Hybrid multiscale registration of planning CT images and daily CBCT images for adaptive and image-guided radiation therapy

Transcription

1 Hybrid multiscale registration of planning CT images and daily CBCT images for adaptive and image-guided radiation therapy Dana Paquin, Doron Levy, and Lei Xing May 10, 2007 Abstract Adaptive radiation therapy (ART) is the incorporation of daily images in the radiotherapy treatment process so that the treatment plan can be evaluated and modified to maximize the amount of radiation dose to the tumor while minimizing the amount of radiation delivered to healthy tissue. Deformable registration between the planning images and the daily images is thus an important component of ART. In this paper, we report our research on deformable registration of planning CT images the daily CBCT images. The registration algorithm is based on a multiscale approach in which the coarse scales of the images are registered with one another using a landmarkbased registration method, and the resulting transformation is fine-tuned by deformably registering the next scales with one another using the landmarkbased transformation as a starting point. This procedure is then iterated, at each stage using the transformation computed by the previous scale registration as the starting point for the current scale registration. We present the results of studies of head-neck and prostate CT-CBCT registration, and validate our registration method with comparisons of similarity measures before and after registration and with visual evaluation of the registration results. We also demonstrate that the multiscale method is more accurate for CT-CBCT registration than ordinary (non-multiscale) B-splines deformable registration. 1 Introduction Image-guided radiation therapy (IGRT) is the use of patient imaging before and during treatment to increase the accuracy and efficacy of radiation treatment. The Department of Mathematics, Stanford University, Stanford, CA ; paquin@math.stanford.edu Department of Mathematics, Stanford University, Stanford, CA ; dlevy@math.stanford.edu Department of Radiation Oncology, Stanford University, Stanford, CA ; lei@reyes.stanford.edu 1

2 goals of IGRT are to increase the radiation dose to the tumor, while minimizing the amount of healthy tissue exposed to radiation. As imaging techniques and external beam radiation delivery methods have advanced, IGRT (used in conjunction with intensity-modulated radiation therapy) has become increasingly important in treating cancer patients. Numerous clinical studies and simulations have demonstrated that such treatments can decrease both the spread of cancer in the patient and reduce healthy tissue complications [7], [8], [24]. IGRT is typically implemented in the following way. CT images are obtained several days or weeks prior to treatment, and are used for planning dose distributions, patient alignment, and radiation beam optimization. Immediately prior to treatment, CBCT images are obtained in the treatment room and are used to adjust the treatment parameters to maximize the radiation dose delivered to the tumor. This enables the practitioner to adjust the treatment plan to account for patient movement, tumor growth or movement, and deformation of the surrounding organs. To adjust the patient position and radiation beam angles based on the information provided by the CBCT images, the CBCT images must first be registered with the planning CT images. Ideally, an adaptive radiotherapy treatment (ART) will eventually be implemented in which the patient alignment and/or radiation beam angles are continually updated in the treatment room to maximize radiation dose to the tumor and minimize radiation to healthy tissue. Such a treatment program would require real-time multi-modality registration of images obtained during treatment with planning images obtained prior to treatment. Thus, registration of images acquired from different machines at different times is an important step in the adaptive treatment process, and incorporation of daily images into the radiotherapy treatment process is an essential step in ART [17], [26]. In a conventional CT imaging system, a motorized table moves the patient through a circular opening in the imaging device. See Figure 1 for an example. The machine depicted in Figure 1 is a Hitachi Medical Systems CXR-16 CT imaging system (image courtesy of Brooks Stevens Design and Hitachi Medical Systems). As the patient passes through the CT system, a source of x-rays rotates around the inside of the circular opening. The x-ray source produces a narrow, fan-shaped beam of x-rays used to irradiate a section of the body. As x-rays pass through the body, they are absorbed or attenuated at different levels, and image slices are reconstructed based on the attenuation process. Three-dimensional images are constructed using a series of two-dimensional slices taken around a single axis of rotation. In a CBCT imaging system, on the other hand, a cone-shaped beam is rotated around the patient, acquiring images incrementally at various angles around the patient. See Figure 2 for an example. The CBCT machine depicted in Figure 2 is an on-board imager (OBI) integrated in a Trilogy TM medical linear accelerator (image courtesy of Varian Medical Systems, Palo Alto, CA). The reconstructed data set is a three-dimensional image without slice artifacts, which can then be sliced on any plane for two-dimensional visualization. CBCT images contain low frequency components that are not present in CT images (similar to inhomogeneity related components in magnetic resonance images). One of the challenges in CT-CBCT image registration is thus to account for artifacts and other components that appear 2

