3D Image-guided Robotic Needle Positioning System for Small Animal Interventions

3D image-guided robotic needle positioning system for smallanimal interventions

Jeffrey S. Baxa) and Christopher S. R. WaringBiomedical Engineering Graduate Program, The University of Western Ontario, London, Ontario N6A 5B9,Canada and Robarts Research Institute, The University of Western Ontario, London, Ontario N6A 5K8, CanadaShi SherebrinRobarts Research Institute, The University of Western Ontario, London, Ontario N6A 5K8, CanadaShawn StapletonPrincess Margaret Hospital/University of Toronto, Radiation Medicine Program, Toronto,Ontario M5G 2M9, Canada

Thomas J. HudsonRobarts Research Institute, The University of Western Ontario, London, Ontario N6A 5K8, CanadaDavid A. JaffrayPrincess Margaret Hospital/University of Toronto, Radiation Medicine Program, Toronto,Ontario M5G 2M9, Canada

James C. Lacefield and Aaron FensterBiomedical Engineering Graduate Program, The University of Western Ontario, London, Ontario N6A 5B9,Canada and Robarts Research Institute, The University of Western Ontario, London, Ontario N6A 5K8, Canada(Received 8 June 2012; revised 13 November 2012; accepted for publication 27 November 2012;published 4 January 2013)Purpose: This paper presents the design of a micro-CT guided small animal robotic needle posi-tioning system. In order to simplify the robotic design and maintain a small targeting error, a novelimplementation of the remote center of motion is used in the system. The system has been developedwith the objective of achieving a mean targeting error of

011909-2 Bax et al.: Image-guided robotic needle positioning system 011909-2

I. INTRODUCTION

Clinical medical robotics is a mature field and dozens of clin-ical robotic systems have been developed for use in a widerange of interventional applications.1, 25 Today, use of medicalrobotics is increasingly becoming part of routine procedures,for example, the da Vinci robot by Intuitive Surgical for radi-cal prostatectomy.2 This growing use and importance of med-ical robotic systems is a stark contrast to the state of roboticsfor use with small animals in preclinical research. A partic-ular preclinical application, which could greatly benefit fromthe use of robotics, is the development of an image-guidedrobotic system for needle interventions. Although a numberof systems have been developed for image-guided clinicalneedle interventions,1 no such systems are in routine preclin-ical use. Rather, suboptimal nonrobotic and nonimage-guidedtechniques remain the norm for small animal needle interven-tions. Techniques typically used for small animal needle inter-ventions require surgical exposure of targets,36 percutaneousinjections through the skin3, 7 or stereotactic devices.8 Surgi-cal exposure suffers from associated surgical mortality andmorbidity, which may confound research results. Both per-cutaneous and surgical techniques are ultimately highly de-pendent on the ability of a human operator to correctly placea needle and suffer from problems of accuracy and repeata-bility. Stereotactic devices are limited to interventions withinthe skull and are limited by the accuracy of anatomic atlasesand localization of external landmarks. The current methodstypically used for small animal needle interventions are rela-tively unsophisticated in comparison to clinical methods.

Imaging technology has outperformed robotics in the de-velopment of specialized small-animal systems for preclini-cal research. Analogous small-animal imaging systems havebeen developed for all the major clinical imaging modalitiesincluding: computed tomography (CT), magnetic resonance(MR), positron emission tomography (PET), single-photonemission computed tomography (SPECT), and ultrasound.These small-animal imaging systems have achieved popularuse and are considered to have greatly contributed to preclini-cal research.9 Micro-CT is a particular imaging modality ofinterest and is commonly used to image small-animals be-cause of its ability to provide submillimeter resolution im-ages of soft tissue and skeletal structures. Beyond preclinicalapplications, micro-CT has also expanded to image humananatomy such as the wrist and tibia.9 Micro-CT scanners areavailable from at least a dozen manufacturers10 with typicalvoxel sizes ranging from 5 m to 450 m and trans-axialfields-of-view ranging from 1 to 20 cm.11, 12

To improve small animal needle interventions, two previ-ous efforts have been made to integrate robotic devices withmicro-CT imaging systems to perform image-guided needleinterventions.1315 In addition, three further systems not ex-plicitly intended for micro-CT imaging-guidance have alsobeen developed for small animal needle interventions andcould be potentially integrated with micro-CT scanners.1620The development of these systems combines the accurate andnoninvasive target localization of imaging with the position-ing accuracy and repeatability of robotic systems. The design

requirements these devices must satisfy are demanding. Aneedle positioning error of


workspace.25 The RCM architecture is well established withinclinical robotic systems.24 In a RCM based robot architec-ture, all of the rotational axes of the robot intersect at a com-mon point in 3D space that is outside the robots structure.This architecture allows for higher angular mobility in a con-fined space such as a scanner bore. A RCM design also hasthe added advantage of allowing needle translation and ori-entation to be decoupled when positioning the needle for in-terventions. The decoupling of translation and orientation nolonger makes it necessary to simultaneously control multipledegrees-of-freedom during the most delicate part of proce-dures: needle insertion.

A comparative analysis was performed to determine theRCM architecture best-suited for the kinematic frame. Thedesigns considered were: the double parallelogram linkage,15the goniometric arc,20 and the spherical linkage.22, 23 Thedouble-parallelogram design is unsuitable for this applica-tion because of its size and number of components required.The goniometric arc is a simpler design than the double-parallelogram. However, it is difficult to manufacture a lin-ear bearing from CT compatible materials to support the toolin a goniometric arc design. The high-density materials typ-ically used in most linear bearings will generate streak arti-facts because of high attenuation and nonferric CT compati-ble materials typically lack rigidity, which results in bearingdeflection and friction. The spherical linkage is advantageousover the other two options due to its simplicity of design. Thespherical linkage design also allows for ease in adjustabilityand calibration to create a precise RCM independent of themanufacturing tolerances in each part of the linkage.

II.A.2. Mechatronic system descriptionIn previous robotic designs, the RCM is typically placed

on the skin surface of the animal during interventions. In thisdesign, the mechatronic system instead positions the RCM tocorrespond to the position of the target inside of the animal.The shift in RCM position from the skin surface to the tar-get position offers two major benefits. First, the needle driverto insert and retract the needle will only be required to posi-tion the needle tip at one position in space (the RCM), sincethe RCM position always corresponds with the target posi-tion. As a result, the needle driver design can be simplifiedand miniaturized to operate within the micro-CT bore. Sec-ond, the needle tip will always be inserted at the position ofthe RCM regardless of the needle angulation. Therefore, nee-dle translation and needle angulation can be decoupled. Asa result, in free air the targeting error becomes independentof the needle orientation and targeting accuracy is no longerdependent on the calibration of the robots kinematics.

