Volume registration using needle paths and point ... - IEEE Xplore

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 22, NO. 5, MAY 2003

653

Volume Registration Using Needle Paths and Point Landmarks for Evaluation of Interventional MRI Treatments Roee S. Lazebnik, Tanya L. Lancaster, Michael S. Breen, Jonathan S. Lewin, and David L. Wilson*, Member, IEEE

Abstract—We created a method for three-dimensional (3-D) registration of medical images (e.g., magnetic resonance imaging (MRI) or computed tomography) to images of physical tissue sections or to other medical images and evaluated its accuracy. Our method proved valuable for evaluation of animal model experiments on interventional-MRI guided thermal ablation and on a new localized drug delivery system. The method computes an optimum set of rigid body registration parameters by minimization of the Euclidean distances between automatically chosen correspondence points, along manually selected fiducial needle paths, and optional point landmarks, using the iterative closest point algorithm. For numerically simulated experiments, using two needle paths over a range of needle orientations, mean voxel displacement errors depended mostly on needle localization error when the angle between needles was at least 20 . For parameters typical of our in vivo experiments, the mean voxel displacement error was 0.35 mm. In addition, we determined that the distance objective function was a useful diagnostic for predicting registration quality. To evaluate the registration quality of physical specimens, we computed the misregistration for a needle not considered during the optimization procedure. We registered an ex vivo sheep brain MR volume with another MR volume and tissue section photographs, using various combinations of needle and point landmarks. Mean registration error was always 0.54 mm for MR-to-MR registrations and 0.52 mm for MR to tissue section registrations. We also applied the method to correlate MR volumes of radio-frequency induced thermal ablation lesions with actual tissue destruction. In this case, in vivo rabbit thigh volumes were registered to photographs of ex vivo tissue sections using two needle paths. Mean registration errors were between 0.7 and 1.36 mm over all rabbits, the largest error less than two MR voxel widths. We conclude that our method provides sufficient spatial correspondence to facilitate comparison of 3-D image data with data from gross pathology tissue sections and histology. Index Terms—Image guided therapy, interventional MRI, medical image processing, radio-frequency thermal ablation, volume registration.

Manuscript received June 11, 2002; revised December 13, 2002. This work was supported in part by National Institutes of Health (NIH) under Grant RO1CA84433. The work of R. S. Lazebnik was supported by a Whitaker Foundation Graduate Fellowship and the CWRU Medical Scientist Training Program. The Associate Editor responsible for coordinating the review of this paper and recommending its publication was D. Hawkes. Asterisk indicates corresponding author. R. S. Lazebnik, T. L. Lancaster, and M. S. Breen are with the Department of Biomedical Engineering, Case Western Reserve University, Cleveland, OH 44106-7207 USA. J. S. Lewin is with the Department of Radiology, University Hospitals of Cleveland, Cleveland, OH 44106 USA. *D. L. Wilson is with the Department of Biomedical Engineering, Case Western Reserve University, 10900 Euclid Avenue, Cleveland, OH 44106-7207 USA (e-mail: [email protected]). Digital Object Identifier 10.1109/TMI.2003.812246

I. INTRODUCTION

M

ANY studies require correlation of image data from medical scanners with measurements from excised tissue including photograph-based gross pathology and histological samples. Such analysis requires accurate spatial correspondence between the volumes of interest, achieved by image registration. Sometimes, insufficient common image features or anatomical landmarks preclude the successful application of many registration methods including point matching of anatomical landmarks or voxel gray scale techniques [1]. In addition, tissue sections and organs can deform during excision and fixation; hence, global rigid body registration methods might be unsuitable. We present a three-dimensional (3-D) registration method that overcomes these limitations. Our method involves insertion of fiducial needles, near the site of interest, and aligns needle paths, and optional point landmarks, to obtain local 3-D volume registration. This method is useful for a variety of animal model imaging studies, where inclusion of such fiducials is feasible. We are investigating two applications. First, we are applying the method to evaluate a minimally invasive interventional magnetic resonance imaging (iMRI) thermal ablation procedure [2], [3]. Studies of MR guided thermal ablation of pathologic tissue, using a radio-frequency (RF) energy source, reveal a change of the MR signal in the vicinity of the heat source [4]. In order to correlate changes observed in the MR signal to cellular response, we register the MR images with tissue sections of a physically sliced animal organ. Using an independent method, we also register histological samples to the tissue slices, facilitating an MR image to histology study [2]. Previous studies that correlate MR thermal ablation images with tissue destruction did not utilize 3-D volume registration techniques and, thus, did not allow for voxel to voxel correlation throughout the volume of interest [5]–[9]. Our method facilitates this comparison. Second, we are studying the spatio-temporal distribution of cancer drug concentrations from specially designed, controlled release implantable millimeter-sized rods using serial computed tomography (CT) studies[10]. Several methods are available for volume registration, but most are inappropriate for our applications. MR images and color photographs of excised tissue are very dissimilar, especially those from muscle tissue, as shown later. Grayscale registration methods, such as mutual information maximization, rely on the presence of at least some common features in both volumes [1]. Surface matching techniques are inappropriate due

