Feasibility study for image guided kidney surgery: assessment of required intraoperative surface for accurate image to physical space registrations

Transcription

1 Feasibility study for image guided kidney surgery: assessment of required intraoperative surface for accurate image to physical space registrations Anne B. Benincasa a, Logan W. Clements a, S. Duke Herrell b, Sam S. Chang b, Michael S. Cookson b, Robert L. Galloway a,c a Dept of Biomedical Engineering, Vanderbilt University, 1225 Stevenson Ctr, Nashville, TN b Dept of Urologic Surgery, Vanderbilt University, A-1302 MCN Nashville, TN c Dept of Neurological Surgery, Vanderbilt University, T-4224 MCN, Nashville, TN ABSTRACT Currently, the removal of kidney tumor masses uses only direct or laparoscopic visualizations, resulting in prolonged procedure and recovery times and reduced clear margin. Applying current image guided surgery (IGS) techniques, as those used in liver cases, to kidney resections (nephrectomies) presents a number of complications. Most notably is the limited field of view of the intraoperative kidney surface, which constrains the ability to obtain a surface delineation that is geometrically descriptive enough to drive a surface-based registration. Two different phantom orientations were used to model the laparoscopic and traditional partial nephrectomy views. For the laparoscopic view, fiducial point sets were compiled from a CT image volume using anatomical features such as the renal artery and vein. For the traditional view, markers attached to the phantom set-up were used for fiducials and targets. The fiducial points were used to perform a point-based registration, which then served as a guide for the surface-based registration. Laser range scanner (LRS) obtained surfaces were registered to each phantom surface using a rigid iterative closest point algorithm. Subsets of each phantom s LRS surface were used in a robustness test to determine the predictability of their registrations to transform the entire surface. Results from both orientations suggest that about half of the kidney's surface needs to be obtained intraoperatively for accurate registrations between the image surface and the LRS surface, suggesting the obtained kidney surfaces were geometrically descriptive enough to perform accurate registrations. This preliminary work paves the way for further development of kidney IGS systems. Keywords: image guided surgery, kidney surgery, partial nephrectomy, abdominal procedures, surface-based registration, iterative closest point algorithm 1. INTRODUCTION There are about 30,000 new cases of kidney cancer detected each year in the U.S., and kidney resection, also known as a nephrectomy, is the only known curative treatment for this type of localized cancer. 1 Recent studies have demonstrated that a partial nephrectomy, either open or laparoscopic, with a clear margin is an effective procedure, especially for tumors less than 4 cm. 2-5 This nephron sparing procedure is imperative when the contralateral kidney is functionally impaired or has been surgically removed. 3, 5 Partial nephrectomies spare kidney function, thus enhancing surgical outcome. However, there are technical challenges associated with these procedures. Such obstacles include adequate intraoperative identification of the tumor, identification and control of the vascular supply, and avoidance of ischemic injury to the normal kidney tissue. 5 Currently, surgeons remove tumor masses using only direct or laparoscopic visualizations. This limited view prolongs the procedure and decreases the likelihood of a clear margin. The less the surgeons are required to disturb the kidney and its surrounding tissue during the procedure, the shorter the recovery time will be for the patient. Thus, there remains a need for surgeons to acquire additional intraoperative visualizations of the patient. Employing image guided surgery (IGS) could provide such representations in the operating room (OR). anne.b.benincasa@vanderbilt.edu

