce CTF scans. In this paper, we estimated the intermediate CT images by sampling the longitude deformation between inhale and exhale status. To better estimate the serial CT images, in the future, other methods nonlinear motion regression model may be used. 5. REFERENCES 1. Parkin, D.M., et al., Global cancer statistics, CA: A Cancer Journal for Clinicians, (2): p Travis, W.D., Pathology of lung cancer. Clinics in Chest Medicine, (1): p Jackman, D.M. and B.E. Johnson, Small-cell lung cancer. The Lancet, (9494): p Pearson, F.G., Non-small Cell Lung Cancer*. Chest, (suppl 3): p. 500S-503S. 5. Tsukada, H., et al., Diagnostic accuracy of CT-guided automated needle biopsy of lung nodules. American Journal of Roentgenology, (1): p Hayashi, N., et al., CT-guided biopsy of pulmonary nodules less than 3 cm: usefulness of the spring-operated core biopsy needle and frozen-section pathologic diagnosis. American Journal of Roentgenology, (2): p Li, H., et al., Diagnostic accuracy and safety of CT-guided percutaneous needle aspiration biopsy of the lung: comparison of small and large pulmonary nodules. American Journal of Roentgenology, (1): p Wood, B.J., et al., Navigation systems for ablation. Journal of Vascular and Interventional Radiology, (8): p. S257-S Katada, K., et al., Guidance with real-time CT fluoroscopy: early clinical experience. Radiology, (3): p Silverman, S.G., et al., CT Fluoroscopy-guided Abdominal Interventions: Techniques, Results, and Radiation Exposure1. Radiology, (3): p Gianfelice, D., et al., Value of CT fluoroscopy for percutaneous biopsy procedures. Journal of Vascular and Interventional Radiology, (7): p Xue, Z., K. Wong, and S.T.C. Wong, Joint registration and segmentation of serial lung CT images for image-guided lung cancer diagnosis and therapy. Computerized Medical Imaging and Graphics, (1): p Mattes, D., et al., PET-CT image registration in the chest using free-form deformations. IEEE Transactions on Medical Imaging, (1): p Warfield, S.K., et al., Real-time registration of volumetric brain MRI by biomechanical simulation of deformation during image guided neurosurgery. Computing and Visualization in Science, (1): p Ferrant, M., et al. Registration of 3D intraoperative MR images of the brain using a finite element biomechanical model. 2000: Springer. 16. Johnson, H.J. and G.E. Christensen, Consistent landmark and intensity-based image registration. IEEE Transactions on Medical Imaging, (5): p Sorzano, C.O.S., P. Thévenaz, and M. Unser, Elastic registration of biological images using vector-spline regularization. IEEE Transactions on Biomedical Engineering, (4): p Duncan, J.S. and N. Ayache, Medical image analysis: Progress over two decades and the challenges ahead. IEEE Transactions on Pattern Analysis and Machine Intelligence, (1): p Huang, X., N. Paragios, and D.N. Metaxas, Shape registration in implicit spaces using information theory and free form deformations. IEEE Transactions on Pattern Analysis and Machine Intelligence, (8): p Rueckert, D., et al., Nonrigid registration using free-form deformations: application to breast MR images. IEEE Transactions on Medical Imaging, (8): p Crum, W.R., T. Hartkens, and D. Hill, Non-rigid image registration: theory and practice. British Journal of Radiology, (Special Issue 2): p. S fsl: FMRIB software library. Website(2011)