Image Registration Accuracy of a 3-Dimensional … anatomic relation to the current transrectal...

Image Registration Accuracy of a 3-Dimensional TransrectalUltrasound–Guided Prostate Biopsy System

Yujun Guo, PhD, Priya N. Werahera, PhD,Ramkrishnan Narayanan, PhD, Lu Li, MS, DineshKumar, MS, E. David Crawford, MD, Jasjit S. Suri, PhD

Objective. For a follow-up prostate biopsy procedure, it is useful to know the previous biopsy locationsin anatomic relation to the current transrectal ultrasound (TRUS) scan. The goal of this study was tovalidate the performance of a 3-dimensional TRUS-guided prostate biopsy system that can accuratelyrelocate previous biopsy sites. Methods. To correlate biopsy locations from a sequence of visits by apatient, the prostate surface data obtained from a previous visit needs to be registered to the follow-up visits. Two interpolation methods, thin-plate spline (TPS) and elastic warping (EW), were tested forregistration of the TRUS prostate image to follow-up scans. We validated our biopsy system using acustom-built phantom. Beads were embedded inside the phantom and were located in each TRUSscan. We recorded the locations of the beads before and after pressures were applied to the phantomand then compared them with computer-estimated positions to measure performance. Results. In ourexperiments, before system processing, the mean target registration error (TRE) ± SD was 6.4 ± 4.5mm (range, 3–13 mm). After registration and TPS interpolation, the TRE was 5.0 ± 1.03 mm (range,2–8 mm). After registration and EW interpolation, the TRE was 2.7 ± 0.99 mm (range, 1–4 mm). Elasticwarping was significantly better than the TPS in most cases (P < .0011). For clinical applications, EWcan be implemented on a graphics processing unit with an execution time of less than 2.5 seconds.Conclusions. Elastic warping interpolation yields more accurate results than the TPS for registration ofTRUS prostate images. Experimental results indicate potential for clinical application of this method.Key words: elastic warping interpolation; phantom validation; prostate cancer; surface-based regis-tration; thin-plate spline interpolation; 3-dimensional transrectal ultrasound.

Received February 2, 2009, from Eigen Inc, GrassValley, California USA (Y.G., R.N., L.L., D.K., J.S.S.);University of Colorado, Denver, Colorado USA(P.N.W.); and University of Colorado HealthSciences Center, Aurora, Colorado USA (D.C.).Revision requested March 23, 2009. Revisedmanuscript accepted for publication May 5, 2009.

Yujun Guo, Ramakrishnan Narayanan, Lu Li,Dinesh Kumar, and Jasjit S. Suri are employees ofEigen Inc, manufacturer of the Artemis system.

Address correspondence to Jasjit S. Suri, PhD (CTO),Eigen Inc, 13366 Grass Valley Ave, Grass Valley, CA95945 USA.

E-mail: [email protected]

AbbreviationsEW, elastic warping; PCa, prostate cancer; PSA,prostate-specific antigen; 3D, 3-dimensional; TPS, thin-plate spline, TRE, target registration error; TRUS, tran-srectal ultrasound

rostate cancer (PCa) is the most common noncu-taneous human malignancy and the secondmost lethal tumor among American men. In2008, an estimated 186,320 men will have a diag-

nosis of PCa; 28,660 of them will die of this disease inthe United States.1 In general, a biopsy is recommendedwhen the patient shows elevated prostate-specific anti-gen (PSA) levels, a possible indicator of underlying malig-nancy. Transrectal ultrasound (TRUS)–guided biopsy isused to remove tissue from the prostate gland for patho-logic classification.2 One of the most perplexing aspects isthat patients with PCa who have similar PSA levels, clini-cal stages, and histopathologic features in their biopsy

© 2009 by the American Institute of Ultrasound in Medicine • J Ultrasound Med 2009; 28:1561–1568 • 0278-4297/09/$3.50

P

Technical Advance

2811jumonline.qxp:Layout 1 10/15/09 1:34 PM 561

tissue can have markedly different clinical out-comes.3 Although localized PCa can becomelethal in some patients,4 most men die with PCarather than of it. Autopsy studies have confirmedhistologically apparent PCa in the prostateglands of approximately 42% of men older than50 years who died of other causes.5 Nevertheless,the 5-year survival rate for American men with adiagnosis of PCa was nearly 100% based onpatients with a diagnosis between 1996 and 2002and followed through 2003.6 Therefore, earlydetection and treatment play important roles inthe clinical management of this disease.