0278-0062/03$17.00 © 2003 IEEE

654


to potential warping of the external surface or potential loss of significant regions during tissue excision. Further, because skin is commonly removed to aid tissue fixation, surface-landmark-based registration is inappropriate. Point-landmark-based methods, utilizing internal anatomical or artificially implanted features, are feasible but require the localization of landmarks common to both volumes. Internal point landmarks are often difficult to locate accurately in both volumes, due to voxel dimensions and/or contrast limitations. An alternative approach is registration based on extended anatomical features such as lines and planes [11]; however, identification of suitable landmarks is not feasible in many cases such as our experiments with quite homogeneous muscle tissue. In light of the above considerations, our method is useful when fiducial needles, or similar structures, can be inserted in vicinity of the volume of interest. In this paper we examine the influence of needle orientation and localization error on our method, develop methods for automatic evaluation of registration quality, and demonstrate our method’s application to images acquired from both an ex vivo specimen and during an iMRI procedure.

II. REGISTRATION ALGORITHM FOR CURVES AND POINTS We consider two volumes of images: a reference volume, which remains unchanged, and a nonreference volume, which will be resliced using volume interpolation. Both volumes feature the same fiducial needles and optional fiducial points. Registration consists of manual selection of points and needle paths and automatic matching of these features using an iterative optimization technique. Manual identification of features is performed using a preprocessing program having a graphical user interface. The program allows an operator to view cross sections of the unregistered volumes side by side, and manipulate image display options including window and level, as well as image magnification. Image slices are not parallel to the needles, allowing the user to choose the point in each slice where the needle intersects it. Each identified point is assigned a 3-D spatial location, in millimeters, relative to an origin located in the upper left corner of the volume’s first slice. The , , and coordinates are chosen such that and lie along the axes of each two-dimensional (2-D) slice, and lies along the axis perpendicular to the 2-D images. For photograph volumes, the coordinate is assumed to be that of the exposed face of each slab, but for MR volumes, it is assumed to lie midway through the slice thickness. After manual identification is complete, we initiate an automated optimization of the six registration parameters required for rigid body transformation: translations along three axes ( , , and ), as well as rotations about those axes [1]. For this we apply the iterative closest point algorithm, where the reference and nonreference volumes are considered the model and data respectively [12]. First, we create a parametric representation of each needle path in the nonreference volume by projecting each and planes and fitting a line to the data needle onto the in each plane using a least-squares criterion. The needle path through the nonreference volume is, thus, represented by two

2–D functions and , which allow the determination of an interpolated 3-D point at any value. We next determine corresponding point pairs between the unregistered volumes. For each of point landmarks, the corresponding point was chosen by the operator. For needle pairs, we examine each chosen point along a needle in the reference along the nonrefvolume and assume that the closest point erence fitted line corresponds. We compute a vector from any point along the line to , and a unit vector along the line. . This equation is exact for The closest point is straight lines; an iterative solution is required for more complex curves. As registration improves with iterative optimization, so does the validity of the assumption that the closest point is the true corresponding point [12]. We now define a scalar objective function that represents regbetween each of istration error in terms of the distances point landmark pairs and between each of the points along each of the needles

(1) We utilize the closed form quaternion-based approach described by Besl and McKay [12] to compute the rigid body transformation matrix that will minimize for the current set of correspondence points. We then apply the matrix to the nonreference volume point landmarks and needle points and repeat the above procedure. Because we take care to acquire images in a similar geometric fashion, the initial guesses for all transformation parameters relative to the reference are all zeros. A tolerance of 1E-8 is the termination criterion. The optimized transformation matrix is used to reslice the nonreference volume using tri-linear interpolation. For registration to a photograph volume, the photographs are the reference and image slices from a medical scanner are interpolated. III. EXPERIMENTAL METHODS We performed three types of experiments. First, we utilized digital phantoms to prove the algorithm and to determine the influence of needle orientation and manual localization on registration error. Second, a physical ex vivo sheep brain specimen was employed to determine the method’s applicability for registration of both MR to MR and MR to tissue sections. Third, we applied the method to image data acquired in iMRI thermal ablation experiments performed on a rabbit thigh model. This last type of experiment involved registration of in vivo MR to tissue section images, and tested all the practical complications of in vivo imaging. A. Imaging and Tissue Handling For MR imaging, we used a clinical open 0.2-T MR imaging system (Siemens Magnetom Open, Erlangen, Ger-weighted many). The rabbits were imaged using a 2-D spin-echo sequence [TR/TE/NEX 624/26/6, field of view (FOV) , slice thickness 3 mm] 180 180 mm, matrix -weighted spin-echo sequence (TR/TE/NEX and a 2-D

