research article methodology for registration of shrinkage ...downloads.hindawi.com › journals ›...

Research ArticleMethodology for Registration of Shrinkage Tumors inHead-and-Neck CT Studies

Jianhua Wang1 Jianrong Dai2 Yongjie Jing3 Yanan Huo4 and Tianye Niu5

1Department of Radiation Oncology Ningbo Treatment Center Ningbo Lihuili Hospital Ningbo Zhejiang 315000 China2Department of RadiationOncology Cancer Institute (Hospital) Chinese Academy ofMedical Sciences Peking UnionMedical CollegeBeijing 100021 China3Key Laboratory of Particle and Radiation Imaging of Chinese Ministry of Education Department of Engineering PhysicsTsinghua University Beijing 100084 China4The Second Affiliated Hospital of Zhejiang University School of Medicine Hangzhou Zhejiang 310009 China5Sir Run Run Shaw Hospital Institute of Translational Medicine Zhejiang University Hangzhou Zhejiang 310016 China

Correspondence should be addressed to Tianye Niu tyniuzjueducn

Received 31 October 2014 Revised 15 January 2015 Accepted 28 January 2015

Academic Editor Panagiotis Mavroidis

Copyright copy 2015 Jianhua Wang et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Tumor shrinkage occurs in many patients undergoing radiotherapy for head-and-neck (HampN) cancer However one-to-onecorrespondence is not always available between voxels of two image sets This makes intensity-based deformable registrationdifficult and inaccurate In this paper we describe a novel method to increase the performance of the registration in presenceof tumor shrinkage The method combines an image modification procedure and a fast symmetric Demons algorithm to registerCT images acquired at planning and posttreatment fractions The image modification procedure modifies the image intensities ofthe primary tumor by calculating tumor cell survival rate using the linear quadratic (LQ) model according to the dose delivered tothe tumor A scale operation is used to deal with uncertainties in biological parameters The method was tested in 10 patients withnasopharyngeal cancer (NPC) Registration accuracy was improved compared with that achieved using the symmetric DemonsalgorithmThe averageDice similarity coefficient (DSC) increased by 21This novelmethod is suitable forHampNadaptive radiationtherapy

1 Introduction

Image registration is becoming a key tool inmodern radiationtherapy In image-guided radiation therapy (IGRT) CTimages acquired before each treatment are registered withplanning CT images to verify patient position In adaptiveradiation therapy similar registrations are needed to segmentstructures automatically and to evaluate the dose receivedin each fraction [1ndash3] Rigid registrations are sufficient insituations where only rigid movements occur Howeverdeformable registrations are required in caseswhere structuredeformation andor tumor shrinkage occur

The majority of patients with head-and-neck (HampN)cancer who undergo fractionated radiation therapy experi-ence significant anatomic changes such as tumors shrinkagechanges in overall body habitus and weight loss [4ndash6]

Tumor shrinkage is mainly due to the death and destructionof cells killed by radiation The loss of these cells causestumor density and tumor volume to decrease This meansthat physical one-to-one correspondence may no longer existbetween voxels of the two image sets This problem severelyimpairs the performance of an intensity-based deformableregistration algorithm The ldquoDemonsrdquo family of algorithmsis such a registration form which can accurately account foranatomical changes at relatively low computational expense[7ndash10] and is a suitable choice for registration ofHampN images[11]

The gas pocket mismatching problem in abdominalimages also provides challenges for the intensity-based de-formable image registration Several methods have been usedto overcome the problems associatedwith lack of correspond-ence and mismatched objects that occur with deformable

Hindawi Publishing CorporationComputational and Mathematical Methods in MedicineVolume 2015 Article ID 265497 9 pageshttpdxdoiorg1011552015265497

2 Computational and Mathematical Methods in Medicine

registration of abdominal CT images [12ndash16] In imagesof the pelvic regions the issue of no correspondence isassociated with the presence of bowel gas Because the gas isnot clinically important (the contents of the bowel neednot be registered) the artificial gas [12 13] or constantintensity masks [14 15] were introduced to create ldquovirtualrdquocorrespondence This improved the registration accuracy forthe surrounding tissues such as the rectal wall or prostatewhich were clinically important organs However in thecase of tumor shrinkage the tumor itself is the treatmenttarget We could not use the constant intensity masks (orartificial intensity pattern) in the tumor where the noncorre-spondence problem occurred For this reason the approachmentioned above is infeasible To our knowledge the issue oflack of correspondence induced by the radiation therapy hasnot been adequately described in the literature

In this paper we describe a novelmethod for dealing withthe effects of radiation on tumor tissues in the HampN duringdeformable registration The method involved modificationof image intensities of the primary tumor by calculatingtumor cell survival rate using the linear quadratic (LQ)modelaccording to radiation dose delivered to the tumor [17]