The mechatronic system achieves this RCM implemen-tation using six motorized degrees-of-freedom. Three lineardegrees-of-freedom are contained in a XYZ linear stage. Thestage is used to translate the entire system, which in turn trans-lates the RCM to the target positions. Two rotational degrees-of-freedom, roll and pitch, are contained in two mirror im-age spherical linkages at the front and the rear of the system.The rotational axes are used to adjust needle orientation dur-

FIG. 1. A schematic representation of the proposed RCM linkage design.For labels L1L5 and l1 l5 straight lines indicate the kinematic overlay andlines with arrows indicate the actual physical links.

ing interventions. The forward and rear linkages are coupledtogether through an elongated shaft assembly. The shaft as-sembly and associated linkage functions as a pantograph toallow the rear linkage to counterbalance the forward linkage.Finally, one linear degree-of-freedom is contained in a nee-dle driver mounted onto the forward spherical linkage. Theneedle driver is used to insert and retract the needle duringinterventions.

Figure 1 shows a schematic representation of the RCMlinkage design and its two constituent parts: the forward (la-beled L1L5) and the rear spherical linkage (labeled l1l4).The spherical linkages can be manipulated using either mo-tors or through manual manipulation of a handle mounted tothe rear spherical linkage. The full range of motion of thespherical linkage is illustrated in Fig. 2. The forward spheri-cal linkage consists of five links (L1L5) and five hinged con-nections (R1-R5) pinned to the base (L0). The rear linkage isa mirror image of the forward linkage and consists of fourlinkage elements (l1-l4) and four hinged connections (r1-r4).The axis of each hinged connection in the spherical linkagesconverges to a common point in space to form a (RCM) atthe forward spherical linkage and (rcm) at the rear linkage.The extra pinned connection (R5) in the forward assemblywas found to be required to obtain a precise RCM indepen-dent of the manufacturing tolerances in each link. Machin-ing tolerances resulted in the misalignment of the axes of thepinned connections (R3 and R4) in the base link (L1 and L5).Adjustment of the pinned connection at (R5) allows for themisalignment between the two axes to be corrected. Figure 3shows the adjustability of the extra pinned connection.

To determine the stability of the RCM following adjust-ment, a precision 6.35 mm diameter tooling ball (29011, Jer-gens Inc., Cleveland, OH) was attached to the linkage and

Medical Physics, Vol. 40, No. 1, January 2013


Path of

Needle Axis

Roll:

Right to Left

Path of

Needle Axis

Roll:

Left to RightPitch:

Up

Path of

Needle Axis

Path of

Needle AxisPitch:Down

FIG. 2. Strobe diagram of the robot spherical linkage illustrating the two possible degrees of freedom: roll and pitch. To control the linkage manually, the userwould manipulate the joystick at the rear.

aligned to the RCM as illustrated in Fig. 3. The location ofthe tooling ball was estimated from the design drawings rel-ative to a fixed point on the arm. The deviation of the actualdistance of the tooling ball to the fixed point from the ex-pected design drawing distance was 12.5 m. The tooling balldisplacement throughout the spherical linkages full range ofmotion was measured using a calibrated XYZ stage (M-462,Newport Corp., Irvine, CA) and attached indicator (Model

Tooling Ball

Joint Motion

Tooling Ball Fixture

FIG. 3. Strobe diagram of the robot forward spherical linkage with links L2and L4 removed. This shows the adjustability of the extra pinned connection,used to create a precise RCM independently of the manufacturing tolerancesin links L1, L5, and L3. To validate the accuracy of the mechanically con-strained RCM, a tooling ball was mounted to the arm L3 as shown.

25164-10, The L. S. Starrett Company, Waite Park, MN). Themeasured tooling ball deviation was


-CT Scanner

Robot Control Box

Needle Positioning Robot

Forward Spherical Linkage

Needle Driver

Mouse Bed

Fiducials

x-y-z Stage

FIG. 4. (Top) Photograph of the robotic apparatus mounted on the CT scan-ner animal couch and control system. The mechatronic apparatus consists ofan XYZ linear stage that supports the spherical linkage. (Bottom) Photographof the forward spherical linkage and attached needle driver. The mouse bed isattached to a double ball joint, which in turn is clamped to the animal couchvia a pair of hollow aluminum rails giving a total of six degrees-of-freedom.The robot coordinate system consists of a z-axis along the bore of the micro-CT, and x- and y-axes aligned to the horizontal and vertical, respectively.

The electronics of the robot are divided into two separatecontrol systems: one system for the XYZ stage and one sys-tem for the spherical linkage and needle driver. The sphericallinkage and needle driver are driven by stepper motors con-trolled by a multiaxis dual-loop controller (MAXNet, Pro-Dex Oregon Micro Systems, Oregon, WA) and stepper mo-tor drivers (BSD-01v2, Interinar Electronics, Tampa, FL). TheXYZ stage is powered by three stepper motors coupled to leadscrews. The stepper motors are controlled by a multiaxis con-troller with integrated stepper motor drivers (DMC2133 withSDM-20242, Galil Motion Controls, Rocklin, CA). The linearencoders for each axis of the XYZ stage feed into the multi-axis controller. Custom closed-loop software on the controllermonitors the encoders to compare the target position of eachaxis against their desired position. Both control systems areinterfaced to a host PC via an Ethernet hub. Custom softwareon the host PC sends programs and commands to each of thecontrollers. The user enters the desired position and orienta-tion of the needle to the PC. The software then uses appropri-ate inverse kinematic equations to calculate the appropriatecommands to send to each controller.

