Initial Investigation of an Automatic Registration Algorithm for Surgical ...

30th Annual International IEEE EMBS Conference Vancouver, British Columbia, Canada, August 20-24, 2008

Initial Investigation of an Automatic Registration Algorithm for Surgical Navigation Gregory J. Bootsma, Jeffrey H. Siewerdsen, Michael J. Daly, and David A. Jaffray 

Abstract—The procedure required for registering a surgical navigation system prior to use in a surgical procedure is conventionally a time-consuming manual process that is prone to human errors and must be repeated as necessary through the course of a procedure. The conventional procedure becomes even more time consuming when intra-operative 3D imaging such as the C-arm cone-beam CT (CBCT) is introduced, as each updated volume set requires a new registration. To improve the speed and accuracy of registering image and world reference frames in image-guided surgery, a novel automatic registration algorithm was developed and investigated. The surgical navigation system consists of either Polaris (Northern Digital Inc., Waterloo, ON) or MicronTracker (Claron Technology Inc., Toronto, ON) tracking camera(s), custom software (Cogito running on a PC), and a prototype CBCT imaging system based on a mobile isocentric C-arm (Siemens, Erlangen, Germany). Experiments were conducted to test the accuracy of automatic registration methods for both the MicronTracker and Polaris tracking cameras. Results indicate the automated registration performs as well as the manual registration procedure using either the Claron or Polaris camera. The average root-mean-squared (rms) observed target registration error (TRE) for the manual procedure was 2.58 +/0.42 mm and 1.76 +/- 0.49 mm for the Polaris and MicronTracker, respectively. The mean observed TRE for the automatic algorithm was 2.11 +/- 0.13 and 2.03 +/- 0.3 mm for the Polaris and MicronTracker, respectively. Implementation and optimization of the automatic registration technique in Carm CBCT guidance of surgical procedures is underway.

I

I. INTRODUCTION

MAGE-GUIDED

SURGERY (IGS) technology involving realtime tracking and navigation has received widespread use over the past few decades [1]-[6]. Traditionally these applications allow localization of interventional tools with respect to computed tomography (CT) and magnetic resonance (MR) images acquired preoperatively. Due to the nature of surgery, the surrounding morphology of the target site undergoes changes which are not accounted for in the preoperative images and the need for intraoperative imaging becomes apparent. Recent research in our lab has shown the potential benefits of intra-operative C-arm cone-beam computed tomography (CBCT) with a flat-panel detector  Manuscript received April 7, 2008. This work was supported in part by the National Institutes of Health under Grant R-01-CA127944. G. J. Bootsma is with the Department of Medical Biophysics, University of Toronto, ON M5G 2M9 CAN (phone: 416-946-4501 x4048; e-mail: gregory.bootsma@ rmp.uhn.on.ca). J. H. Siewerdsen, is with the Ontario Cancer Institute, Princess Margaret Hospital, Toronto, ON M5G 2M9 CAN. M. J. Daly is with the Ontario Cancer Institute, Princess Margaret Hospital, Toronto, ON M5G 2M9 CAN. D. A. Jaffray is with the Department of Radiation Physics, Princess Margaret Hospital, Toronto, ON M5G 2M9 CAN.

978-1-4244-1815-2/08/$25.00 ©2008 IEEE.