Currently, there are 2 important clinical chal-lenges: (1) diagnosis of clinically threatening can-cer7 and (2) selection of a suitable treatmentregimen.8 Prostate biopsies are subjected to seri-ous sampling errors. The success of prostatebiopsies largely depends on the size and locationof the tumor rather than the clinical importanceof the disease.9 There have been efforts to findoptimal locations for prostate biopsies, but theselection does not guarantee that malignancywill be detected in the first session. If the initialbiopsy results are negative, a second biopsy isusually recommended when PSA levels remainelevated. Studies have shown that up to 10% ofcases with initial negative biopsy results mayproduce positive results during a subsequentbiopsy.10 For a subsequent biopsy, it is useful toknow the previous biopsy locations so that thephysician may plan the current procedure byrevisiting or avoiding some locations.

Because of the relatively long latency of PCa, aconsiderable proportion of men with localizedPCa are subject to overdiagnosis and receiveunnecessary therapy10,11 with attendant morbid-ity, coupled with substantial cost escalationsfrom detection of minor tumors via aggressivescreening.12,13 Because most cases of PCa cur-rently diagnosed by prostate biopsies have anintermediate Gleason score, with either good orpoor clinical outcomes, some of these patientsmay be treated with focal therapy as opposed tomore aggressive treatments, eg, surgery and radi-ation.14 Patients with disease localized to oneside of the prostate can be treated with focal ther-apy, thereby eliminating the usual side effectsassociated with surgery and radiation. The suc-cess of focal therapy largely depends on the

screening procedure (brachytherapy and tem-plate-guided saturation biopsy) and the ability tofocus treatment in relation to specific locationswithin the prostate.15 Therefore, it is necessary toaccurately relocate initial biopsy locations (thosehaving malignancies) during focal treatment.

However, it is very difficult for a physician torelocate the initial biopsy locations obtainedacross the TRUS images during a subsequentbiopsy or therapy session. Although template-guided transperineal biopsy procedures mayprovide some degree of relocation accuracy, cur-rently there is no mechanism to accurately relo-cate biopsy locations performed transrectally.This is partly due to use of a live 2-dimensionalTRUS image, whereas a 3-dimensional (3D) TRUSimage of the gland may provide anatomic featuresthat are more easily discerned. Furthermore, thegland tends to move or deform because of exter-nal physical disturbances, discomfort intro-duced by the procedure, changes due to cancerprogression, therapy, or intrinsic peristalsis. Thequality of the image also depends on the typeand particular settings of the machine. It is,therefore, necessary to find the correspondencebetween TRUS images so that the previous biop-sy sites can be identified on an ultrasound scanduring subsequent visits. Hence, there is a clini-cal need for a 3D TRUS-guided prostate biopsysystem in which initial biopsy locations can beaccurately relocated during follow- up visits.

A 3D TRUS-guided biopsy scheme was initiallydemonstrated.16 It uses a mechanical trackerhaving 4 degrees of freedom. An integral person-al computer–based workstation can registerbiopsy locations in 3D space and accurately relo-cate them in follow-up visits. In this study, weevaluate the accuracy and utility of 2 image inter-polation methods in this 3D TRUS-guidedprostate biopsy system using tissue phantoms.

Materials and Methods

System OverviewThe 3D TRUS-guided biopsy system is parti-tioned into the following subsystems: imageacquisition, prostate segmentation, target plan-ning, tracking, and reporting. Three-dimensionalimage registration completed as a part of targetplanning is a surface-based registration technique.

1562 J Ultrasound Med 2009; 28:1561–1568

Image Registration Accuracy of 3D TRUS-Guided Prostate Biopsy

2811jumonline.qxp:Layout 1 10/15/09 1:34 PM Page 1562

This aligns the segmented surface of the current3D prostate gland with that computed from theprevious visit so that previous biopsy sites canbe interpolated onto the current 3D TRUS vol-ume. Only surface information is used in 3DTRUS image segmentation because ultrasoundimage quality is poor in certain regions withinthe prostate and makes intensity-based registra-tion methods impractical.