II.B. Robot calibrationII.B.1. Coordinate system calibration

The coordinate system of the robot was calibrated to en-sure accurate control of the robot position. The coordinatesystem of the robot is aligned to the three axes of the XYZlinear stage and tracks the position of the devices RCM. TheRCM position was calibrated by repositioning the XYZ linearstage at 7.5 mm increments along each axis: covering a totalrange of 15.0 mm, 15.0 cm, and 7.5 mm. At each po-sition, the encoder count for each of the XYZ linear stagesthree axes was recorded and the distance traveled by thestage measured using a depth gauge (Model 2776S, MitutoyoCanada, Toronto, ON) mounted onto a calibrated manualXYZ stage (M-462, Newport Corp., Irvine, CA). The encoderstep size for the XYZ linear stage was calculated by averag-ing the measurements along each axis. The flatness, straight-ness, and perpendicularity of each axis of the XYZ linearstage were also verified using a granite surface plate (GradeB-18, The L. S. Starrett Company, Waite Park, MN), indica-tor (Model 25164-10, The L. S. Starrett Company, Waite Park,MN) and the same calibrated manual XYZ stage used earlierto determine the stability of the RCM. Flatness was evalu-ated by running the indicator along the surface of the stagefor each of the three translational axes and recording any de-viations. Straightness was evaluated by placing the indicatoronto a surface perpendicular to a translational axis. The stagewas then advanced a known distance along the axis and thisdistance compared to the distance measured by the indica-tor. The straightness measurement was completed three times,once for each of the translational axes. The perpendicularitywas evaluated by placing the indicator against surfaces paral-lel to a translational axis. The stage was then advanced alongthe axis and the indicator recorded for any motion. The de-viation of the flatness, straightness, and perpendicularity was:1.81 m, 0.77 m, and 1.1 min (1 min = 1/60), respec-tively. These values were all measured over 20 mm of stroke.

The angular orientation of each of the two arms (Fig. 1,L1 and L2) in the forward spherical linkages was also cal-ibrated. The robotic system was first placed onto a granitesurface plate, which served as a reference plane. Each of thetwo arms was independently oriented so that one arm was in aplane perpendicular and one arm was in a plane parallel to thegranite surface. Each arm was adjusted to within 2.5 m ofthe perpendicular or parallel plane using the same calibratedmanual XYZ stage and indicator as earlier. The encoder val-ues of the two encoders in the spherical linkage were recordedwith the arms in this orientation. The two arms were then ro-tated 90 once. The encoder values of the second orientationwere recorded giving the step size and absolute reference ofthe encoder home position.

II.B.2. Optical needle tip calibration to RCMTo complete an intervention, the RCM of the robot is trans-

lated to correspond with the localized position of the target.It is therefore essential that the needle tip reaches the RCMwhen it is at the inserted position. Thus, a method to calibrate



x

y

x

y

FIG. 5. Composite image of the calibration photographs showing the roll ofthe needle throughout its full range of motion. The right photograph presentsa close-up view of the segmented needle tip locations.

accurately the needle tip to the RCM is critical to the suc-cess of the robotic system. Waspe et al. previously devel-oped a method to evaluate the RCM calibration of a needlein a robotic system for small animal interventions.14, 15 Themethod involved photographing the needle tip at several an-gles using a high-resolution macro lens. Unfortunately, thiscalibration method is insufficient for our robotic system RCMimplementation. The method by Waspe et al. only accountsfor the location of the needle axis and not the needle tip. Al-though the needle axis may travel very closely to the RCM,the distance of needle tip from the RCM may be quite largedepending upon the needle insertion depth. Thus, the needlemay miss the target even though the reported calibration errorwas quite small. To avoid this problem, the calibration methodof Waspe et al. was modified to account for the needle tipposition.

Calibration of the robot was again completed using a CCDcamera (EOS-1D Mark IV, Canon Canada Inc., Mississauga,ON) and a high-resolution macro lens (MP-E 65 mm f/2.81-5x, Canon Canada Inc., Mississauga, ON). The camera wasaligned to the robot using a carpenters square to align the bar-rel of the lens to the long axis of the robot and vertically usingthe bubble levels of the camera tripod. A 20 mm length of 27gauge drill stock (Model 3009A239, McMaster-Carr, Aurora,OH) sharpened to form a conical tip was mounted onto theneedle driver. To minimize the effects of lens distortion, theneedle was centered in the photograph by moving the robotstages along the X and Y axes. We did not directly estimatethe error arising from lens distortion, but lens distortion was

accounted for through estimation of the photograph pixel size.The needle was advanced a known distance using the robotstages and the total number of pixels traveled in the photo-graph counted. The pixel size was estimated to be the knowndistance traveled using the stage divided by the pixel count.This method of estimation will help to average out the effectsof lens distortion. The pixel size of the images captured usingthis camera and lens was approximately 1.0 m.

The insertion depth of the needle was adjusted using a set-screw on the needle driver. The set-screw was iteratively ad-justed until the tip motion in the roll axis was minimized in thecamera viewfinder. The needle was photographed five timesat approximately equally spaced angular positions along theroll axis over a range of 120. Figure 5 displays a compositeimage of the calibration photographs and the segmented nee-dle tip locations. Once all the photographs had been acquired,the needle tip was segmented in each image using a semiauto-mated algorithm developed in MATLAB (The Mathworks Inc.,Natick, MA). A Sobel edge detector was first applied to theneedle tip images. The identified edge points from the Sobeledge detector, which followed the outer edge of the needletip, were detected based on a user initialization. A linear leastsquares regression was applied to the detected points to de-termine the lines of best fit for each of the two edges of theneedle tip. The angular bisector of the two lines of best fit wasthen calculated (Fig. 6). The needle tip location was finally de-termined by calculating the intersection of the bisector withthe detected points of the needle edge. This process was re-peated for each photograph to yield a set of tip positions. Thecalibration error was then calculated by determining the to-tal range of movement of the needle tip over the full range ofmotion.

Unlike the method introduced by Waspe et al., imagesof the needle tip photographs were acquired in only a sin-gle plane perpendicular to the systems roll axis. Photographswere not acquired in the plane perpendicular to the pitch axisto avoid redundancy. As discussed in Sec. II.A.2, the sys-tem was found to maintain a stable RCM with a deviation


RCM drift, results from the roll and pitch directions shouldbe similar, as the needle tip would be expected to track thesurface of a sphere. The center of the sphere will be locatedat RCM and the radius represents the calibration error in theneedle tip insertion depth. In this case, to characterize the nee-dle tip calibration photographs from only one plane should besufficient. The roll axis was selected for the calibration evalu-ation since it provides approximately twice the range of mo-tion of the pitch axis.