2 The goal of IGS is to provide surgeons with an accurate location of pathology in real time during a procedure. The efficacy of IGS has been demonstrated in areas such as neuro, ENT, spine, and non-spine orthopedic surgeries. All of these applications are based on the use of rigid anatomical landmarks on or near the site of surgery. Recent developments at Vanderbilt and other research programs have begun to make inroads into IGS of soft tissue. However, with open abdominal procedures, no such landmarks exist. Therefore, in abdominal procedures such as liver resections, anatomical features are used to drive registrations using surface descriptions. These surface properties are extracted preoperatively from a CT scan via segmentation and a 3D image volume of the organ of interest can be constructed. It has been demonstrated that surface acquisition can be accurately obtained in the OR using a laser range scanner (LRS).6 This method avoids contact with the patient, provides quick acquisition time, and produces a structured output data set. Nevertheless, these surface-based registrations rely upon an initial alignment given by a point-based registration using anatomical features as fiducials. Applying current IGS techniques, as those used in liver cases, to nephrectomy procedures presents a number of complications. Most notably is the limited field of view of the intraoperative kidney surface, which constrains the ability to obtain a surface delineation that is geometrically descriptive enough to drive a surface-based registration. Thus, using a LRS to obtain surface descriptions is hindered by the large amount of fat surrounding the kidney and the incision size made by the surgeon. This project aims to assess the feasibility of image guided kidney surgery with respect to the performance of image to physical space registration. The paper will focus on the validation of surface-based registrations using various surface amounts on a realistic kidney phantom arranged in two orientations representative of typical nephrectomies: one in the laparoscopic position and the other in the traditional partial nephrectomy position. Varying the amount of surface used to drive registrations will resemble varying amounts of visible surface data during a partial nephrectomy. Also, this paper will test sequential versus random sections of the phantom surface indicative of whether the threshold of accurate surface-based registrations is based on a percentage of total surface points or acquisition of descriptive surface properties. 2. METHODS In order to test the feasibility of extending the current image-guided surgery framework to kidney procedures, a realistic, to-scale kidney phantom was generated using silicon rubber (Smooth-On, The phantom accurately modeled typical geometrical surface properties such as curvature and smoothness. Two different phantom orientations were used to simulate the different orientations usually presented in the OR. For traditional partial nephrectomies, the patient is right/left lateral with the smooth, round back of the kidney facing upwards. To model this, a cradle was constructed using plexiglass and nylon screws to hold the phantom upright as depicted in Figure 1. Nine fiducial markers were screwed into the cradle and the centroids of these markers served as fiducial and target point sets. Markers 2, 5, 7, 8, and 9 served as targets and the other four were used as fiducials. Figure 1. Cradle constructed to provide typical orientation of kidney seen by surgeons during partial nephrectomies Figure 2. Laparoscopic orientation of kidney phantom The other setup had the phantom lying in its side and is show in Figure 2. This orientation was used to represent the view typically seen in laparoscopic cases. CT images of both phantom orientations were acquired. The kidney phantom CT images were segmented manually using Analyze AVW 6.0 (Mayo Clinic, Rochester MN, From the segmented images, the marching cubes algorithm was used to generate an initial approximation of the kidney

3 phantom s surface. 7 Fast RBF Toolbox (Farfield Technologies, was then used to define a parametric version of the marching cubes surface. This smooth surface was considered as the image surface. Fiducial point sets for the laparoscopic orientation were compiled from the CT image volume using anatomical features on the kidney phantom such as the ureter, renal artery and renal vein. For both phantom orientations, the fiducial points were used to perform a point-based registration, which then served as a guide for a surface-based registration. Physical surfaces were obtained using a LRS (3-D Digital Corp., Sandy Hook CT, and were registered to the image surface. The surface-based registrations used a rigid iterative closest point (ICP) algorithm formulated by Besl and McKay 8. In order to decrease closest point search times, k-d dimensional trees were used in the 9, 10 ICP implementation. These registrations were validated in order to characterize the effect of limited LRS field of view on the robustness of the surface-based registrations Laparoscopic orientation validation First, subsets consisting of a varying number