Both semiautomated and fully automatedimage segmentation techniques are available inthis system.16 Figure 1 shows the process flow forthe selection of biopsy sites based on a previousbiopsy report. Figure 2 illustrates the essentialcomponents of the surface-based registrationtechnique adapted into our system.17 Figure 3shows a sample graphical user interface designto load a previous biopsy plan onto the currentultrasound scan. Once a patient’s previous visitis selected, the corresponding previously seg-mented prostate surface is registered to the cur-rently segmented surface. After this, the previousbiopsy sites are interpolated on the current vol-ume based on the correspondence establishedthrough registration.

Two interpolation techniques are implement-ed. One is based on a thin-plate spline (TPS)method, whereas the other is based on elasticwarping (EW). The TPS method uses the conceptof minimizing the “bending” energy of a thinsheet of metal.18 In 3D cases, given 2-point setsP and Q, each has n points. Each point (pi, i = 1. . .n) in P corresponds to 1 point (qi, i = 1 . . .n) inQ. Thin-plate spline interpolation is describedby 3(n + 4) parameters, which include 12 globalaffine motion parameters and 3n coefficientsfor correspondences of the n control points(Equation 1):

(1)

The definition of Uij (i = 1 . . .n, j = 1 . . .n) is

(2)

J Ultrasound Med 2009; 28:1561–1568 1563

Guo et al

Figure 1. Process flow diagram for the mapping of biopsy sites from a previous patient visit onto thecurrent visit.

Figure 2. Surface-based registration flow chart.


These parameters are computed by solving thelinear system.18 Elastic warping uses modeling ofelastic sheets, which are warped by an externalforce applied to points (x, y, z) in data set D1, sothat they are deformed to the coordinates of theircorresponding points (f u (x, y, z), f v (x, y, z), f w (x,y, z)) in data set D2.19 Given a set of n correspond-ing points, EW interpolation is used to find thesolution of function (U, V, W ) to the equationsdescribing the deformation of an elastic sheet:

(3)

Where U (X ) = [U V W ]T, X = (x, y, z), and q(X ) =q(x, y, z) is unity when there is a correspondenceand 0 otherwise. Equation 3 is discretized, and theresulting linear system is solved iteratively.

Evaluation ProtocolWe used 5 custom-built tissue phantoms to evalu-ate the accuracy of the 3D TRUS-guided biopsyscheme. A typical tissue phantom design is illus-trated in Figure 4. Spherical beads comprising 2-mm stainless steel balls are planted inside eachtissue phantom at random locations to emulateseveral targeted biopsy sites and can be identifiedon ultrasound scans. As shown in Figure 5, theultrasound signal was locally distorted by the pres-ence of the beads; the beads, however, were iden-tifiable, and the segmentation method wasunaffected. The evaluation procedure is as follows:

1. A 3D TRUS image scan of tissue phantom isacquired.

2. The spherical beads in the ultrasound scanare identified. The centers of spherical beads aresaved and designated as P. Then, either the semi-automated segmentation process or the fullyautomated segmentation process is performed,and the segmented prostate volume is obtainedand designated as a floating 3D image (labeled A).

3. Next, mechanical pressure is applied in anarbitrary direction on the tissue phantom. Thismimics anatomic deformations (shape and sizevariations) or movement of the prostate. Theaforementioned steps are repeated to obtain 3DTRUS images, with the spherical beads centersdesignated Q. The deformed tissue phantom sur-face is segmented and designated as a target 3Dimage (labeled B).

4. The floating 3D image, A, is registered to thetarget 3D image, B, and then a nonrigid defor-mation between the 2 images is obtained.16

Now the 3D image, A, is aligned to the target 3Dimage, B. Therefore, the set of points, P, insidethe floating 3D image, A, is accordinglydeformed and registered as points P′ inside thetarget 3D image, B.

5. The target registration error (TRE) is definedas the mean euclidean distance (D) betweencorresponding locations of the ground truthand the computer estimation. Before registra-tion, the TRE is computed between P and Q.After registration, the TRE is computed betweenP′ and Q, as illustrated in Figure 6. The TRE isused to assess the accuracy of the registrationtechnique.