3 Figure 1: A conventional spiral CT imaging system. Figure 2: A CBCT imaging system. in one of the modalities but not in the other. Numerous efficient and accurate methods for deformable image registration have been published [1], [12], [19], [20], [25]. See [3], [5] for a survey of deformable image registration techniques. Most image registration algorithms can be classified as either landmark-based or intensity-based. Landmark-based techniques [18] rely on the identification of corresponding points or features in the images to be registered. Although such methods are computationally easy to implement, the detection and mathematical characterization of anatomical landmarks is not fully automated. Thus, the landmarks must be user-supplied, and this can be a time-consuming and difficult process. Intensity-based methods, on the other hand, are fully automatic and do not require extraction or identification of physical features in the images. Mean squares and normalized correlation similarity measures are typically used for images of the same modality, and mutual information [4], [21] is used for registration of images of different modalities. In general, fast and fully automatic multi-modality deformable registration is still a challenging problem and is a subject of current 3

4 research [6], [11]. In our previous work [16], we developed a fast, accurate, and robust hybrid image registration technique which combines features of landmark-based registration and free-form deformable registration techniques. Our technique was motivated by the idea that correspondence between certain structures in images, such as bony structures, can be easily identified visually, while the correspondence between other regions, such as soft tissue, is less easy to see. Bony structures undergo rigid transformations, while soft tissue and muscular structures undergo deformable transformations. Thus, mapping of the two regions can be approached using different techniques. We have shown that the accuracy of our hybrid technique is independent of the precise correspondence of landmarks, and thus provides a significant improvement over ordinary landmark-based techniques. Additionally, we have demonstrated that our technique is robust when the images to be registered contain a substantial amount of noise. In this current work, we register planning CT and daily CBCT images using our hybrid technique. The motivation for applying our hybrid registration algorithm to CT-CBCT registration is that artifacts that appear, for example, in CBCT images but not in CT images can be treated as noise. Since our hybrid algorithm accurately registers extremely noisy images, we expect that it will also provide accurate registration of CT and CBCT images. The structure of this paper is as follows. In Section 2, we discuss our multiscale registration algorithm. In Section 3, we present several examples of CT-CBCT registration to demonstrate the accuracy of the hybrid registration technique. Section 4 concludes. 2 Materials and Methods We study registration of planning CT and daily CBCT images obtained from the Stanford University Medical Center. The CT images were acquired using a GE Discovery-ST Scanner (GE Medical Systems, Miluakee, WI), and the CBCT images were acquired using an on-board imager (OBI) integrated in a Trilogy TM medical linear accelerator (Varian Medical Systems, Palo Alto, CA). The data set included one head-neck patient case and one prostate patient case. The head-neck case consists of 80 planning CT images and 120 daily CBCT images, and the prostate case consists of 80 planning CT images and 60 daily CBCT images. To study the problem of deformable registration of planning CT and daily CBCT images, the following methods are investigated in this paper. In the following discussion, A represents the fixed image (in these examples, the CT image) and B represents the moving image (CBCT). 2.1 Hybrid multiscale registration Our hybrid registration algorithm is implemented as follows. 1. Decompose the images to be registered into a hierarchical series of coarse and fine scales using the multiscale (BV, L 2 ) image decomposition of Tadmor et al. 4