II.B.3. Fixture based needle tip calibrationA calibration fixture was introduced to simplify the nee-

dle tip calibration process. The fixture consists of a Delrinplastic block with a flatness verified to within 25.4 m andmounts directly onto the front of the system (Fig. 7). The fix-ture was mounted to the robot using two 6.35 aluminum shafts(Fig. 7). One of the shafts was machined to have an approxi-mate 200 m eccentricity in its diameter. By rotating the alu-minum shaft, the top surface of the calibration fixture can beadjusted upward and downward to match the RCM location.The correct height of the block was determined by mount-ing the fixture to the robot and advancing the needle driver tothe fully inserted position. A needle was then placed into theloosened needle holder of the driver. The needle was slowlylowered in the holder until its tip was at the surface of the Del-rin block. The needle tip was identified as being at the blocksurface by lightly sliding a 25.4 m steel shim back and forthwhile lowering the needle. The needle tip caught the shim andno longer allowed it to slide freely when in contact. The steelshim provided a means for the user to feel when the needle isin contact with the shim by sliding it back and forth and alsoprotected the Delrin block from being damaged by the nee-dle tip. Without the steel shim, over repeated calibrations theneedle tip may damage the surface of the Delrin block intro-ducing imperfections to the surface flatness. Such imperfec-tions would diminish the ability of the block to calibrate theneedle tip depth to m accuracy. The needle holder was thentightened to fix the inserted needle tip to correspond with theblock surface. The calibration block was then detached from

Delrin Block Shim

Aluminum Shaft

Needle

FIG. 7. Photograph of the calibration fixture used to set the needle tip posi-tion at the RCM when the needle driver is in its forward position.

the robot and the needle tip observed using the viewfinderof the camera and macro lens. This process underwent sev-eral iterations of adjusting the calibration fixture until needlemotion was minimized in the camera viewfinder. With needlemotion minimized, the eccentric shaft was fixed with a set-screw to maintain the appropriate calibration fixture position.The calibration fixture can then be mounted to the robot andused to calibrate the needle tip as needed.

The calibration fixture possesses the advantage of reduc-ing the time required to complete needle calibration. Opticalcalibration of the needle with a camera requires an iterativeprocess of adjusting the needle insertion depth every time theneedle is calibrated. Iterative calibration is time consumingand requires approximately 30 min to complete. On the otherhand, use of the calibration fixture requires only one iterativecalibration of the fixture itself. Once the fixture itself is cal-ibrated, it can be mounted to the robot and used to calibratethe needle tip in less than 5 min. The results of the calibrationusing the fixture were validated using the camera and macrolens. The needle was again imaged at 5 approximately equallyspaced positions in a plane perpendicular to the roll axis. Theneedle tips were then localized in each of the photographsusing the previously described technique in Sec. II.B.2. Thecalibration error was calculated by determining the total rangeof movement of the needle tip in each plane. The calibrationresults of the calibration fixture were compared to the itera-tive optical calibration results to determine the difference inthe accuracy of the two methods.

II.B.4. Needle tip repeatabilityThe repeatability of the needle driver was characterized to

ensure the stability of the needle tip calibration. The needletip will quickly become uncalibrated if the needle driver doesnot consistently and reproducibly place the needle tip to thecorrect depth. To characterize the needle driver repeatability,the needle orientation was adjusted to be fully upright usingthe spherical linkage encoders. The needle was then insertedand retracted a total of nine times while maintaining a con-stant needle orientation. The error in needle depth was inde-pendently characterized using the optical calibration methoddescribed in Sec. II.B.2. The repeatability was characterizedby determining the standard deviation of the needle tip posi-tion in the photographs.

II.C. Robot to micro-CT robot registrationII.C.1. Registration process

A two-stage registration process was developed to regis-ter the coordinate system of the robot to the micro-CT scan-ner (eXplore Ultra Locus, General Electric Healthcare Bio-sciences London, ON, Canada). The two-stage registrationwas developed to achieve a balance between attaining a highquality registration with the time requirements of the enduser to complete a preclinical intervention. Completion ofthe primary first step of the registration requires that a re-movable 6.35 mm borosilicate fiducial bead (McMaster-Carr,



Cleveland, OH)be mounted onto the device at the RCM. Inaddition, an array of six more borosilicate bead secondaryfiducials was mounted onto the robot below the animal bed(fiducial array visible in Fig. 4). With the RCM fiducial beadattached, the robot was positioned at four different locationswithin the micro-CT bore and a CT image was obtained eachtime. Using registration software developed in MATLAB (TheMathworks Inc., Natick, MA), the rigid body transformationbetween the two sets of coordinates was calculated by com-paring the position of the RCM fiducial in robot coordinates toits position in each of the four images. This primary registra-tion can be used alone to guide the robot to targets in micro-CT images. However, if the robotic system is removed andthen reattached to the micro-CT bed, this primary registrationis no longer valid due to variability in robot reattachment.14Unfortunately, repeating the primary registration each timethe robot is reattached to the micro-CT to complete an in-tervention would be time consuming and laborious. To avoidconstant repetition of the primary registration a secondaryregistration was developed.

The secondary registration takes advantage of the sixborosilicate secondary fiducials attached to the robot belowthe animal bed (Fig. 4). During primary registration, thesesix secondary fiducials are imaged along with the RCM fidu-cial bead. One of the scans acquired for the primary regis-tration is of the robot at its home position. To complete thesecondary registration, the reattached robot and six secondaryfiducials are imaged with the robot at its home position. Theregistration software is then used to calculate the rigid bodytransformation using the secondary fiducial positions in theprimary registration home scan and the secondary fiducialpositions in the secondary registration home scan. This sec-ondary registration will account for and correct the variabilityintroduced when the robot is reattached to the micro-CT scan-ner. A target in the micro-CT images can then be localized inthe robot coordinates by applying both the primary and sec-ondary registrations. Through the secondary registration, theend user is only required to acquire one image at the initiationof an intervention rather than four. Furthermore, imaging ofthe fiducials for secondary registration can be simultaneouslyacquired while imaging the small animal to localize targets.Since the small animal must always be imaged, the secondaryregistration does not increase the total number of scans re-quired and allows for the primary registration to be reusedacross multiple interventions. Unfortunately, use of the sec-ondary registration will also reduce overall targeting accuracysince two registration errors, from the primary and secondary,will be combined. The primary registration can be used by it-self to improve targeting accuracy; however, this will be at theexpense of increasing total registration time. Use of the com-bined registration reduces time for procedures at the expenseof accuracy. The end-user must determine which registrationprocess will best suit their application needs.