6. The above protocol is repeated severaltimes, and the TRE is computed. The mean andSD of the TRE are calculated to determinewhether the system meets the required perfor-mance standards for clinical applications.

7. The above steps are repeated the same num-ber of times for the 2 interpolation techniques,TPS and EW, respectively. The mean and SD ofthe TREs are calculated and compared.

Target registration error metrics are calculatedfrom the measurements obtained using the fol-lowing equations. For each set of experiments,the euclidean distance between P and Q is com-puted using Equation 4, where N is the totalnumber of beads in the phantom. The mean ( D )and SD (σ) from all of the experiments were com-puted using Equations 5 and 6, respectively, asfollows:

(4)

(5)

(6)

1564 J Ultrasound Med 2009; 28:1561–1568



Results

Five custom-built tissue phantoms having 3 to 5beads randomly located within each phantomwere used to evaluate the accuracy of the system.For each tissue phantom, 10 to 15 TRUS scanswere obtained using the Falcon 2101 ultrasoundmachine with a type 8667 transducer (5–10 MHz;BK Medical, Herlev, Denmark). The first scan wasperformed without any pressure. Remainingscans were performed with varying mechanicalpressure applied in an arbitrary direction on thetissue phantom. All scans were segmented, andthen the segmented surfaces from the samephantom were registered with a correspondingdeformed phantom in pairs. Figure 7 shows anexample of overlapped surfaces before and afterregistration as well as the locations of the beads(both computer estimated and ground truth).

Altogether, the evaluation procedure wasrepeated 100 times for each phantom, and thenthe TREs for each trial were recorded both beforeand after registration. The same procedure wasperformed for both the TPS interpolation andEW methods. Table 1 lists the TRE (both meanand SD) for all phantom trials. The results for theTPS and EW are shown side by side for compari-son. Before system processing, the TRE was 6.4 ±4.5 mm (range, 3–13 mm). After registration andTPS interpolation, the TRE was 5.0 ± 1.03 mm(range, 2–8 mm). After registration and EW inter-polation, the TRE was 2.7 ± 0.99 mm (range, 1–4mm). Figure 8 illustrates the TREs for trials onphantom 4. In most cases, EW outperformed theTPS method (2-tailed t test, P < .0011).

Discussion

We have presented a 3D prostate biopsy systemthat can be used to map previous biopsy sitesonto a current ultrasound scan. This is veryimportant during a prostate biopsy procedurebecause the urologist may want to either avoidor rebiopsy previous sites.

A custom-built phantom was used to simulatethe prostate deformation between 2 scans.Because the identification from the second scancan serve as the ground truth, it can be used tomeasure the performance of the registration sys-tem. In our experiments, 2 interpolation meth-

ods were implemented, and the results werecompared. Elastic warping outperformed theTPS method in most cases. Both the TPS and EWare approximations of prostate tissue property.However, the TPS is used to solve a given finitenumber of equations, more suitable to a collec-tion of scattered points marking distinct surfacefeatures, whereas EW tends to preserve theshape and relative position of given point setsbecause of its smoothness.19 Because the struc-ture of the segmented surface is fairly consistenton a case-by-case basis, EW is more appropriatefor our system. The EW approach is also morerobust compared with the TPS because the solu-

J Ultrasound Med 2009; 28:1561–1568 1565

Guo et al

Figure 3. Biopsy planning graphical user interface. A, The previous visit to loadcan be selected from the bottom right display quadrant. B, After the selected visitis loaded, corresponding biopsy sites (white spheres) and surfaces (blue-green) aredisplayed.

A

B


tion of the former is solved iteratively, whereasthe TPS approach may construct a singularmatrix and may lead to a singular matrix error.This error may result in a less accurate TRE forthe TPS approach.