2 Materials and Methods

21 Data Extraction The study included 10 patients withnasopharyngeal cancer (NPC) who underwent intensity-modulated radiation therapy (IMRT) at Cancer Institute(Hospital) of Chinese Academy of Medical Sciences ChinaPatients were immobilized by custom-made thermoplasticmasks Two CT image sets were acquired for each patient(before the start of treatment and at the completion of thetreatment course) using a Philips Brilliance Big Bore 16-sliceCT scanner The image sets were composed of 3mm thickslices the matrix size of each slice was 512 times 512 and thepixel size was approximately 1mmThe initial image sets wereloaded into a Pinnacle version 82g system (Philips MedicalSystems Cleveland OH) for treatment planning To improvethe registration speed the number of slices for each imagepair was reduced so that it just covered the whole primarytumor According to our standard treatment protocol theprimary tumorwas prescribed 70ndash726GyThe tumor volumewas contoured by a single physician for all planning andposttreatment CTs Contouring was undertaken to evaluatethe registration results rather than to help the deformationprocess

The CT images contours and treatment plan doses foreach patient were exported from Pinnacle using the DICOMRTprotocolThe exported dose array was a 3D dosemapThedose voxel size was 4 times 4 times 4mm3

The study was approved by the local ethical committeeand informed consent was obtained from all patients

22 Image Preparation The posttreatment CT was used asthe static image and the planning CT as the moving imagefor each image pair This represented a worst-case scenariofor tumor shrinkage to test our proposed algorithm Thecouch region was manually selected in the transverse planeand the voxels CT numbers were set to that of air (ie

minus1000HU) Voxels CT numbers below the empirically deter-mined minus500HU were also set to minus1000HU This was doneto reduce the interference from nonuniformities outside thepatient

Both the planning CT images and the posttreatment CTimages were cropped on the transverse plane to restrict theregions of interest (ROI) to the headThe images were resam-pled and determined to have the same voxel dimensions of2 times 2 times 3mm3 using the nearest interpolation

23 Rigid Registration Rigid registration was performed inadvance to improve the speed and accuracy of the deformableregistration A minimum threshold of around 500HU wasapplied to the two images so only the bony structures of theskeleton remainedThe rigid registration determined a trans-lation that minimized the correlation coefficient betweenvoxels in the two images To evaluate the accuracy of therigid registration we created a set of simulated translationsof patient CT images Compared to the introduced (known)shifts the mean residual shifts (error) were 11mm withinthe voxel size of the patient images This rigid alignmentprovided the basic initialization for subsequent deformableregistration

24 Deformable Registration The ldquoDemonsrdquo algorithm is animage intensity-based deformable registrationmethod that iswidely used in medical practice as it demands relatively lowcomputational expense A variant of this algorithm proposedby Wang was used in the present study [18]

In the original Demons algorithm the displacement foreach voxel is obtained using the spatial gradient of statictarget image intensities The displacement field is then reg-ularized using a Gaussian smoothing filter to suppress noiseand preserve the geometric continuity of the moving imageThis iterative process alternates between calculation of thedisplacement field and regularization By introducing anldquoactive forcerdquo Wang et al modified the standard Demonsalgorithm to obtain faster convergence and improve registra-tion performance The updated deformation field 119863

119899for the

current iteration was written as follows

119863119899= 119866120590lowast(119863

119899minus1+ (1198681015840minus 119868)

sdot (

997888

nabla119878

100381610038161003816100381610038161003816

997888

nabla119868

100381610038161003816100381610038161003816

2

+ 1198962(119868 minus 119868

1015840)2

+

997888

nabla119898

100381610038161003816100381610038161003816

997888

nabla1198681015840

100381610038161003816100381610038161003816

2

+ 1198962(119868 minus 119868

1015840)2

))

(1)

where 119866120590is the Gaussian kernel lowast denotes the convolution

operator and the width of the Gaussian kernel 120590 was fixedto 1 119863

119899minus1is the deformation field at iteration 119899 minus 1 1198681015840 is the

intensity of the moving image and 119868 is the intensity of the

Computational and Mathematical Methods in Medicine 3

Post-RT CT

Resampling and rigid registration

Intensity modification

Demons registration

Fixed imageMoving image

Fixed image

Rescaled image

Moving image

Planning CT

Output deformable field

Planningdose array

Resampling

Figure 1 Method workflow consisting of the following proce-dures rigid registration intensity modification intensity scale anddeformable registration

static image the respective gradient images are denoted by997888

nabla1198681015840 and 997888nabla

119868 119896 is a normalization factor and we used 119896 = 04

25 Intensity-Modification Procedure (IMP) In patientsundergoing radiation therapy including HampN cancer itis important that target volumes are accurately registeredDue to the cells killed by radiation the density and even thevolume of primary tumor may decrease Unless this one-to-one correspondence is addressed a large registration errorwill occur We thus proposed to change the image intensitiesfor the target volume in the planning CT image (movingimage) based on the linear quadratic (LQ) model Theintensity modification was only done for the planning tumorand all other voxels of planning CT remained unchanged