II.C.2. Registration softwareThe fiducial bead mounted onto the robot RCM was seg-

mented and its centroid was determined in each of the four im-

ages acquired for the primary registration. Segmentation wasaccomplished through a threshold based region-growing. Thecenter of the segmented RCM fiducial was then calculated us-ing a squared-intensity-weighted centroiding. The centroid-ing algorithm used was found in simulated images to havean error of


FIG. 8. View of a reconstructed CT scan used for the needle targetingexperiment.

16 s anatomical scan at 140 kVp and 20 mAs. The imagewas reconstructed to yield an image with 153.9 m isotropicvoxel size (example shown in Fig. 8). The image of the phan-tom was then visualized within MATLAB. Ten image voxelcoordinates within the phantom were manually localized andselected as target positions. For each target set, the location ofeach of the 10 target voxel coordinates within robot coordi-nates was calculated using either a combined registration or aprimary only registration. For each target, the robot RCM wasplaced at the target location, the needle inserted, and an im-age acquired. The angle of the needle was set perpendicularto the mouse platform using a toolmakers square. The needlewas next inclined approximately 60 to the long axis of therobot as a means to minimize artefacts created from the nee-dle by reducing the total material along x-ray paths. Needleangulation was maintained constant at this angle throughoutthe targeting experiments to allow needle alignment accuracyto be quantified independently of angle.

For each acquired image, the distance of the needle fromtarget voxel coordinates was determined by first segmentingthe inserted needle using a threshold-based region growingalgorithm. The centerline of the needle in the image was es-timated using principal components analysis (PCA) to fit a3D line to the segmented needle. The needle alignment accu-

racy, which refers to the accuracy of alignment of the needleinsertion axis with the target in the image was quantified bycalculating the distance of each 3D fitted line to the targetvoxel. The distance of the line to the target voxel representsthe accumulation of errors from a number of sources includ-ing: robot positioning error, registration error, and needle de-flection. However, the line distance to the target voxel doesnot account for errors in needle depth along the needle pathand its associated variability. The error in needle depth mustbe independently characterized using an optical method dur-ing robot calibration as described in Sec. II.B.

II.D.3. Needle angulation accuracyWe used a gelatin phantom to evaluate the variation in

needle positioning over the full range of needle angulation.The gelatin phantom was placed onto the animal bed of therobot and the robot oriented in seven different positions (seeTable II) covering the angular range of the robot motion. Theangles in the table represent the extent of the range of mo-tion of the robot before colliding with the platform support-ing the phantom (see Fig. 4). At each orientation, the nee-dle was inserted into the gelatin phantom and imaged by themicro-CT scanner. In each image, the needle centerline wascalculated using the same technique previously described inSec. II.D.2. Ideally, all seven centerlines should intersect ata common point in space (i.e., the RCM) if no variation inneedle positioning is present with angulation. An iterative so-lution was used to calculate the point in space with the small-est sum of squares distance from each of the centerlines. Aniterative approach was selected to solve the problem since itcould be quickly implemented and since the calculation wasonly performed once to characterize the robot performance.The variation in needle positioning over the full range of nee-dle angulation was then quantified by calculating the distanceof each centerline to the point of best fit.

II.E. Preclinical application

All in vivo imaging was performed under a protocol ap-proved by the University Health Network Animal Care andUse Committee. Measurements where performed in two fe-male SCID mice, each bearing a single subcutaneous humancervix carcinoma tumor (ME180). A tumor was established inthe first mouse by suturing a 23 mm3 tumor fragment alongwith a 1.5 mm radio-opaque pellet (Beekley Co., Bristol, CT)into the dorsal subcutaneous tissue. A tumor was establishedin the second mouse by suturing a 23 mm3 tumor fragmentinto the subcutaneous tissue of the hind limb. The experimentwas performed once the tumors reached approximately 1 cmin diameter. The mice were anesthetized using a 2% by vol-ume isoflurane-oxygen mixture, the hair removed from theirtumors areas, and immobilized in a supine position with theirfront and back paws taped to the custom built mouse platformmounted on the robot. The mice were imaged using a standardanatomical imaging protocol consisting of a 16 s scan with an80 kVp tube voltage and 60 mAs tube current. The imageswere reconstructed to yield a 153.9 m isotropic voxel size.



FIG. 9. Photograph of the experimental setup used for the animal interventions (left) outside, and (right) inside of the bore of the CT scanner.

The experimental setup used in each of the small animal ex-periments is shown in Fig. 9.

The mouse bearing the dorsal tumor with the implantedradio-opaque pellet was used to evaluate the robots abilityto perform image guided needle placement in vivo. Similarto in vitro experiments, a 20 mm length 23 G needle wasmounted on the needle driver. The tumor was immobilized bytaping it onto a plastic block and surrounding it with piecesof rigid foam. A preneedle insertion CT scan was performedto visualize the radio-opaque pellet. The robot RCM wasplaced at the center of mass (CM) of the radio-opaque pel-let and the needle was inserted. A postneedle insertion scanwas acquired to confirm successful contact with the target.This process was repeated for three angles of insertion chosenrandomly.

Interstitial fluid pressure (IFP) measurements were per-formed in the mouse bearing the hind limb subcutaneoustumor using the wick-in-needle technique.30 Measurementswere made using a 23 G needle containing both a front andside port connected to a pressure transducer (Model P23XL,Harvard Apparatus, Canada), which in turn was connectedto a data acquisition system (PowerLab 4/35 with LabChartPro, ADInstruments Pty Ltd., USA) through 50 cm of PE20polyethylene tubing (Becton Dickinson, Franklin Lakes, NJ,USA). The entire system was flushed with a heparin sul-phate/saline solution (1:10). A preneedle insertion scan wasperformed, the tumor identified, and a position manually se-lected near the boundary between the normal tissue and thetumor. The needle was inserted to the target and an unsta-ble measurement was observed. We then retracted the nee-dle by approximately 3 mm to obtain a more stable mea-surement. We added this to illustrate the value of perform-ing the experiment inside the CT bore, which allowed us tocorrect the needle position based on image feedback and ob-tain a more stable measurement. This highlights the impor-tance of using an accurate positioning system under imageguidance to perform reliable IFP measurements. IFP mea-surements were made as the needle was inserted to the firstposition, retracted to the second position, and then main-tained for 30 s after reaching the second target location. Post-needle injection scans were performed at both positions toconfirm the location of the needle inside the tumor at bothpositions.