of points of the total LRS surface (25,938 points) were segmented out to serve as a measure for the various views seen by the LRS, and these partial surfaces can be seen in Figure 3. Since current methods of IGS in soft tissue are so heavily reliant on the initial pose provided by point-based registration, the physical fiducial points were perturbed by a random normalized vector with magnitudes of 5, 10, 15, 20, and 25 mm. This methodology results in both a translation and rotation of the initial pose. These perturbations should reveal the effects of poor initial alignments given by the point-based registration. Next, each surface subset was registered to the image surface using the perturbed initial alignments. The rotations and translations from those surface-based registrations were used to transform the rest of the LRS surface not included in the subset in order to assess the accuracy of using partial surfaces to estimate the rest of the LRS surface s registration. The root mean square (RMS) of the distances between closest points on the image surface and the transformed points on each LRS Figure 3. Laparoscopic Orientation (whole) 6550 Segmentation of kidney into subsets of LRS surface surface without the subset was calculated to serve as a measure for error. The RMS error was averaged over 500 trials for each magnitude of perturbation. It is expected that RMS distances will decrease by increasing the amount of points in the LRS surface subset. RMS errors significantly greater than those found with using the total LRS surface should reflect inaccurate surface-based registrations Traditional orientation validation Similarly, subsets of the total LRS surface (11,802 points) were constructed to emulate the views seen in the OR. The LRS surface was divided into 6 patches as seen in Figure 4. Various combinations of these 6 patches were used as subsets of the LRS surface to examine sequential versus random patch combinations. As with the other phantom orientation, the physical fiducial points were perturbed by a normalized random vector of magnitudes 5, 10, 15, 20 and 25 mm to simulate poor initial alignments. The various surface patch combinations were registered to the image surface using the perturbed initial alignment. The translation and rotation from the surface-based registration given by each patch combination was used to transform the rest of the LRS surface as well as the five targets. Mean RMS errors were calculated as before, but this time using 1000 iterations. In addition, due Figure 4. Traditional Orientation Segmentation of kidney into 6 patches used to create subsets of LRS surface

4 to the ease of attaching targets on this phantom setup, target registrations errors (TRE) were averaged over 1000 trials for each magnitude of perturbation. This study should determine whether the threshold for an accurate registration is based on the amount of points in the subset or on sequential patches that capture enough of a surface's descriptive characteristics Laparoscopic orientation 4. RESULTS Mean of RMS error for 500 iterations (mm) RMS error vs. number of points used at varying magnitudes of perturbation (18%) 6550 (25%) 6496 (25%) 7683 (30%) (42%) (50%) (61%) (76%) 5 mm 10 mm 15 mm 20 mm 25 mm (100%) Number of points used for surface registration x 10 4 Figure 5. Laparoscopic Orientation Means of RMS errors over different magnitudes of perturbation Table 1 Mean ± standard deviation of RMS error for laparoscopic orientation of phantom Magnitude of perturbation (mm) # of points (% of total LRS surface) (18%) ± ± ± ± ± (25%) ± ± ± ± ± (25%) ± ± ± ± ± (30%) ± ± ± ± ± (42%) ± ± ± ± ± (50%) ± ± ± ± ± (61%) ± ± ± ± ± (76%) ± ± ± ± ± (100%) ± ± ± ± ± These preliminary results suggest a little less than half of the LRS surface is needed to drive an accurate surface-based registration. The mean values of RMS errors in millimeters can be seen in Figure 5. With 10,960 points (~42% of the total LRS surface), the RMS error drops significantly for all magnitudes of perturbation. In addition, the standard deviation also drops significantly with this surface subset, further supporting the accuracy of the registration. Table 1