The workflow of the proposed method is shown in Figure1 We implemented all procedures described here using in-house program written in Matlab (version 71 MathWorks)software Based on the planning dose array we modified theimage intensities within the primary tumor using the LQmodel The details of this procedure are described below

251 Data Import The primary tumor volume contours asmanually delineated during the routine treatment processwere loaded into the planning CT images The dose arraywas resampled to have the same spacing as the planning CTThe dose array and the planning CT were aligned accordingto their positions in the DICOM patient coordinate systemThus each element in the dose array represented the dose tobe delivered to the corresponding voxel in the planning CTimages

252 Image Intensities Modification We calculated voxelintensity within the primary tumor using the LQ model Forevery slice in the planning CT images 119868

0represented the

intensity of a voxel within the primary tumor By definitionthe relationship between 119868

0(CT numbers expressed in

Hounsfield units) and the corresponding linear coefficient 1205830

was calculated as

1198680=

1000 (1205830minus 120583119908)

120583119908

(2)

Rearranging this gives

1205830=

1198680120583119882

1000

+ 120583119882 (3)

where 120583119882

is the attenuation coefficient of water (approxi-mately 01928minus1)

After the primary tumor cells in a given voxel wereirradiated at dose 119863 the intensity value would decease to 119868

119878

(HU) with the corresponding attenuation coefficient 120583119878 119868119878

was converted to 120583119878by the equation

120583119878=

119868119878120583119882

1000

+ 120583119882 (4)

As shown in (5) 120583119878is proportional to the number of

primary tumor cells 119873119878in a given voxel that survives

irradiation and 1205830is proportional to the number of cells in

that voxel prior to irradiation1198730

120583119878=

1205830119873119878

1198730

= 1205830SF (5)

where according to the LQ model the survival fraction SF isgiven by

SF = exp(minus120572119863(1 + 119889

120572120573

)) (6)

In this situation cell proliferation is negligible 119889 and 119863are the dose per fraction and the total dose to the voxelrespectively 120572 and 120573 are the LQ parameters For the purposesof illustration we assumed that 120572 was 033Gyminus1 and 120572120573was 10Gy [19ndash22] Note that the result is not sensitive to thevariation of 120572 and 120572120573 values as demonstrated in the resultsection Using (3) (4) (5) and (6) we obtained

119868119878= [(

1198680120583119882

1000

+ 120583119882) exp(minus120572119863(1 + 119889

120572120573

)) minus 120583119882] times

1000

120583119882

(7)

253 Voxel Intensity Scale In the planning CT image 1198780

was defined as the sum of voxel values within the primarytumor and 119878 denoted the sum of the voxel values within thecorresponding region in the posttreatment CT image Theintensity value (119868

119878) of the primary tumor voxel was scaled as

follows

1198681015840=

119868119878119878

1198780

(8)

Equations (7) and (8) were used to modify the intensity ofevery voxel of primary tumor in the planning CT imagesFigure 2 shows an axial slice of a planning image before andafter image modification This procedure was coupled withWangrsquos ldquoDemonsrdquo algorithm to calculate the displacementfield

26 Evaluation To quantitatively evaluate the performanceof the proposedmethod theDice similarity coefficient (DSC)was calculated for the tumor For two segmentations on the


(a) (b) (c)

Figure 2 Example of intensity-modification procedure (a) An axial slice of a planning image (b)The same image after intensitymodificationThis modified planning image can be accurately registered using deformable image registration (c) The corresponding slice from theposttreatment CT image

target images given by the deformed contours using thecomputed motion fields and manual contours respectivelytheir corresponding volumeswere denoted by119881

119889and119881119898The

DSC was then defined as [23]

DSC =2119881119889cap 119881119898

119881119889+ 119881119898

times 100 (9)

DSC ranged between 0 and 100 A DSC value of 0 indicatedtwo completely uncorrelated images and aDSC value of 100indicated a perfect match

To assess the effectiveness of the proposed intensity mod-ified procedure (IMP) we compared results of deformableimage registration with and without intensity modificationwhen all other parameters were set to the same We alsocalculated the DSC for the tumor performed only with therigid registration between these 10 pairs of CT images TheWilcoxon signed-rank testwas used to compare eachmethod

3 Results

31 Registration Example To compare the effectiveness ofthe intensity modification we performed deformable imageregistration in 10 pairs of head CT images with and withoutthe intensity-modification procedure An example is shownin Figure 3 The target volume change for this case wasfound to be 33 (from 1559 cc to 1047 cc) The left row ofFigure 3 shows the planning CT in axial and sagittal viewstogether with manually delineated tumor volume (greencontours) The right row shows the posttreatment CT Bothsets of contours were overlaid in these images the deformedcontours without IMP (shown in blue) and the deformedcontours with IMP (shown in red) As it can be seen in thefigure the deformed contours without using IMP clearly didnot match with the reduced tumor target due to the lack ofone-to-one correspondence