III. RESULTSTable I below is a summary of the results obtained in each

of the experiments previously described in the methods. Thesection number of each experiment in Sec. II is providedin the table. Figure 10 provides a bar graph to help visu-alize the contributions from the error sources and their as-sociated variability. The largest contribution to the error isfrom the registration of the robot to the micro-CT and thelargest contribution to the total variability is from the nee-dle angulation experiment where a single target is approachedfrom multiple angles. Table II provides a summary of the re-sults of the needle angulation accuracy experiments.

III.A. Preclinical application

Postneedle insertion CT projection images showed that theneedle was successfully delivered to the target (Fig. 11). Mea-surement of the distance between the needle tip and the sur-face of the pellet was difficult due to partial volume effectsand beam hardening artifacts caused by the radio-opaque pel-let and steel needle. A real time cine-CT scan of the last nee-dle placement showed that the shift in pellet position was dueto contact with the needle a well as tissue deformation duringneedle insertion.

Robot guided IFP measurements were made at two loca-tions in the subcutaneous hind limb tumor. In the first loca-tion, the postneedle insertion CT scan demonstrated that thefront port of the needle was straddling the boundary betweentumor and healthy tissue, and the side port was located inthe center of the tumor [Fig. 12(a)]. The needle was thenretracted 3 mm and a CT scan showed that both the frontand side ports of the needle were inside the tumor volume[Fig. 12(b)]. The accuracy and stability of the wick-in-needletechnique requires that both ports of the needle are ex-posed to a similar external pressure. If the pressure at oneport is substantially lower than the other, fluid flow is di-rected out of the IFP system and a decreasing pressure ismeasured. Region III of Fig. 12(c) demonstrates this effect.When the needle was retracted to the second position shownin Fig. 12(b), the IFP measurement stabilized [Fig. 12(b),region IV].



TABLE I. Summary of the results obtained for each experiment.

Error (m) Test

II.B.2 Optical needle tip calibration error to RCM OpticalRoll planex 43y 28II.B.3. Fixture based needle tip calibration error to RCM OpticalRoll planex 36y 70II.B.4. Needle driver repeatability Optical needle 9II.C. Robot registration Mechanical comparator and micro-CTPrimaryFREprimary 21 6TREprimary 31 12SecondaryFREsecondary 70 25TREsecondary 79 14CombinedTREcombined 139 63II.D.2. Robot needle alignment accuracy Micro-CTPrimary registration alignment accuracyTotalerror 131 25Combined registration alignment accuracyTotalerror 206 20II.D.3. Needle angulation accuracy Micro-CT angle (


TABLE II. Summary of needle angulations used to test needle deflection atdifferent angles of attack. The variables and represent the angle of theprimary and secondary link (L1 and L2, Fig. 1) measured by the encoders.The angle of attack represents the angle between the needle axis and thenormal vector to the surface of the phantom.

Scan Angle of attack (deg) Error (m)

1 45 135 11 342 25 155 14 513 25 115 21 244 66 155 21 1105 5 114 30 346 60 178 25 567 105 172 50 189

IV. DISCUSSIONIV.A. Robot calibration

The calibration error is reported in Table I by measur-ing the range of pixels traveled by the needle tip in eachset of photographs. The needle tip is expected to track thesurface of a sphere and as a result the x error should beapproximately twice the y error as the needle tip tracks acircular path in the photographs with a diameter of x andradius of y. However, this is not completely true for theobtained results and can be attributed to several error sources.First, the use of photography to validate the calibration re-sults in a three-dimensional path being projected onto a two-dimensional plane causing the circular path to distort. Errorin positioning the camera to be truly perpendicular to each ofthe rotational axes will also result in the needle paths failingto track a circular path. Finally, due to machining tolerances,the needle will not be perfectly straight. Bending in the nee-dle will result in the needle tip shifting as the system is movedthroughout its full range of motion. Depending upon the di-rection of needle bending, the error in either x or y may beexaggerated. This is evident in the results of the fixture basedcalibration where y was found to be much larger than x.

Needle

Pellet

Tumor

FIG. 11. Projection views obtained from two sequential CT scans thatdemonstrate the ability to perform image-guided needle placement in vivo.(Left) A preneedle insertion image highlighting the location of the needle,the tumor (red outline), and the radio-opaque pellet. (Right) A postneedleinsertion image showing the needle making contact with the radio-opaquepellet.

IV.B. Robot registrationAs expected, the TRE of the secondary registration was

larger than the primary registration by a factor of approx-imately 2.5. In part, this discrepancy can be attributed tothe secondary registration containing twice the segmentationand centroid errors of the primary registration. Image voxelsize plays an important role in registration and the use ofsmaller voxel sizes should yield lower registration errors. Forcomparison, Waspe et al.14 is the only other methoddeveloped to register a robotic system with a micro-CT imag-ing system. The reported FRE and TRE of the registrationprocess were 96 m and 210 m, respectively. Both the pri-mary and combined registrations offer a reduction in errorover this method.

IV.C. Robot positioning accuracyThe RCM deviation is known to be a maximum of

12.5 m through the robots full range of motion. In compar-ison, the mean error from the point of best fit for the robotangulation testing was 72 62 m. Therefore, the varia-tion in angular targeting accuracy is largely not the result ofmechanical errors. Rather, this error would largely be due toneedle deflection in the tissue-mimicking phantom. The Pear-son product-moment correlation coefficient for needle angleof attack and error was calculated from the data in Table IIand found to be r = 0.77. This coefficient indicates that nee-dle deflection increases with larger oblique angles of attack.These results are not surprising: the larger the angle of at-tack, the more obliquely the needle penetrated the phantomsurface and the greater the proportion of the needle within thephantom. Both of these conditions are conducive to needledeflection. The large degree of variability in the results canin part be attributed to needle axis misalignment arising froma bent needle. All needles will be slightly bent prior to in-sertion as the result of manufacturing tolerances. Pre-existingneedle bending will result in different errors between resultswith an equal angle of attack but with different roll and pitchangles. Depending upon the roll and pitch angle, the direc-tion of the needle bending will shift relative to the phantomand introduce variation to the interaction of the phantom andneedle tip. This behavior can be observed in scans 3 and 4(Table II).