LAZEBNIK et al.: VOLUME REGISTRATION USING NEEDLE PATHS AND POINT LANDMARKS

3362/68/8, FOV 180 180 mm, matrix , slice thickness 3 mm). The sheep brain was imaged using a 2-D -weighted spin-echo sequence (TR/TE/NEX 624/26/50, , slice thickness FOV 125 125 mm, matrix 3 mm). For tissue sectioning and photography, we utilized a custom built apparatus described by Breen et al. [2]. The specimen was positioned in a polymeric geofoam block and held in place by tissue embedding wax (50 C–54 C Parablast X-tra, Oxford Labware, St. Louis, MO). A coloring agent was added to the translucent wax to increase contrast between tissue and wax background. The Styrofoam block was attached to a plastic platform with glue. Using a linear displacement device, the Styrofoam block was accurately advanced, in 3 mm increments, through the specimen’s entirety. After each step, we guided a 12.8-in tissue-slicing knife (Tissue-Tek Accu-Edge Semi-Disposable Autopsy Knife System, Sakura Finetek, Japan) along vertical supports to slice the front face of the block of tissue. The newly exposed face was photographed using a digital camera (Sony, Tokyo, Japan) and the image resolution was 0.1 mm/pixel. Tissue slicing was always performed approximately perpendicular to one of the fiducial needles. B. Evaluation of Registration Several methods were used to evaluate the quality of registration. For the digital phantoms, we knew the exact transformation parameters, as well as the estimated parameters. From the two transformations, we computed the average voxel displacement error over a region of interest. For registration of physical specimens, we employed other measures. First, we qualitatively compared registration of identifiable anatomical features. Using custom software, we displayed each slice of the reference and registered nonreference volume side by side. We then outlined various anatomical features on the reference images, copied the outline to the nonreference images, and noted how well the outline encloses the same feature. Second, for each point in the nonreference volume we locate the closest point along the corresponding path and compute the mean among all such distances. Third, when a needle path not utilized in the optimization was available, we locate corresponding points for this path as before and compute the mean distance between these points. This last error is not directly minimized during registration, and can thus, assay registration quality in the vicinity of the optimized needle paths. C. Effect of Needle Orientation and Localization Error on Registration Accuracy As described earlier, if needles are parallel then registration can fail. Using digital phantoms, we examined registration error as a function of the angle between the needles. For simplicity, we examined a worst case scenario of coplanar needles. That is, non co-planar configurations will never give parallel needles and will tend to restrict registration parameters more than co-planar ones. The phantom volume consisted of 13 slices, each containing mm, 64 64 voxels, where each voxel was and two “virtual” needles. This volume is approximately four times larger than the later described in vivo lesions and, thus,

655

more than encompasses the region of interest to be registered. The needles were co-planar and separated by an angle . The needle trajectories would intersect below the simulated volume. To eliminate the confounding error from manually localizing needle centers, we obtained error free coordinates of ten points along a straight-line equation for each needle. Each point was located at the intersection of the needle with the center of the slice thickness, digitized to the nearest voxel. The real valued reference points were translated and rotated in 3-D using a set of rigid body transformation parameters, yielding unregistered points for our tests. Transformation parameters were randomly chosen from uniform distributions ranging from 45 to 45 mm (equal to the in-plane FOV) and 45 to 45 . These ranges more than encompassed values encountered during the imaging experiments we performed. The two volumes were registered, using our software, and the process repeated for values of between 0 and 45 . At each value of , we used 1000 sets of six randomly generated transformation parameters. As described earlier, registration error was the mean Euclidean distance between voxels transformed using the optimized and actual parameters, over the entire simulated volume. We conducted a similar experiment that included the influence of needle localization error on registration accuracy. We generated digital phantoms as above and added, to each sampled reference and nonreference needle point, randomly generated displacements along both the and directions using a Gaussian distribution with standard deviation . This procedure was designed to roughly mimic manual user error in selecting a point within a slice. We registered 1000 volumes for each combination of and values, which ranged from 0.25 to 1.0 mm and from 5 to 45 , respectively. Again, these values exceeded what we find in practice. We computed the registration error as above. D. Registration of Ex Vivo Specimen Data A formalin preserved sheep brain that included the brainstem (Nebraska Scientific, Omaha, NE) was utilized to test both MR to MR and MR to tissue section registration. Because all images were acquired on fixed tissues, the only deformation that might be present occurred as a result of mechanically slicing the tissue. Needle centers were manually identified in MR images and tissue section photographs. To test optional point landmarks in MR-to-MR registration, we placed four doughnut shaped MR fiducial markers (IZI Medical Products, Baltimore, MD) in the FOV, with three on the specimen surface and another on the platform holding the specimen. The centers of the doughnut markers were easily identified in MR images and utilized as point landmarks. Three 18-G, 5-cm biopsy needles were inserted into the specimen, one through the cerebellum and the cerebrum’s long axis, and two at approximately 30 to the first needle. An MR volume was acquired perpendicular to the first fiducial needle and another perpendicular to the second needle. By acquiring images perpendicular to needles, one needle was always perpendicular to the image slice direction. This minimized the uncertainty in locating that needle’s center, as discussed later. Following MR imaging, the sheep brain was sectioned and photographed.