(FPD), offering sub-mm spatial resolution and soft-tissue visibility at low radiation dose [7]-[10]. The use of intra-operative imaging provides a means for updating the surgical navigation system with the morphological changes, but requires an additional registration of the navigation system with the updated images. This can add significant time to the surgical procedure. In an attempt to increase the efficiency of the registration procedure we have developed a novel automatic registration algorithm that automatically colocalizes the fiducials in the CBCT image and the camera (MicronTracker and Polaris) coordinate systems for use in a rigid body point-based registration algorithm. Previous researchers have created algorithms for automatic localization of fiducials in the image space [11], [12], but these algorithms require manual localization of the fiducials in the camera space for use in the registration of a surgical navigation system. An automatic registration procedure using a laser range scanner has also been developed [13], but this system does not explicitly incorporate the use of tracked surgical tools and requires additional segmentation of the image data prior to registration. Commercial navigation systems are available (Stryker, Kalamazoo, MI, USA) with automatic registration, but these systems use a skin surface matching approach that is subject to skin shifts and is not well suited to invasive procedures such as medial maxillectomy, skull base surgery, etc. Our automatic registration technique is integrated into an in-house software system (Cogito) developed in object-oriented C++ using open source libraries (Coin3D and SIM Voleon) for 3D visualization. Cogito supports both the Polaris and MicronTracker cameras and has been integrated with augmented reality tools. II. MATERIALS AND METHODS A.C-Arm Volumetric Imaging A prototype CBCT imaging system based on a mobile isocentric C-arm was developed in collaboration with Siemens Medical Solutions as an experimental platform for preclinical applications in image-guided interventions. The C-arm was modified to replace the x-ray image intensifier with a large-area flat-panel detector (FPD), motorized orbital drive for rotation of C-arm during image acquisition, a method for geometric calibration [9], and a computer control system for image acquisition and 3D reconstruction. The FPD is a PaxScan 4030CB (Varian, Palo Alto, CA) designed for real-time radiographic/fluoroscopic imaging [14], with a 2048x1536 (~40x30 cm2) active matrix of aSI:H photodiodes and thin-film transistors at 194 μm pixel

3638

pitch and a 600 μm thick CsI:Tl scintillator. The detector support multiple gain modes, with dual-gain interlaced utilized for all image acquisitions for the following experiments. The C-arm imaging geometry has a source-to-detector distance of 125.5 cm and a source-to-isocenter distance of 63.7 cm resulting in a magnification factor of ~1.97 and a field of view (FOV) of ~20x20x15cm3 at the isocenter. The image acquisition for the CBCT reconstructions consisted of 200 projections collected in ~60 s over the maximum orbital arc of ~178°, resulting in approximately 1.1 projections per degree. The volume reconstructions were performed using a modified FDK algorithm [15] for 3D filtered backprojection. The reconstructions used in the experiments had 256x256x192 voxels with a isotropic voxel size of 0.8mm. B.Optical Tracking Cameras The navigation software, Cogito, supports two different stereoscopic tracking cameras: the MicronTracker and the Polaris. In the case of the MicronTracker, multiple cameras can be used to create a large FOV by combining the FOV of each individual camera.

(a)

(c)

(b)

(d)

(e)

(f)

Fig. 1. Example hardware used in registration experiments: stereoscopic (a) MicronTracker and (d) Polaris cameras; (b) Tag and (f) IR reflective fiducial place on mount; (c) MicronTracker and (e) Polaris tracked tool.

The MicronTracker (Fig. 1.a) cameras used were the H40 (15 Hz measurement rate, 0.20 mm rms accuracy) and the H60 (15 Hz measurement rate, 0.35 mm rms accuracy). The main difference between the two models is the FOV and calibration accuracy, with the H60 having a slightly larger FOV but lower calibration accuracy than the H40. The MicronTracker, uses visible light for tracking high contrast

checkered markers. Fig. 1.c shows the tracked tool used during the registration experiments. Custom multi-modal fiducials, called Tags (Fig. 1.b), were designed that can be easily localized by the MicronTracker and in the CBCT volume for use with the auto-registration algorithm. The Tags consist of a plastic body (3mm thick, 35mm length, and 20mm height) created with a rapid prototyping machine and affixed with a checkered pattern which is tracked by the MicronTracker. To make it easily localized in the CBCT volume a 2 mm diameter steel ball bearing (BB) was implanted at a location geometrically related to the tracked checker pattern. The Polaris (Fig. 1.d) camera (60 Hz measurement rate, 0.35 mm rms accuracy) is an infrared (IR) stereoscopic camera that tracks passive reflective spheres under infrared exposure provided by a ring of IR LEDS around the lenses of the two cameras. Fig. 1.e shows the tracked tool used during the registration experiments for the Polaris camera. The passive reflective spheres were used as a multi-modal fiducial for localization in both the camera's coordinate system and the CBCT volume. A custom plastic mount for the reflective spheres was created using a rapid-prototyping machine (Fig. 1.f). C.Manual Registration Procedure As a basis of comparison for the auto-registration procedure, a conventional manual registration was performed using five CT/MR compatible adhesive skin fiducials (IZI Medical Products Corp., Baltimore MD, USA) attached to an anthropomorphic head phantom prior to CBCT image acquisition. After CBCT imaging, the fiducials were localized manually both in the CBCT image, FCBCT, and camera space, FCAM, using a tracked tool. A rigid registration (rotation, scale, and translation) was computed between FCBCT and FCAM using a least-squares solution where the rotation was solved using eigenvalue decomposition [16]. Once the coordinate systems had been registered, the transformation was updated according to the rotation and translation of a tool rigidly affixed to the phantom. To test the target registration error (TRE) of the manual registration, 10 plastic markers were placed randomly on the surface of the phantom and localized using the navigation system after the manual registration was performed. The location of the markers reported by the navigation system, xj, were compared to the actual location of the marker fiducials determined manually in the CBCT image, yj, to determine the observed TRE root-mean-square (rms) , where n