Clinical application of either method willdepend on computing time. For a 15-minutebiopsy session, for example, it is desirable to havea registration procedure done in 20 seconds. Thecurrent central processing unit time for the regis-tration procedure on an Intel Core 2 processor(Intel Corporation, Santa Clara, CA) with a 2.66-GHz clock speed is 150 seconds; this is unaccept-able in clinical practice. When the registrationalgorithm was implemented on a graphics pro-cessing unit (an 8800 GT video board running at640 MHz and accessing 512 MB of onboard RAM;NVIDIA Corporation, Santa Clara, CA), the run-ning time was reduced to 12 seconds. For theinterpolation procedure, the TPS method can beperformed onboard the central processing unitin 0.3 second, whereas the EW method requires15 seconds. However, after EW migrated to agraphics processing unit implementation, itstotal time was reduced to 2.5 seconds. This meetsthe clinical requirement.

1566 J Ultrasound Med 2009; 28:1561–1568


Figure 4. Customized phantom design. A, Four views of thedesign. B, beads distributed in the customized phantom.

A

B

Figure 5. Ultrasound scan with the identified bead (red).

Figure 6. Measurement of the registration error. Top, Overlaidscans before registration. Bottom, Overlaid scans after registration.


In addition to the errors introduced by theinterpolation methods, there are other sources oferror contributing to the overall TRE. A beadidentification error is introduced by the operatorwhile locating the beads in the phantom scan;this is due to poor ultrasound scan image qualityand the software tool used. Table 2 lists theexperimental results for the bead identificationerror. Three operators were requested to identifythe beads inside 6 phantom scans, respectively.The locations were recorded and compared. Theoverall bead identification error was 0.8682 ±0.2377 mm.

Segmentation error is also important becausethe segmented prostate volume is always differ-ent from the actual phantom volume. For exam-ple, in our experiments, the customized phantomhas a designed volume of 44 cm3, whereas theaverage segmented volume is 36.77 cm3: a seg-mentation error of 8.07%. Also, the error due tosurface registration adds to the TRE. Image cali-bration and acquisition errors can also con-tribute to errors in the TRE.

J Ultrasound Med 2009; 28:1561–1568 1567

Guo et al

Figure 7. Phantom validation. A, Prostate model surface(white) and simulated biopsy sites (red). B, Deformed model sur-face (green) overlapped onto the original surface (white). Biopsysites from different sources are illustrated in different colors,such as original biopsy sites (red), deformed sites (yellow), andregistered sites (blue).

A

B

Figure 8. Comparison of TREs before and after processingusing different interpolation methods.

Table 2. Bead Identification Error for 6 Phantom Trials

Trial Maximal, mm Minimal, mm Average, mm

1 1.3265 0.2500 0.71122 1.6247 0.7586 1.0073 1.2384 0.4223 0.74574 2.5616 0.2162 1.18025 2.1363 0.3063 0.87586 0.9619 0.2268 0.6886

Each trial had a different number of beads; the maximal, minimal, andaverage identification errors for each trial are listed.

Table 1. Comparison of TREs Before and After System Processing(Registration and Interpolation)

PostprocessingTRE, mm Preprocessing TPS EW

Dµ 6.396 5.060 2.727σ 4.500 1.630 0.995

Dmaxµ 6.915 6.220 3.784σ 4.626 1.840 1.288

Dminµ 5.880 3.787 1.995σ 4.390 1.960 0.857

Target registration errors from 2 interpolation methods (TPS and EW) arealso listed; EW gave better results.


There are other variabilities that may also intro-duce errors due to (1) probe or transducer settings,(2) the scanning method during the ultrasoundscan or navigation, and (3) the changes in themorphologic shape due to medications andother factors. Issue 1 can be compensated by ourhardware interfaces,20 whereas the other 2 issuesrelated to the pressure of the probe or transduceron the prostate gland and the shape deformationcan be mitigated by physician training andmotion compensation procedures.21

In conclusion, we have presented a 3D imageregistration system in which previous patientbiopsy sites can be mapped onto a currentultrasound scan. Our system is based on arobust, surface-based registration algorithm.The transformation method to project biopsysites after registration is fast and accurate. Twointerpolation methods were implemented andcompared, and has been shown that EW per-forms better than the TPS method. The regis-tration system reported here is currently beingintegrated in our Artemis system, a US Food andDrug Administration 510(k)–approved 3DTRUS-guided prostate biopsy system developedby Eigen Inc (Grass Valley, CA). The phantom-based validation results have been published,20

and the product is now at hospital sites for clini-cal validation.