Figure 4 shows the dense displacement fields with andwithout IMPThe arrows indicate the displacement in 3D but

Table 1 Statistics for Dice similarity coefficient (DSC) for the tumorusing rigid registration and deformable image registration with andwithout the intensity-modification procedure (IMP)

Patient DSC (rigid) DSC (IMP) DSC (no IMP)1 867 916 8802 789 951 8233 735 909 6074 869 968 8625 525 782 6346 715 894 5397 768 940 8258 793 958 7519 804 952 87510 760 933 807Mean 763 920 760

are projected as 2D images for display purposes A vectorfield was used to assess the result and detect errors It isevident from these displays that the displacement vectors(within the smaller region around the primary tumor wherenoticeable shrinkage occurs) are more chaotic discontin-uous and abrupt when IMP is not applied (Figure 4(a))The voxels in this region exhibit unrealistic displacementdue to the lack of correspondence This is improved whenthe deformable registration is embedded with the intensity-modification procedure (Figure 4(b))

32 Registration Statistics The DSC values calculated for thetumor using rigid and deformable registrationwith andwith-out IMP are summarized in Table 1 The Wilcoxon signed-rank test showed that there was little difference between therigid registration and the deformable registration methodwithout using the IMP method (119875 = 0721) Neverthelessthere was a significant improvement when the IMP method


Planning CT

(a)

Posttreatment CT

(b)

Figure 3 The CT images of one patient with the registered contours overlaid (a) Planning CT in sagittal and axial slices and the manualcontours overlaid (b)The corresponding slices of posttreatment CTwith the deformed contours overlaidThe red and blue contours representthe segmentation results with (red) and without (blue) the IMP

(a) (b)

Figure 4 Displacement field calculated using the deformable registration algorithm overlaid on an axial slice Displacement field (a)computed without using the intensity-modification procedure (IMP) and (b) computed using the IMP

was used over the rigid registration (119875 lt 0005) orthe deformable registration method without using the IMPmethod (119875 lt 0005) The images with rigid registration hada mean overlap value of DSC = 763 The mean value ofthe calculated DSC was 760 for deformable registration

without IMPThe application of the IMP leads to amean valueof DSC = 920 On average the improvement in DSC was21 with IMP in these 10 cases

In this study hypothesesweremade on the values of120572 and120572120573 ratioThe uncertainty in radiosensitivity parametersmay

6 Computational and Mathematical Methods in MedicineD

ice s

imila

rity

coeffi

cien

t (D

SC)

050 3 5 10 15 20 3025 35 40

05506

06507

07508

08509

1095

120572120573 (Gy)

120572 = 01Gyminus1 120572 = 10Gyminus1120572 = 033Gyminus1

Figure 5 Sensitivity of the Dice similarity coefficient (DSC) withrespect to 120572 and 120572120573 for patient 2

affect the reliability of the outcome Therefore a sensitivityanalysis was performed to quantify the uncertainty of DSCwhen different parameter combinations were applied in thestudy A typical example is shown in Figure 5We adjusted thekey parameters (120572 and 120572120573) in a large range of possible valuesand provided the DSC results It indicated that our method isnot sensitive to the radiosensitivity parameters For exampleDSC slightly increased from 936 to 953 even with morethan one order of magnitude change on 120572 and 120572120573 values (120572from 01 Gyminus1 to 1 Gyminus1 and 120572120573 from 3Gy to 40Gy)

The computing time to perform the deformable registra-tion varied depending on the number of registered imageslices For a typical planning and posttreatment CT pair therun time for the full registration was approximately 15minThese results were obtained on a Dell desktop PC with dual213 GHz Intel Xeon processors and 225GB of RAM

As our focus was the accuracy of registration we madeno special effort to improve the speed of the registration pro-cess However time saving could be achieved by improvingcomputer hardware such as GPU-based implementation ofthe Demons algorithm [24 25]

4 Discussion

Previous studies have shown that nonrigid anatomicalchanges can occur in HampN during the course of fraction-ated radiation therapy [26] The Demons algorithm hasbeen shown to be a good choice for registration of HampNimages [11] However the issue of primary tumor densitychangevolume shrinkage due to radiation treatment has notbeen addressed by the authors In our experience registrationaccuracy using this algorithm is reduced under these circum-stances

In the current study we used a modification to theDemons algorithm which adjusted the intensities of theprimary tumor voxels according to the LQmodel Qualitativeand quantitative results demonstrated that the proposedmethod increased the performance of the registration inpresence of tumor shrinkage