5 C1(A) Coarsest Scale of Image A Original Image A (Fixed Image) C2(A)... Cm(A) Finest Scale of Image A C1(B) Coarsest Scale of Image B Original Image B (Moving Image) C2(B) Cm(B) Finest Scale of Image B Figure 3: The multiscale image decomposition. C 1(A) C 1(B) Landmark Registration φ landmark Figure 4: The coarse-scale landmark-based registration step of the hybrid multiscale registration algorithm. [23]. See Figure 3. This decomposition separates an image into a series of scales in which the first scale consists only of the main shapes and general features of the image, and each additional scale includes increasingly finer amounts of detail and texture (until the final scale is a close approximation to the original image). In this paper, as in our previous works, we decompose each image into 12 scales. We note that the decomposition can be implemented in either 2 or 3 dimensions. For the details of the 2D implementation, see [23]. For details of the 3D implementation, see [15]. 2. Register the first coarse scales C 1 (A) and C 1 (B) of the images with one another using a standard landmark-based registration algorithm [18]. See Figure 4. The landmarks are taken as bony structures. Let φ landmark denote the resulting transformation. Since bony structures undergo rigid transformations, we require φ landmark to be an affine transformation (consisting only of translation, rotation, scaling, and shearing). Since the coarse scales contain only the main shapes and general features of each image, the identification of landmarks in the coarse scales is a relatively easy task, and can potentially be implemented automatically (for example, using the methods of [22]). 3. Next, use φ landmark as the starting point to deformably register the next coarse scales C 2 (A) and C 2 (B) with one another. In this work, we use φ landmark as the 5

6 φ landmark C 1 (A) Deformable Registration C 1 (B) φ 1 C 2 (A) φ 1 Deformable Registration φ 2 C 2 (B) C m (A) φ m 1 Deformable Registration C m (B) φ m Figure 5: The hybrid multiscale registration algorithm. input transformation for a B-splines free-form deformable registration [2], [10] algorithm. Let φ 2 denote the resulting transformation. 4. Use φ 2 as the starting point to deformably register C 3 (A) and C 3 (B) with one another. Let φ 3 denote the transformation obtained upon registering C 3 (A) with C 3 (B). Iterate this method, at each stage using the transformation computed by the previous scale registration algorithm as the starting point for the current registration. See Figure 5. For all of the B-splines deformable registration steps, we use a spline of order 3 and a grid size of Evaluation To visually evaluate the accuracy of our hybrid multiscale technique, we will compare checkerboard comparisons of corresponding slices of the CT and CBCT images after ordinary B-splines deformable registration and after hybrid multiscale registration. To quantitatively evaluate the accuracy of our hybrid multiscale registration algorithm for registration of CT and CBCT images, we shall compare the mutual information measure between the images before registration, after ordinary deformable registration, and after hybrid multiscale registration. Recall that mutual information is an information-theoretic metric which computes the amount of information that one random variable (here, image intensity) gives about another random variable (here, intensity values of another image). The mutual information of two images A and B is defined as follows: MI(A, B) = h(a) + h(b) h(a, B), (1) where h(x) = p(x) ln p(x) dx is the entropy of the random variable x, h(x, y) = p(x, y) lnp(x, y) dx dy is the joint entropy of the random variables x and y, p(x) 6

7 is the probability density of the random variable x, and p(x, y) is the joint probability density function of x and y. Mutual information has been studied extensively as a method of comparing images originating from multiple modalities in [4], [21]. The mutual information of two images increases when the images are geometrically aligned. If x and y are independent, then p(x, y) = p(x)p(y), so h(x, y) = h(x) + h(y). However, However, if there is any dependency (as would be the case if x and y are intensity values of images of the same object), then h(x, y) < h(x)+h(y). Note that MI(A, A) = h(a), so that the maximum possible mutual information between two images is the entropy of one of the images. 3 Results In each of the following examples, we present the results of our hybrid multiscale registration algorithm for CT-CBCT head-neck and prostate registration. In each example, we illustrate a slice of the CT image (upper left), the corresponding slice of the CBCT image (upper right), a checkerboard comparison of the images after ordinary B-splines deformable registration (lower left), and a checkerboard comparison of the images after hybrid multiscale registration (lower right). We have highlighted areas of misregistration in the checkerboard images after ordinary registration with arrows. In particular, we notice misalignment on the edges and in the bony structure regions. To demonstrate the accuracy and applicability of our method, we have illustrated example slices which contain different anatomical features. 3.1 Head-neck CT-CBCT registration In Figures 6, 8, and 9, we present three head-neck registration examples. To highlight the misregistration obtained in the bony structures after ordinary deformable registration, in Figure 7 we present a magnified view of the misaligned areas for example 1 after ordinary deformable registration and after hybrid multiscale registration. 3.2 Prostate CT-CBCT registration In Figures 10, 12, 13, and 14, we present four prostate registration examples. To highlight the misregistration obtained in the bony structures after ordinary deformable registration, in Figure 11 we present a magnified view of the misaligned areas for example 1 after ordinary deformable registration and after hybrid multiscale registration. 3.3 Mutual information similarity measures In Figures 15 and 16, we present the mutual information similarity measures between the planning CT s and daily CBCT s before registration (circles), after ordinary B- splines deformable registration (squares), and after hybrid multiscale registration 7