TRE=

1 ∑ ∣x − y j∣2 . n j=1 j

(1)

The manual registration was repeated five times on the same CBCT image, and the observed TRE was calculated for each trial.

3639

D.Automatic Registration Procedure The automatic registration algorithm was tested in a similar method to the manual registration. Fig. 2 shows the experimental setup for scanning the phantom using the Carm during the Polaris automatic registration experiment (Fig. 3). The automatic registration algorithm was tested for both the MicronTracker and the Polaris. Five fiducials were

iterations, N, in the algorithm is selectable by the user and should be chosen according to the size of the volume and template; for the Polaris and MicronTracker experiments N was set to 2 and 1, respectively. At each iteration, both the template and the volume are down-sampled using an average binning by a factor of 2n-i. This averaged binning gives an image pyramiding [17], [18] approach to the search and decreases the computational time required. Currently sub-voxel resolution is not incorporated into the scheme but could easily be added. On the first iteration, i=0, of the algorithm, when searching for the Polaris fiducial, the search region used is the entire volume and on subsequent iterations the search region is limited to areas around the points returned by the best match optimization procedure. In the case of the MicronTracker fiducials the search region is limited on the first iteration to regions around a volume mask created by thresholding the volume. The threshold values are set to find regions of intensity associated with the steel BB.

Fig. 2. Experimental setup for the Polaris automatic registration procedure.

placed in a similar arrangement to those used in the manual registration. After the fiducials were affixed a CBCT image with the fiducials in the FOV was acquired. The position of the fiducials in the cameras coordinate system, FCAM, was then acquired. The FCAM values were then passed into the auto-registration algorithm along with the CBCT image which would return an estimated location of the fiducials in the CBCT coordinate system, FCBCT. The same method as in the manual registration was used for performing registration between the two coordinate systems. Once registered, a tracked reference tool rigidly affixed to the phantom, was used to account for any motion of the phantom or camera. The same experiment to calculate the observed rms TRE in the manual registration was performed. E.Automatic Registration Algorithm The automatic registration algorithm (Fig. 4) is an iterative template matching algorithm that uses constraints to optimize the fiducial localization in the CBCT coordinate system, FCBCT. The algorithm considers a pre-defined fiducial template and a volume image that contains the object of interest plus the fiducials. The number of

Fig. 3. Procedure for performing automatic registration: (a) placement of either IR reflective or Tag fiducials on phantom; (b) CBCT acquisition using the C-arm with fiducials within the FOV; (d) localization of fiducials in camera's coordinate system, FCAM; (e) auto-locilization of fiducials in CBCT coordinate system, FCBCT, using algorithm; (f) registration of navigation system; (g) navigation and evaluation of TRE.

The template matching is a translation-only search using normalized cross correlation (NCC) as

3640

NCC u ,v ,w=

(2)

∑x , y , z [ f  x , y , z − f u , v , w ][t  x−u , y− v , z −w−t ]  ∑ x , y , z [ f  x , y , z − f u , v ,w ]2 ∑x , y , z [t  x−u , y−v , z−w −t ]2 where f is the image and the sum is over x, y, z under the template positioned at u, v, w, t is the mean of the template, f u , v , w is the mean of the image under the template. The reason for using a translation-only search is that the templates used are rotationally invariant. Future implementations could use a three-dimensional version of the fast normalized cross-correlation [19] to increase computational efficiency. For the Tags used in these experiments, the search considers a 2mm diameter spherical BB (rotationally

Once the template matching is complete, the position of the fiducials localized in the camera reference frame, FCAM, are transformed into a set of geometrical constraints to be used in conjunction with the NCC values to determine the best possible matches as well as sort each point according to its associated FCAM fiducial. In the case that i