5 shows the means and standard deviations of the RMS errors. Surfaces with percentages greater than 42% of the LRS scan produced similar results. Higher RMS values resulted for surfaces with fewer point sets. For 4,586 (18%), 6,496 (25%) and 6,550 (25%) points, RMS errors were very large (greater than a few millimeters) for all magnitudes of perturbation. Increasing the magnitude of perturbation increased the error since poor initial alignments will worsen registration results if the registrations are not robust. This is most evident with the 7,683 subset (30%) where for small magnitudes of perturbation the RMS error was on the order of 1 mm, but for higher magnitudes became greater than a few millimeters. The standard deviation also greatly increased for the 7,683 subset (30%), revealing its inability to yield an accurate registration. The perturbation effect was negligible with surfaces of 10,960 points (42%) and higher, implying that their surface-based registrations are robust. These findings suggest that accurate surface-based registrations require obtaining fractions of the surface that include 42% and higher. Therefore, when a kidney is oriented on its side, approximately half of that surface is geometrically descriptive enough to drive robust surface-based registrations Traditional orientation Mean TRE measured over 1000 iterations (mm) (10%) 2 (13%) 1 & 4 (26%) TRE vs. # of points at varying magnitudes of perturbation 2 & 5 10 (37%) 3 & 6 (38%) 2 4 & 5 (61%) & 4 (40%) 2 4 & 6 (53%) & 6 (74%) 4 & 6 (40%) whole 1 & 3 1 & & & & 6 (63%) (25%) (35%) (40%) (60%) Number of points used for surface-based registration 5 mm 10 mm 15 mm 20 mm 25 mm Figure 6. Traditional Orientation Means of TRE over different magnitudes of perturbation

6 Table 2 Mean ± standard deviation of TRE (mm) for traditional orientation of phantom Patch of LRS surface used Magnitude of perturbation (mm) patch # # of points (% of total LRS surface) (10%) ± ± ± ± ± (13%) ± ± ± ± ± and (25%) ± ± ± ± ± and (26%) ± ± ± ± ± and (35%) ± ± ± ± ± and (37%) ± ± ± ± ± and (38%) ± ± ± ± ± and (40%) ± ± ± ± ± ,2 and (40%) ± ± ± ± ± ,3 and (40%) ± ± ± ± ± ,4 and (53%) ± ± ± ± ± ,5 and (60%) ± ± ± ± ± ,2,4 and (61%) ± ± ± ± ± ,3,4 and (63%) ± ± ± ± ± ,3,5, and (74%) ± ± ± ± ± whole (100%) ± ± ± ± ± The results for the traditional orientation reveal more information on the criteria required for accurate registrations. Not all combinations were tried, but representative data is shown in Figure 6. Unlike the laparoscopic results, there is not a clear drop off in the error after a certain percentage of points are acquired. In this case, obtaining about 40% of the total LRS surface points yields varying results depending on the location of those points. Using as little as 35% of the LRS surface with patch combination 1 & 6 resulted in low TREs for all magnitudes of perturbation as well as low standard deviations. The means and standard deviations of the TREs calculated for the combinations of patches tried are shown in Table 2. Additionally, patch combination 1 & 3 (25%) produced smaller TREs than using larger portions of the LRS surface such as combinations 2 & 5 (37%) and 3 & 6 (38%). This result seemed to follow a trend that combinations of patches that included a section from the front and back produce much lower TREs with little variance, regardless of proper initial alignment. For example, the left front (patch 1) and the right back (patch 6) sections yielded the most favorable TREs for any 2 patches combined. The curvature information from the front and back is needed to lock the surface in. On the other hand, surface patch combinations that did not contain patches from both ends of the surface (1; 2; 1 & 4; 2 & 5; and 3 & 6) generated high TREs (the spikes in the graph) with larger standard deviations, which increased with higher magnitudes of perturbation. This finding suggests that the sides of the kidney are not as geometrically descriptive as the front and back. However, if a patch combination contains information from the front and back and the sides this will further decrease the mean TRE for all magnitudes of perturbation. Such patch combinations include 2, 4 & 6; 1, 2 & 3; and 4, 5 & 6. This outcome is also why the patch combination of 1 & 6 produced such low TREs. Thus, patch combinations that capture the most curvature produce the most favorable surfacebased registrations. The RMS error results followed the same pattern as the TRE results with the exception of patches 2 & 5 and 3 & 6. RMS data are shown in Figure 7 and Table 3. The RMS error for combination 2 & 5 was lower than for combination 3 & 6, whereas the TRE for combination 2 & 5 was higher than combination 3 & 6. The RMS errors were on the order of four millimeters, whereas TREs were much larger, on the order of 30 millimeters. Nevertheless, the RMS errors yielded the same implications as the TRE data, insinuating that the RMS errors for the laparoscopic orientation are a good estimate of actual TREs.