8 Figure 6: Head and neck example 1. Planning CT image (upper left), daily CBCT image (upper right), checkerboard comparison after ordinary B-splines deformable registration (lower left), checkerboard comparison after hybrid multiscale registration (lower right). 8

9 Figure 7: Head and neck example 1. A magnified view of the checkerboard comparison of the bony structures after ordinary deformable registration (upper) and after hybrid multiscale registration (lower). 9

10 Figure 8: Head and neck example 2. Planning CT image (upper left), daily CBCT image (upper right), checkerboard comparison after ordinary B-splines deformable registration (lower left), checkerboard comparison after hybrid multiscale registration (lower right). 10

11 Figure 9: Head and neck example 3. Planning CT image (upper left), daily CBCT image (upper right), checkerboard comparison after ordinary B-splines deformable registration (lower left), checkerboard comparison after hybrid multiscale registration (lower right). 11

12 Figure 10: Prostate example 1. Planning CT image (upper left), daily CBCT image (upper right), checkerboard comparison after ordinary B-splines deformable registration (lower left), checkerboard comparison after hybrid multiscale registration (lower right). 12

13 Figure 11: Prostate example 1. A magnified view of the checkerboard comparison of the bony structures after ordinary deformable registration (upper) and after hybrid multiscale registration (lower). 13

14 Figure 12: Prostate example 2. Planning CT image (upper left), daily CBCT image (upper right), checkerboard comparison after B-splines deformable registration (lower left), checkerboard comparison after hybrid multiscale registration (lower right). 14

15 Figure 13: Prostate example 3. Planning CT image (upper left), daily CBCT image (upper right), checkerboard comparison after B-splines deformable registration (lower left), checkerboard comparison after hybrid multiscale registration (lower right). 15

16 Figure 14: Prostate example 4. Planning CT image (upper left), daily CBCT image (upper right), checkerboard comparison after B-splines deformable registration (lower left), checkerboard comparison after hybrid multiscale registration (lower right). 16

17 A f t e r m u lt is c a le r e g is t r a t io n A f t e r o r d i n a r y r e g is t r a t io n B e f o r e r e g is t r a t i o n Figure 15: The mutual information similarity measures between the head-neck planning CT s and daily CBCT s before registration, after B-splines deformable registration, and after hybrid multiscale registration (crosses) for all slices considered for the head-neck and prostate registration examples. For all cases, the similarity measures increased after hybrid multiscale registration. 4 Conclusions We have demonstrated the accuracy of our hybrid multiscale registration algorithm [16] for registration of planning CT images and daily CBCT images with several head-neck and prostate registration examples. The hybrid multiscale algorithm is motivated by the observation that different anatomical structures in the images to be registered undergo different types of transformations, and thus mapping of the different regions should be approached differently. We use a multiscale implementation to increase computational efficiency and accuracy and to prevent the registration algorithm from reaching local minima. References [1] R. Bajcsy and S. Kovacic, Multiresolution elastic matching, Computer Vision Graphics and Image Processing, 46, pp. 1 21, [2] F. L. Bookstein, Principal Warps: Thin-Plate Splines and the Decomposition of Deformations, IEEE Transactions on Pattern Analysis and Machine Intelligence, 11, no. 6, pp ,