656


Multiple registration experiments were performed. We registered the two MR volumes using all combinations of only two of the three needles, three needles, only doughnut fiducial points, and all combinations of two needles with all fiducial points. We also registered the first MR volume to the physically sectioned volume using all combinations of two needles. E. Registration of In Vivo Thermal Ablation Data We applied our method to image volumes from MR guided RF thermal ablation experiments. These experiments tested all the realistic confounds of in vivo MR imaging, including needle placement, manual needle localization, and tissue photography, as well as the excision, fixation, and slicing of the tissue. The experiments were performed following a protocol approved by the Institutional Animal Research Committee. In each experiment, a New Zealand white rabbit was anesthetized, its abdomen and thigh shaved, and its legs secured to a custom-built Plexiglas support to prevent movement of the right thigh. The animal was positioned within the gantry of 12 the 0.2-T MR imaging system described above. Two 8 cm wire mesh grounding pads (Radionics, Burlington, MA), coated with conductive gel (Aquasonic 100: Parker Laboratories, Orange, NJ), were placed on the rabbit’s abdomen. The thighs were placed within a 12–cm-diameter multiturn solenoid receive-only coil. Subsequently, a MR-compatible 17-G copper RF electrode, with a 20-mm exposed tip (Radionics, Burlington, MA), was inserted percutaneously into the thigh muscle using MR image guidance. Once the electrode tip was positioned, a lesion was induced in the muscle tissue by increasing the local temperature, using RF electric current between the electrode tip and ground pads. RF energy was applied for 3 min using a 100-W RF generator operating at 500 kHz (model RFG-3C, Radionics, Burlington, MA). The tip of the RF electrode was maintained at 90 C 2 C. A thermistor within the electrode tip provided accurate instantaneous temperature information. Immediately after ablation, a 22-G 10-cm needle was inserted into the right thigh, near the thermal lesion, approximately 30 to the RF electrode. A second identical needle was also inserted in the vicinity of the lesion, approximately parallel to the electrode’s path. We later utilize this needle’s orientation to physically slice the tissue perpendicular to the lesion. Prior to imaging, the RF electrode was removed from the thigh to prevent the appearance of a susceptibility artifact -weighted images were acquired followed in the lesion. -weighted images. A 0.2-mmol/kg IV injection of by gadolinium contrast medium (gadopentate dimeglumine: Berlex Laboratories, Wayne, NJ) was administered 5 min prior -weighted image acquisition. The image volumes, to the approximately centered about the lesion and acquired perpendicular to the electrode path, consisted of 13 slices. The rabbit was sacrificed approximately 30 min post imaging using a barbiturate overdose technique. The entire pelvis and back limbs of the rabbit were dissected from the subject’s body and placed in a formalin bath for a week. Following fixation, the right thigh was dissected from the femur bone and placed in the formalin bath for an additional week. Subsequently, the thigh was sectioned approximately perpen-

Fig. 1. Voxel displacement error, as a function of , the angle between the needles, and the standard deviation (d) of the in-plane needle localization error. Each point represents the median registration error among 1000 registration trials where error is the mean Euclidean distance between corresponding voxels transformed using the optimized and actual parameters, over the entire simulated volume. Past a value of 20 , the error saturates and depends only upon (d). The dashed box encompasses the operating range typical for our in vivo iMRI experiments. Note that the lines connecting points are not model fitted.

Fig. 2. Registration success as a function of , the angle between the needles, and the standard deviation d of the needle localization error. A digital phantom was used as described in Section III, and successful registration was defined as a mean voxel error