The intensity-modification procedure used in this studyis essentially a preprocessing method that can be used inconjunction with other intensity-based deformable registra-tion algorithms While the focus of the current work was toevaluate deformable registration in HampN region the methodmay also be adapted to other anatomical regions wheretumor shrinking occurs (eg in lung cancer) In lung cancernoncoplanar beams may be used and more slices should beincluded along both the superior and inferior directions tothe tumor

The LQ model is generally effective in describing tumorresponse to radiation and is widely used in experimental andclinical radiobiology So we chose this model to calculatethe NPC cell killing by radiation The cell killing process isa complicated biological process However this preliminarymodel could serve as a basis for more complex models todeal with the problem of tumor shrinkage Considering thenonhomogeneous dose distributions in IMRT we assumedthat the primary tumor volume was composed of a seriesof subvolumes (voxels) each one receiving a homogeneousdose In this setting the basic LQ model may be used toestimate the NPC cell killing by radiation We also recognizethat further studies would be desirable to validate thisassumption

The radiation therapy parameters values used in ourstudy were derived from the literature [19ndash22] The ratio120572120573 was assumed to be 10Gy which is a nominal valuefor most tumors and 120572 was set at 033Gyminus1 Radiobiologyparameters have a high degree of uncertainty caused byinterpatient variations tumor heterogeneity and effects ofhypoxia and chemotherapy To deal with this uncertaintywe used a scale procedure within the LQ model to fix thepotential problems caused by the patient and tumor specificbiological parameters The scale adjustments were based oninformation from pretreatment CT images The sensitivityanalysis showed that our method was not sensitive to theradiosensitivity parameters It may also be possible to obtainbiological information for individual patients noninvasivelyusing functional imaging [27 28]

In the formulation for the deformation field we havechosen 120590 to be 1 and 119896 to be 04 based on previous studiesdone by other investigators [7 8 11 14 18] and our own initialtesting experience The impacts of the 120590 and 119896 parameterswere not investigated thoroughly in this preliminary studyHowever these specific values work well in most of ourcases Adjusting parameters by a trial-and-error methodwas also tested No significant improvement was observedcompared to the fixed values In addition our emphasis ison the comparison of the various schemes and not the finalperformance Therefore we used the same set of parametersfor all experiments without multiresolution adjustments

Validation of deformable image registration remains adifficult task due to the lack of the ground truth Ourdeformable registration algorithm was validated by simulat-ing the deformation on patient CT image We applied a 2nd-order polynomial transformation to the original HampN image(Figure 6(a)) and deformed its shape by more than 5mm onaverage (Figure 6(b)) Our algorithm automatically generatedthe deformation field and deformed the original image to


(a) (b)

(c) (d)

(e)

Figure 6 (a) Original image (b) original image transformed by a ldquopolynomialrdquo transformation (c) deformed image derived from (a) usingthe Demons algorithm (d) difference image of (b) and (a) (e) difference image of (b) and (c)


the mathematically transformed one (Figure 6(c)) The dif-ference images (Figures 6(d) and 6(e)) were respectivelyFigures 6(b)-6(a) and Figures 6(b)-6(c) Figure 6(e) showedthat the original CT image registeredwith themathematicallydeformed CT image with little difference Quantitative vali-dation results showed that more than 90 of the voxels werewithin 2mm of their intended shifts Future work includesa further improvement to the performance of IMP methodby implanting fiducial markers into the patient for furthervalidation

5 Conclusion

We developed and tested a novel method for performingdeformable registration between planning and posttreatmentCT images in the HampN region The technique was ableto account for tumor responses to radiation therapy bymodifying image intensities of the primary tumor voxelsaccording to the LQ model and to deal with the interpatientheterogeneity of radiobiology parameters by a scaling factorThe preliminary tests resulted in higher registration accuracythan current methods indicating a role in HampN adaptiveradiation therapy

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

Thiswork is supported by theNational Natural Science Foun-dation of China (NSFC Grants nos 81201091 and 81200662)the Fundamental Research Funds for the Central Universitiesin China the Fund Project for Excellent Abroad ScholarPersonnel in Science andTechnology and theNatural ScienceFoundation of Zhejiang Province (Grant no LY12H12010)

References

[1] D Sarrut ldquoDeformable registration for image-guided radiationtherapyrdquo Zeitschrift fur Medizinische Physik vol 16 no 4 pp285ndash297 2006

[2] J Orban de Xivry P Castadot G Janssens et al ldquoEvaluationof the radiobiological impact of anatomic modifications duringradiation therapy for head and neck cancer can we simplysummate the doserdquo Radiotherapy and Oncology vol 96 no 1pp 131ndash138 2010