18 A f t e r m u lt is c a le r e g is t r a t io n A f t e r o r d i n a r y r e g is t r a t io n B e f o r e r e g is t r a t i o n Figure 16: The mutual information similarity measures between the prostate planning CT s and daily CBCT s before registration, after B-splines deformable registration, and after hybrid multiscale registration [3] L. Brown, A Survey of Image Registration Techniques, ACM Computing Surveys, 24, no. 4, pp , [4] A. Collignon, D. Vadermeulen, P. Suetens, and G. Marchal, 3d multi-modality medical image registration based on information theory, Computational Imaging and Vision, vol. 3, pp , [5] W. R. Crum, T. Hartkens, and D. L. G. Hill, Non-rigid image registration: theory and practice, The British Journal of Radiology, 77, pp , [6] J.C. Gee, A. Maintz, and M.W. Vannier, Biomedical Image Registration: Revised Papers, Second International Workshop, WBIR 2003 (Philadelphia, PA, USA: Springer). [7] C.W. Hurkmans, B.C.J. Cho, E. Damen, et al., Reduction of cardiac and lung complication probabilities after breast irradiation using conformal radiotherapy with or without intensity modulation, Radiotherapy Oncology, vol. 62, pp , [8] R. Jacob, A. Hanlon, E. Horowitz, et al., The relationship of increasing radiotherapy dose to reduced distant metastases and mortality in men with prostate cancer, Cancer, vol. 100, pp ,

19 [9] S. Lee, G. Wolberg, K.-Y. Chwa, and S. Y. Shin, Image metamorphosis with scattered feature constraints, IEEE Transactions on Visualization and Computer Graphics, 2, pp , [10] S. Lee, G. Wolberg, and S. Y. Shin, Scattered data interpolation with multilevel B-splines, IEEE Transactions on Visualization and Computer Graphics, 3, pp , [11] B. Likar and F. Pernus, A hierarchical approach to elastic registration based on mutual information, Image and Vision Computing, 19, pp , [12] W. Lu, M. Chen, G.H. Olivera, K. Ruchala and T.R. Mackie, Fast free-form deformable registration via calculus of variations, Physics in Medicine and Biology, 49, pp , [13] J. Modersitzki, Numerical Methods for Image Registration, Oxford, [14] D. Paquin, D. Levy, E. Schreibmann, and L. Xing, Multiscale image registration, Mathematical Biosciences and Engineering, 3, no. 2, pp , [15] D. Paquin, D. Levy, and L. Xing, Multiscale deformable registration, IEEE Transactions on Medical Imaging, submitted. [16] D. Paquin, D. Levy, and L. Xing, Hybrid multiscale landmark and deformable image registration, Mathematical Biosciences and Engineering, accepted for publication. [17] H. Rehbinder, C. Forsgren, and J. Lof, Adaptive radiation therapy for compensation of errors in patient setup and treatment delivery, Medical Physics, 31, pp , [18] K. Rohr, Landmark-based image analysis, Computational Imaging and Vision, Kluwer Academic, Dordrecht, [19] E. Schreibmann and L. Xing, Narrow band deformable registration of prostate magnetic resonance imaging, magnetic resonance spectroscopic imaging, and computed tomography studies, International Journal of Radiation Oncology Biology Physics, 62, pp , [20] sc J.P. Thirion, Image matching as a diffusion process: an analogy with Maxwell s demons, Medical Image Analysis, 2, pp , [21] P. Viola, W. Wells, H. Atsumi, S. Nakajima, and R. Kikinis, Multimodal volume registration by maximization of mutual information, Medical Image Analysis, vol. 1, no. 1, pp , [22] E. Schreibmann and L. Xing, Image registration with auto-mapped control volumes, Medical Physics, 33, no. 4, pp ,

20 [23] E. Tadmor, S. Nezzar, and L. Vese, A multiscale image representation using hierarchical (BV, L 2 ) decompositions, Multiscale Modeling and Simulations, 2, no. 4, pp , [24] M. Tubiana and F. Eschwege, Conformal radiotherapy and intensitymodulated radiotherapy: Clinical data, Acta Oncology, vol. 39, pp , [25] H. Wang, L. Dong, J. O Daniel, R. Mohan, A.S. Garden, K.K. Ang, D.A. Kuban, M. Bonnen, J.Y. Chang, and R. Cheung, Validation of an accelerated demons algorithm for deformable image registration in radiation therapy, Physics in Medicine and Biology, 50, pp , [26] D. Yan, F. Vicini, J. Wong, and A. Martinez, Adaptive radiation therapy, Physics in Medicine and Biology, 42, pp ,