A source of error typically neglected in the literature is therearward deflection of the robot itself during needle insertion.This error is neither accounted for in the robot needle align-ment error nor in the calibration error. The robot deflectionis the result of the entire machine shifting due to reactionaryforces acting on the needle driver as it advances the needlethrough the tissue. To determine the magnitude of this deflec-tion in our design, a brass weight, which exerted 10 N of forcewas mounted to the robot RCM. The brass weights were sus-pended from the end of the spherical linkage while it was fullyextended and the attached tooling ball displacement was mea-sured using a dial indicator as a comparator. A force of 10 Nwas chosen instead of 2 N because the resultant deflectionmay have been too small for the indicator to measure. The



0 10 20 30 40 50 60 70-5

0

5

10

15

20

25

30

35

Time (s)

Pres

sure

(mm

Hg)

I II III IV

(a) (b)

(c)

Tumor Boundary

Side Port

Front Port

Tumor Boundary

Side Port

Front Port

FIG. 12. Wick-in-needle measurements of IFP demonstrating the importance of needle placement for stable and accurate results. (a) The front port of the IFPneedle is straddling the tumor boundary (outline), while the side port is in the center. (b) Both the front and side ports are within the tumor boundary (outline).(c) Results of IFP measurements showing (I) the preneedle insertion baseline (II) the signal as the needle is inserted; (III) the measured IFP at the position shownin (a); and (IV) the measured IFP at the position shown in (b).

deflection of the robot with the brass weight was measuredusing an indicator to be 151 m. However, the needle driveris capable of delivering a maximum force of 2 N, which corre-sponds with a rearward robot deflection of 30 m. This illus-trates the need to make the robot as rigid as possible. Althoughour system is suitable for inserting needles into soft tissue, therigidity of this device would need to be improved for applica-tions like drilling into harder materials such as bone.

The needle alignment error is defined as the shortest dis-tance of the needle axis to the target voxel in a tissue-mimicking phantom. Unfortunately, this metric does not pro-vide any information about the error in needle depth or itsassociated variability. Furthermore, the reported errors weremeasured at a constant angle and do not account for variationin the needle alignment accuracy due to needle angulation.The reported needle alignment errors therefore underestimatethe true targeting error. The needle alignment errors can becorrected to better represent the true targeting error by us-ing the errors measured during needle calibration and needleangulation testing, which do account for these other factors.

Since these errors are all independent, their means and stan-dard deviations can be added in quadrature to estimate the truetargeting error.

Combining the measured needle alignment error with thetip calibration error and angulation error, the resultant target-ing error for both the primary and combined registration tech-niques would be 142 41 m and 213 38 m, respectively.Including the presence of a 2 N axial load, the targeting er-ror would be 149 41 m and 218 38 m, respectively.These estimates of targeting error are better representative ofthe true targeting error of the robotic system. Even with theinclusion of additional error sources, the targeting accuracyof the robot is approximately equal to the imaging voxel sizeof 153.9 m. This targeting accuracy makes the robot poten-tially useful for targeting small vessels with a high degree ofconfidence. The secondary registration can be used to reducethe time requirements of interventions. However, use of thesecondary registration resulted in a poorer targeting accuracy,which was greater than the image voxel size. This methodwould be useful for targeting larger structures like the left or



right ventricle of a mouses heart or a large tumor. Since thevariability in targeting is relatively low in comparison to themean error, the targeting accuracy could be further improvedto achieve finer targeting accuracies using micro-CT scannerswith smaller voxel sizes. One approach to improve the tar-geting accuracy is introduced by Ramrath et al.20 to measurethe magnitude and direction of needle misalignment using ahigh-resolution camera. An appropriate correction can thenbe applied when positioning the needle to reduce error fromneedle misalignment.

IV.D. Preclinical application

Under image guidance, the robot was able to success-fully target a 23 G needle to a 1.5 mm radio-opaque pelletimplanted in a subcutaneous tumor. Tissue deformation wasobserved during the initial penetration and retraction of theneedle, and could potentially result in missing the intendedin vivo target. While the effect of tissue deformation was neg-ligible in our ability to target a 1.5 mm radio-opaque pellet,it likely worsens with smaller targets and with proximity ofthe target to the skin (where the observed tissue deformationwas the largest). Using real-time image guidance it may bepossible to reduce, if not eliminate, the effect of tissue defor-mation. To accomplish this, the needle driver would have toadvance or retract the needle tip beyond the preset target. Thisapproach can accommodate for target motion if the target ismoving along the path of the advancing needle axis. However,the kinematic error present in the mechanical linkage wouldadd to the targeting error as the needle tip moves farther thanthe RCM and mechanical set point.

The wick-in-needle technique requires proper placementof the needle for reliable IFP measurements in small tumors.Both the front and side ports of the IFP needle must be insidethe tumor volume, which becomes difficult in small animaltumors with diameters between 5 and 10 mm. The averagedistance between the front and side port of our IFP needlewas approximately 5 mm. Therefore, a great deal of uncer-tainty in manually placing the IFP needle in mouse tumorssmaller than 10 mm is expected. For example, we have foundthat performing repeated manual needle placement in an intra-muscular ME180 tumor 7 mm in diameter results in IFP val-ues that differ by a factor of 5. In this study, we have shownthat the robotic position system in combination with imageguidance provides an accurate method to guide needle place-ment, and reliably perform IFP measurements. Additionally,the design of the robot allows for spatial mapping of IFP overthe tumor volume and is an application we plan to explore inthe future.

V. CONCLUSIONThe design of a micro-CT guided needle positioning sys-

tem for small animal intervention has been presented. Thesystem has been developed with the objective of achievinga mean targeting error of


review of the literature, Int. J. Oral Maxillofac Surg. 41(11), 13111324(2012).

2K. B. Stitzenberg, Y. Wong, M. E. Nielsen, B. L. Egleston, and R. G. Uzzo,Trends in radical prostatectomy: Centralization, robotics, and access tourologic cancer care, Cancer, 118.1, 5462 (2011).

3T. de la Cueva, A. Naranjo, E. de la Cueva, and D. Rubio, Refinement ofintrathymic injection in mice, Lab Anim. 36(5), 2732 (2007).

4M. El-Ghamari, F. Bergmann, B. M. Schmied, J. Weitz, and A. Ulrich,Islet cells contribute to pancreatic carcinogenesis in an animal model,Pancreas 40(2), 242246 (2011).