[3] P Castadot X Geets J A Lee and V Gregoire ldquoAdaptive func-tional image-guided IMRT in pharyngo-laryngeal squamouscell carcinoma is the gain in dose distributionworth the effortrdquoRadiotherapy and Oncology vol 101 no 3 pp 343ndash350 2011

[4] J L Barker Jr A S Garden K K Ang et al ldquoQuantifica-tion of volumetric and geometric changes occurring duringfractionated radiotherapy for head-and-neck cancer using anintegrated CTlinear accelerator systemrdquo International Journalof Radiation Oncology Biology Physics vol 59 no 4 pp 960ndash970 2004

[5] E KHansenM K Bucci JMQuivey VWeinberg and P XialdquoRepeat CT imaging and replanning during the course of IMRT

for head-and-neck cancerrdquo International Journal of RadiationOncology Biology Physics vol 64 no 2 pp 355ndash362 2006

[6] Q Wu Y Chi P Y Chen D J Krauss D Yan and A Mar-tinez ldquoAdaptive replanning strategies accounting for shrinkagein head and neck IMRTrdquo International Journal of RadiationOncology Biology Physics vol 75 no 3 pp 924ndash932 2009

[7] J PThirion ldquoImagematching as a diffusion process an analogywith Maxwellrsquos demonsrdquo Medical Image Analysis vol 2 no 3pp 243ndash260 1998

[8] H Wang L Dong J OrsquoDaniel et al ldquoValidation of an accel-erated lsquodemonsrsquo algorithm for deformable image registration inradiation therapyrdquo Physics in Medicine and Biology vol 50 no12 pp 2887ndash2905 2005

[9] R Varadhan G Karangelis K Krishnan and S Hui ldquoA frame-work for deformable image registration validation in radiother-apy clinical applicationsrdquo Journal of Applied Clinical MedicalPhysics vol 14 no 1 article 4066 2013

[10] N Stanley C Glide-Hurst J Kim et al ldquoUsing patient-specificphantoms to evaluate deformable image registration algorithmsfor adaptive radiation therapyrdquo Journal of Applied ClinicalMedical Physics vol 14 no 6 pp 177ndash194 2013

[11] P Castadot J A Lee A Parraga X Geets B Macq and VGregoire ldquoComparison of 12 deformable registration strategiesin adaptive radiation therapy for the treatment of head and necktumorsrdquo Radiotherapy and Oncology vol 89 no 1 pp 1ndash122008

[12] M Foskey B Davis L Goyal et al ldquoLarge deformation three-dimensional image registration in image-guided radiation ther-apyrdquo Physics in Medicine and Biology vol 50 no 24 pp 5869ndash5892 2005

[13] S Gao L Zhang H Wang et al ldquoA deformable image regis-tration method to handle distended rectums in prostate cancerradiotherapyrdquo Medical Physics vol 33 no 9 pp 3304ndash33122006

[14] A Godley E Ahunbay C Peng and X A Li ldquoAutomatedregistration of large deformations for adaptive radiation therapyof prostate cancerrdquoMedical Physics vol 36 no 4 pp 1433ndash14412009

[15] D Yang S R Chaudhari S M Goddu et al ldquoDeformable reg-istration of abdominal kilovoltage treatment planning CT andtomotherapy daily megavoltage CT for treatment adaptationrdquoMedical Physics vol 36 no 2 pp 329ndash337 2009

[16] B Rodriguez-Vila F Garcia-Vicente and E J Gomez ldquoMethod-ology for registration of distended rectums in pelvicCT studiesrdquoMedical Physics vol 39 no 10 pp 6351ndash6359 2012

[17] J F Fowler ldquoThe linear-quadratic formula and progress infractionated radiotherapyrdquo British Journal of Radiology vol 62no 740 pp 679ndash694 1989

[18] H Wang L Dong M F Lii et al ldquoImplementation and valida-tion of a three-dimensional deformable registration algorithmfor targeted prostate cancer radiotherapyrdquo International Journalof RadiationOncology Biology Physics vol 61 no 3 pp 725ndash7352005

[19] M Avanzo J Stancanello G Franchin et al ldquoCorrelation of ahypoxia based tumor control model with observed local controlrates in nasopharyngeal carcinoma treated with chemoradio-therapyrdquoMedical Physics vol 37 no 4 pp 1533ndash1544 2010

[20] R M Wyatt A H Beddoe and R G Dale ldquoThe effects ofdelays in radiotherapy treatment on tumour controlrdquo Physics inMedicine and Biology vol 48 no 2 pp 139ndash155 2003


[21] S Webb ldquoOptimum parameters in a model for tumour controlprobability including interpatient heterogeneityrdquo Physics inMedicine and Biology vol 39 no 11 pp 1895ndash1914 1994

[22] J F Fowler ldquoIs there an optimumoverall time for head and neckradiotherapy A review with newmodelingrdquoClinical Oncologyvol 19 no 1 pp 8ndash22 2007