7 Mean of RMS error for 500 iterations (mm) (10%) 2 (13%) RMS error vs. # of points at varying magnitudes of perturbation 1 & 4 (26%) 1 & 3 (25%) 2 & 5 (37%) 3 & 6 (38%) 1 & 6 4 & 6 (40%) (35%) 1 2 & 3 (40%) Number of points used for surface-based registration Figure 7. Traditional orientation Means of RMS errors over different magnitudes of perturbation 1 3 & 4 (40%) 2 4 & 6 (53%) 4 5 & 6 (60%) 2 4 & 5 (61%) & 6 (63%) & 6 (74%) whole Table 3 Mean ± standard deviation of RMS errors for traditional orientation Region of LRS surface used Magnitude of perturbation (mm) patch # # of points (% of total LRS surface) (10%) ± ± ± ± ± (13%) ± ± ± ± ± and (25%) ± ± ± ± ± and (26%) ± ± ± ± ± and (35%) ± ± ± ± ± and (37%) ± ± ± ± ± and (38%) ± ± ± ± ± and (40%) ± ± ± ± ± ,2 and (40%) ± ± ± ± ± ,3 and (40%) ± ± ± ± ± ,4 and (53%) ± ± ± ± ± ,5 and (60%) ± ± ± ± ± ,2,4 and (61%) ± ± ± ± ± ,3,4 and (63%) ± ± ± ± ± ,3,5 and (74%) ± ± ± ± ± ,3,5, and (100%) ± ± ± ± ± 0.004

8 5. DISCUSSION These preliminary experiments suggest that image guided kidney surgery is feasible given that a geometrically descriptive surface is available intraoperatively. The laparoscopic orientation of the phantom gave promising results in that just a little less than half of the total LRS surface was needed to accurately predict a surface-based registration for the rest of the surface. Thus, when obtaining a surface in the OR when the kidney is in the orientation as shown in Figure 2, 42% of the surface visible in that figure should be acquired to ensure an accurate registration. This orientation has an unfair advantage over the traditional partial nephrectomy orientaiton because it exposes more descriptive properties of the kidney. This view of the surface contains anatomical features, such as the ureter and the renal aretery and vein. These features provide a more geometrically descriptive surface, resulting in lower errors. For instance, the surfaces with points 10,960 and higher contained the ureter and the renal aretery and vein, thus postulating an explanation for the success of these surface subsets. The lack of rigidly attached fiducials, visible in a CT scan and intraoperatively, will result in poor fiducial localization errors (FLE) and thus, poor initial alignments since the pointbased registration is formed using the fiducials. These errors could be minimized by constructing a similar setup as the traditional orientation that rigid affixes the kidney with fiducials and targets. Such a setup might result in even less than 42% of the LRS surface being necessary for accurate registrations. Additionally, a target and thus TRE computations will increase the validity of the laparoscopic orientation s results. Nevertheless, the RMS error calculations appear to be sufficient based on the comparison of TRE to RMS error for the traditional orientation. The findings from the traditional orientation experiment expose that a threshold for an accurate registration should not be determined by a percentage of surface points, but rather geometric surface properties, such as curvature. Obtaining roughly 40% of the surface resulted in various TREs: some low enough to regard the registration as accurate, yet some too high. Regions 1, 3, 4 and 6 contained more curvature information since they were part of the front and back of the kidney. Any combination of regions that contained at least one region from the front and back produced better registrations, regardless of the percentage of points. For example, using region combination 1 & 6 produced low TREs since it contained information from the front back and both sides of the kidney. The ends of the kidney were needed to lock the surface into place and ensure an accurate registration for the rest of the kidney. However, regions 2, 3, 5, & 6 and 1, 2, 4, & 5 produced relatively low TREs considering they contained a larger percentage of the LRS surface. These two region combinations did not contain information from both the front and back of the surface. Thus, points that covered more of the kidney surface, although not necessarily sequentially as in the laparoscopic orientation, produce lower TREs. Regions that did not contain information from both the front and back and were less