References

1. Jemal A, Siegel R, Ward E, et al. Cancer statistics, 2008. CACancer J Clin 2008; 58:71–96.

2. Hodge KK, McNeal JE, Terris MK, Stamey TA. Random sys-tematic versus directed ultrasound guided transrectal corebiopsies of the prostate. J Urol 1989; 142:71–74.

3. Gann PH, Han M. The natural history of clinically localizedprostate cancer. JAMA 2005; 293:2149–2151.

4. Lerner SE, Blute ML, Bergstralh EJ, Bostwick DG, Eickholt JT,Zincke H. Analysis of risk factors for progression in patientswith pathologically confined prostate cancers after radicalretropubic prostatectomy. J Urol 1996; 156:137–143.

5. Sakr WA, Grignon DJ, Crissman JD, et al. High grade pro-static intraepithelial neoplasia (HGPIN) and prostatic ade-nocarcinoma between the ages of 20–69: an autopsystudy of 249 cases. In Vivo 1994; 8:439–443.

6. Division of Cancer Control and Population Sciences,National Cancer Institute. Surveillance, Epidemiology, andEnd Results Program, 1975–2003. Rockville, MD: NationalCancer Institute; 2006.

7. Kawata N, Miller GJ, Crawford ED, et al. Laterally directedbiopsies detect more clinically threatening prostate cancer:computer simulated results. Prostate 2003; 57:118–128.

8. Damber JE, Aus G. Prostate cancer. Lancet 2008; 371:1710–1721.

9. Crawford ED, Hirano D, Werahera PN, et al. Computermodeling of prostate biopsy: tumor size and location—notclinical significance—determine cancer detection. J Urol1998; 159:1260–1264.

10. Lubeck DP, Grossfeld GD, Carroll PR. A review of measure-ment of patient preferences for treatment outcomes afterprostate cancer. Urology 2002; 60:72–77.

11. Meng MV, Elkin EP, Harlan SR, Mehta SS, Lubeck DP, CarrollPR. Predictors of treatment after initial surveillance in menwith prostate cancer: results from CaPSURE. J Urol 2003;170:2279–2283.

12. Wilson SS, Crawford ED. Screening for prostate cancer:current recommendations. Urol Clin North Am 2004;31:219–226.

13. Cooperberg MR, Moul JW, Carroll PR. The changing faceof prostate cancer. J Clin Oncol 2005; 23:8146–8151.

14. Barqawi A, Crawford ED. Focal therapy in prostate cancer:future trends. BJU Int 2005; 95:273–274.

15. Crawford ED, Wilson SS, Torkko KC, et al. Clinical stagingof prostate cancer: a computer-simulated study oftransperineal prostate biopsy. BJU Int 2005; 96:999–1004.

16. Shen F, Shinohara K, Kumar D, et al. Three-dimensionalsonography with needle tracking: role in diagnosis andtreatment of prostate cancer. J Ultrasound Med 2008;27:895–905.

17. Shen D, Davatzikos C. An adaptive-focus deformablemodel using statistical and geometric information. IEEETrans Pattern Anal Mach Intell 2000; 22:906–913.

18. Bookstein FL. Principal warps: thin-plate splines and thedecomposition of deformations. IEEE Trans Pattern AnalMach Intell 1989; 11:567–585.

19. Davatzikos C, Prince JL, Bryan RN. Image registration basedon boundary mapping. IEEE Trans Med Imaging 1996; 15:112–115.

20. Wei L, Werahera PN, Shinohara K, et al. Accurate andreproducible 3-dimensional transrectal ultrasound–guidedprostate biopsy for diagnosis and treatment of prostatecancer. Grand Rounds Urol 2008; 2:15–19.

21. Shen F, Narayanan R, Suri JS. Rapid motion compensationfor prostate biopsy using GPU. In: Proceedings of the IEEEEngineering in Medicine and Biology Society Conference.Piscataway, NJ: Institute of Electrical and ElectronicsEngineers; 2008:3257–3260.

1568 J Ultrasound Med 2009; 28:1561–1568



Image Registration Accuracy of a 3-Dimensional … anatomic relation to the current transrectal...

Documents

Transcript of Image Registration Accuracy of a 3-Dimensional … anatomic relation to the current transrectal...