[23] L R Dice ldquoMeasures of the amount of ecologic associationbetween speciesrdquo Ecology vol 26 no 3 pp 297ndash302 1945

[24] R Castillo E Castillo R Guerra et al ldquoA framework forevaluation of deformable image registration spatial accuracyusing large landmark point setsrdquo Physics in Medicine andBiology vol 54 no 7 pp 1849ndash1870 2009

[25] G C Sharp N Kandasamy H Singh and M Folkert ldquoGPU-based streaming architectures for fast cone-beam CT imagereconstruction and demons deformable registrationrdquo Physics inMedicine and Biology vol 52 no 19 pp 5771ndash5783 2007

[26] S Nithiananthan K K Brock M J Daly H Chan J C Irishand J H Siewerdsen ldquoDemons deformable registration forCBCT-guided procedures in the head and neck convergenceand accuracyrdquo Medical Physics vol 36 no 10 pp 4755ndash47642009

[27] S M Bentzen and V Gregoire ldquoMolecular imaging-based dosepainting a novel paradigm for radiation therapy prescriptionrdquoSeminars in Radiation Oncology vol 21 no 2 pp 101ndash110 2011

[28] F Z Yetkin and D Mendelsohn ldquoHypoxia imaging in braintumorsrdquo Neuroimaging Clinics of North America vol 12 no 4pp 537ndash552 2002

Submit your manuscripts athttpwwwhindawicom

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


MEDIATORSINFLAMMATION

of


Behavioural Neurology

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary and Alternative Medicine

Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom


registration of abdominal CT images [12ndash16] In imagesof the pelvic regions the issue of no correspondence isassociated with the presence of bowel gas Because the gas isnot clinically important (the contents of the bowel neednot be registered) the artificial gas [12 13] or constantintensity masks [14 15] were introduced to create ldquovirtualrdquocorrespondence This improved the registration accuracy forthe surrounding tissues such as the rectal wall or prostatewhich were clinically important organs However in thecase of tumor shrinkage the tumor itself is the treatmenttarget We could not use the constant intensity masks (orartificial intensity pattern) in the tumor where the noncorre-spondence problem occurred For this reason the approachmentioned above is infeasible To our knowledge the issue oflack of correspondence induced by the radiation therapy hasnot been adequately described in the literature

In this paper we describe a novelmethod for dealing withthe effects of radiation on tumor tissues in the HampN duringdeformable registration The method involved modificationof image intensities of the primary tumor by calculatingtumor cell survival rate using the linear quadratic (LQ)modelaccording to radiation dose delivered to the tumor [17]

2 Materials and Methods

21 Data Extraction The study included 10 patients withnasopharyngeal cancer (NPC) who underwent intensity-modulated radiation therapy (IMRT) at Cancer Institute(Hospital) of Chinese Academy of Medical Sciences ChinaPatients were immobilized by custom-made thermoplasticmasks Two CT image sets were acquired for each patient(before the start of treatment and at the completion of thetreatment course) using a Philips Brilliance Big Bore 16-sliceCT scanner The image sets were composed of 3mm thickslices the matrix size of each slice was 512 times 512 and thepixel size was approximately 1mmThe initial image sets wereloaded into a Pinnacle version 82g system (Philips MedicalSystems Cleveland OH) for treatment planning To improvethe registration speed the number of slices for each imagepair was reduced so that it just covered the whole primarytumor According to our standard treatment protocol theprimary tumorwas prescribed 70ndash726GyThe tumor volumewas contoured by a single physician for all planning andposttreatment CTs Contouring was undertaken to evaluatethe registration results rather than to help the deformationprocess

The CT images contours and treatment plan doses foreach patient were exported from Pinnacle using the DICOMRTprotocolThe exported dose array was a 3D dosemapThedose voxel size was 4 times 4 times 4mm3

The study was approved by the local ethical committeeand informed consent was obtained from all patients

22 Image Preparation The posttreatment CT was used asthe static image and the planning CT as the moving imagefor each image pair This represented a worst-case scenariofor tumor shrinkage to test our proposed algorithm Thecouch region was manually selected in the transverse planeand the voxels CT numbers were set to that of air (ie

minus1000HU) Voxels CT numbers below the empirically deter-mined minus500HU were also set to minus1000HU This was doneto reduce the interference from nonuniformities outside thepatient

Both the planning CT images and the posttreatment CTimages were cropped on the transverse plane to restrict theregions of interest (ROI) to the headThe images were resam-pled and determined to have the same voxel dimensions of2 times 2 times 3mm3 using the nearest interpolation