than 50% of the surface were poor predictors for an accurate registration since they were not geometrically descriptive enough. Thus, a surgeon must unveil at least half of the surface that would be obtained if no surrounding tissue obstructed the kidney surface. Future experiments of the registration validation should include 1000 trials for the laparoscopic orientation. For this paper, 500 trials were chosen to shorten computation time. Tests similar to the traditional orientation should be completed to analyze the effects of point acquisition position on the laparoscopic orientation. These tests will elucidate the reason for the success of registrations from surface subsets containing at least 42% of the LRS scan. More possible combinations of the 6 regions for the traditional orientation should be tested to create a more complete data set. Additionally, different combinations of fiducials and targets could be used to see if the location of the targets varies the TREs. Since targets 7 and 8 are relatively far from the surface, larger TREs can be expected. Moving the targets closer to the surface should decrease TREs for all regions. Future studies should also create more realistic phantoms that model vasculature and perfusion effects as well as surrounding tissues such as fat tissue. 6. CONCLUSIONS The results presented here suggest that the surface obtained intraoperatively during a traditional partial nephrectomy is vital for the applicability of IGS to kidney procedures. If a surgeon is unable to remove enough of the surrounding tissue to expose a geometrically descriptive surface, then the resulting surface-based registrations will not be robust enough to predict an accurate registration for the entire kidney. Thus, under optimal conditions, such as maximized visible surface, image-guided kidney surgery is feasible. However, further steps must be completed to fully understand the requirements necessary to deliver accurate surface-based registrations. Nevertheless, this preliminary work paves the way for further development of an image-guided kidney surgery system.

9 ACKNOWLEDGEMENTS This work was supported by the NIH R44 Grant No. CA The authors would like to thank Debbie Deskins, Dahl Irving, and Jerry DeWitt in Vanderbilt University s Department of Radiology for their aid in acquiring CT images of the phantom. In addition, the authors would like to acknowledge the nursing staff of urology for their assistance in the OR. Some of the code was provided by Prashanth Dumpari, Dr. David Cash, and Dr. Tuhin Sinha. A number of algorithms developed in this work were developed using the Visualization Toolkit ( Also, the authors would like to thank Dr. Phillip Bao and Matthew Stockstill for their help in constructing the phantom. REFERENCES [1] Drucker B. J., "Renal cell carcinoma: current status and future prospects," Cancer Treatment Reviews, 31, pp , [2] Orvieto M., Chien G.W., Tolhurst S.R., Rapp D.E., Steinberg G.D., Mikhail A.A., Brendler C.B., Shalhav A.L., "Simplifying laparoscopic partial nephrectomy: technical considerations for reproducible outcomes," Adult Urology, 66, pp , [3] Fergany A. F., Hafez K.S., Novick A.C., "Long-term results of nephron sparing surgery for localized renal cell carcinoma: 10-year followup," Journal of Urology, 163, pp , [4] Godley P. A., Ataga K.I., "Renal cell carcinoma," Current Opinion in Oncology, 12, pp , [5] Vogelzang N. J., Stadler W.M., "Kidney Cancer," Lancet, 352, pp , [6] Cash D. M., Sinha T.K., Chapman W.C., Terawaki H., Dawant B.M., Galloway R.L., Miga M.I., " Incorporation of a laser range scanner into image-guided liver surgery: Surface acquisition, registration, and tracking," MEDICAL PHYSICS, 30, pp , [7] Lorensen W. E., Cline H.E., "Marching Cubes: A high resolution 3D surface construction algorithm," ACM Computer Graphics, 21, pp , [8] Besl P. J. a. M., N.D., "A method for registration of 3-D shapes," IEEE Trans. Pattern Anal. Mach. Intell., 14, pp , [9] Zhang Z. Y., "Iterative point matching for registration of free-form curves and surfaces," Int. J. Compu. Vis., 13, pp , [10] Friedman J. H., Bentley J.L., Finkel R.A., "An algorithm for finding best matches in logarithmic expected time," ACM Trans. Math. Softw., 3, pp , 1977.