5M. L. Springer, R. E. Sievers, M. N. Viswanathan, M. S. Yee, E. Fos-ter, W. Grossman, and T. Yeghiazarians, Closed-chest cell injections intomouse myocardium guided by high-resolution echocardiography, Am. J.Physiol. Heart Circ. Physiol. 289(3), H1307H1314 (2005).

6Q. Li, B. Li, X. Wang, A. Leri, K. P. Jana, Y. Liu, J. Kajstura, R. Baserga,and P. Anversa, Overexpression of insulin-like growth factor-1 in miceprotects from myocyte death after infarction, attenuating ventricular dila-tion, wall stress, and cardiac hypertrophy, J. Clin. Invest. 100(8), 19911999 (1997).

7R. Blair-Handon, K. Mueller, and S. Hoogstraten-Miller, An alternativemethod for intrathymic injections in mice, Lab Anim. 39(8), 248252(2010).

8V. Horsley and R. H. Clarke, The structure and function of the cerebellumexamined by a new method, Brain 31(1), 45124 (1908).

9G. C. Kagadis, G. Loudos, K. Katsanos, S. G. Langer, and G. C. Niki-foridis, In vivo small animal imaging: current status and future prospects,Med. Phys. 37(12), 64216442 (2010).

10S. J. Schambach, S. Bag, L. Schilling, C. Groden, and M. A. Brockmann,Application of micro-CT in small animal imaging, Methods 50(1), 213(2010).

11S. H. Bartling, W. Stiller, W. Semmler, and F. Kiessing, Small animalcomputed tomography Imaging, Curr. Med. Imaging Rev. 3(1), 4559(2007).

12C. T. Badea, M. Drangova, D. W. Holdsworth, and G. A. Johnson,In vivo small-animal imaging using micro-CT and digital subtractionangiography, Phys. Med. Biol. 53(19), R319R350 (2008).

13S. A. Nicolau, L. Mendoza-Burgos, L. Soler, D. Mutter, and J. Mare-saux, In vivo evaluation of a guidance system for computer assistedrobotized needle insertion devoted to small animals, Proc. Med. ImagingAugmented Reality 5128, 241250 (2008).

14A. C. Waspe, D. D. McErlain, V. Pitelka, D. W. Holdsworth, J. C. Lacefield,and A. Fenster, Integration and evaluation of a needle-positioning robotwith volumetric microcomputed tomography image guidance for small an-imal stereotactic interventions, Med. Phys. 37(4), 16471659 (2010).

15A. C. Waspe, H. J. Cakiroglu, J. C. Lacefield, and A. Fenster, Design,calibration and evaluation of a robotic needle-positioning system for smallanimal imaging applications, Phys. Med. Biol. 52(7), 18631878 (2007).

16P. Kazanzides, J. Chang, I. Iordachita, J. Li, C. Clifton Ling, andG. Fichtinger, Development of an image-guided robot for small animalresearch, Comput. Aided Surg. 12(6), 357365 (2007).

17J. Chang, W. Bixiu, P. Kazanzides, P. Zanzonico, R. D. Finn, G. Fitchtinger,and C. Clifton Ling, A robotic system for 18F-FMISO PET-guided intra-tumoral pO2 measurements, Med. Phys. 36(11), 53015309 (2009).

18M. J. Hwang, O. Bebek, F. Liang, B. Fei, and M. Cenk Cavusoglu, Kine-matic calibration of a parallel robot for small animal biopsies, 2009 IEEE-RSJ International Conference on Intelligent Robots and Systems (IEEE,2009), pp. 41044109.

19O. Bebek, M. J. Hwang, and M. Cenk Cavusoglu, Design of a ParallelRobot for Needle Based Interventions on Small Animals, IEEE/ASMETrans. Mechatron. 18(1), 6273 (2013).

20L. Ramrath, U. G. Hofmann, and A. Schweikard, A robotic assistant forstereotactic neurosurgery on small animals, Int. J. Med. Rob. Comput.Assist. Surg. 4(4), 295303 (2008).

21K. C. Graham, L. A. Wirtzfeld, L. T. MacKenzie, C. O. Postenka,A. C. Groom, I. C. MacDonald, A. Fenster, J. C. Lacefield, and A. F. Cham-bers, Three-dimensional high-frequency ultrasound imaging for longitu-dinal evaluation of liver metastases in preclinical models, Cancer Res.65(12), 52315237 (2005).

22J. Bax, D. Cool, L. Gardi, K. Knight, D. Smith, J. Montreuil, S. Shere-brin, C. Romagnoli, and A. Fenster, Mechanically assisted 3D ultrasoundguided prostate biopsy system, Med. Phys. 35(12), 53975410 (2008).

23K. Ma, A. Kornecki, J. Bax, Y. Mundt, and A. Fenster, Development andvalidation of a new guidance device for lateral approach stereotactic breastbiopsy, Med. Phys. 36(6), 21182129 (2009).

24R. H. Taylor, J. Funda, B. Eldridge, K. Gruben, D. LaRose, S. Gomory,M. Talamini, L. A. Kavoussi, and J. H. Anderson, A telerobotic assis-tant for laparoscopic surgery, IEEE Eng. Med. Biol. Mag. 14(3), 279288(1995).

25R. H. Taylor and D. Stoianovici, Medical robotics in computer-integratedsurgery, IEEE Trans. Rob. Autom. 19(5), 765781 (2003).

26A. Patwardhan, Subpixel position measurement using 1D, 2D and 3D cen-troid algorithms with emphasis on applications in confocal microscopy, J.Microscopy-oxford 186, 246257 (1997).

27K. S. Arun, T. S. Huang, and S. D. Blostein, Least-squares fitting of two3D point sets, IEEE Trans. Pattern Anal. Mach. Intell. 9(5), 698700(1987).

28J. M. Fitzpatrick, J. B. West, and C. R. Maurer, Jr., Predicting error inrigid-body point-based registration, IEEE Trans. Med. Imaging 17(5),694702 (1998).

29L. K. Ryan and F. S. Foster, Tissue equivalent vessel phantoms for in-travascular ultrasound, Ultrasound Med. Biol. 23(2), 261273 (1997).

30H. O. Fadnes, R. K. Reed, and K. Aukland, Interstitial fluid pressure in ratsmeasured with a modified wick technique, Microvasc. Res. 14(1), 2736(1977).


3D Image-guided Robotic Needle Positioning System for Small Animal Interventions

Documents

Transcript of 3D Image-guided Robotic Needle Positioning System for Small Animal Interventions