23 Rigid Registration Rigid registration was performed inadvance to improve the speed and accuracy of the deformableregistration A minimum threshold of around 500HU wasapplied to the two images so only the bony structures of theskeleton remainedThe rigid registration determined a trans-lation that minimized the correlation coefficient betweenvoxels in the two images To evaluate the accuracy of therigid registration we created a set of simulated translationsof patient CT images Compared to the introduced (known)shifts the mean residual shifts (error) were 11mm withinthe voxel size of the patient images This rigid alignmentprovided the basic initialization for subsequent deformableregistration

24 Deformable Registration The ldquoDemonsrdquo algorithm is animage intensity-based deformable registrationmethod that iswidely used in medical practice as it demands relatively lowcomputational expense A variant of this algorithm proposedby Wang was used in the present study [18]

In the original Demons algorithm the displacement foreach voxel is obtained using the spatial gradient of statictarget image intensities The displacement field is then reg-ularized using a Gaussian smoothing filter to suppress noiseand preserve the geometric continuity of the moving imageThis iterative process alternates between calculation of thedisplacement field and regularization By introducing anldquoactive forcerdquo Wang et al modified the standard Demonsalgorithm to obtain faster convergence and improve registra-tion performance The updated deformation field 119863

119899for the

current iteration was written as follows

119863119899= 119866120590lowast(119863

119899minus1+ (1198681015840minus 119868)

sdot (

997888

nabla119878

100381610038161003816100381610038161003816

997888

nabla119868

100381610038161003816100381610038161003816

2

+ 1198962(119868 minus 119868

1015840)2

+

997888

nabla119898

100381610038161003816100381610038161003816

997888

nabla1198681015840

100381610038161003816100381610038161003816

2

+ 1198962(119868 minus 119868

1015840)2

))

(1)

where 119866120590is the Gaussian kernel lowast denotes the convolution

operator and the width of the Gaussian kernel 120590 was fixedto 1 119863

119899minus1is the deformation field at iteration 119899 minus 1 1198681015840 is the

intensity of the moving image and 119868 is the intensity of the


Post-RT CT



Demons registration


Fixed image

Rescaled image

Moving image

Planning CT


Planningdose array

Resampling



nabla1198681015840 and 997888nabla






0represented the




was calculated as

1198680=

1000 (1205830minus 120583119908)

120583119908

(2)


1205830=

1198680120583119882

1000

+ 120583119882 (3)

where 120583119882



119878



120583119878=

119868119878120583119882

1000

+ 120583119882 (4)





120583119878=

1205830119873119878

1198730

= 1205830SF (5)


SF = exp(minus120572119863(1 + 119889

120572120573

)) (6)


119868119878= [(

1198680120583119882

1000

+ 120583119882) exp(minus120572119863(1 + 119889

120572120573

)) minus 120583119882] times

1000

120583119882

(7)




follows

1198681015840=

119868119878119878

1198780

(8)




(a) (b) (c)



119889and119881119898The


DSC =2119881119889cap 119881119898

119881119889+ 119881119898

times 100 (9)



3 Results








Planning CT

(a)

Posttreatment CT

(b)


(a) (b)






ice s

imila

rity

coeffi

cien

t (D

SC)

050 3 5 10 15 20 3025 35 40

05506

06507

07508

08509

1095

120572120573 (Gy)






4 Discussion









(a) (b)

(c) (d)

(e)




5 Conclusion




Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















Post-RT CT



Demons registration


Fixed image

Rescaled image

Moving image

Planning CT


Planningdose array

Resampling



nabla1198681015840 and 997888nabla






0represented the




was calculated as

1198680=

1000 (1205830minus 120583119908)

120583119908

(2)


1205830=

1198680120583119882

1000

+ 120583119882 (3)

where 120583119882



119878



120583119878=

119868119878120583119882

1000

+ 120583119882 (4)





120583119878=

1205830119873119878

1198730

= 1205830SF (5)


SF = exp(minus120572119863(1 + 119889

120572120573

)) (6)


119868119878= [(

1198680120583119882

1000

+ 120583119882) exp(minus120572119863(1 + 119889

120572120573

)) minus 120583119882] times

1000

120583119882

(7)




follows

1198681015840=

119868119878119878

1198780

(8)




(a) (b) (c)



119889and119881119898The


DSC =2119881119889cap 119881119898

119881119889+ 119881119898

times 100 (9)



3 Results








Planning CT

(a)

Posttreatment CT

(b)


(a) (b)






ice s

imila

rity

coeffi

cien

t (D

SC)

050 3 5 10 15 20 3025 35 40

05506

06507

07508

08509

1095

120572120573 (Gy)






4 Discussion









(a) (b)

(c) (d)

(e)




5 Conclusion




Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b) (c)



119889and119881119898The


DSC =2119881119889cap 119881119898

119881119889+ 119881119898

times 100 (9)



3 Results








Planning CT

(a)

Posttreatment CT

(b)


(a) (b)






ice s

imila

rity

coeffi

cien

t (D

SC)

050 3 5 10 15 20 3025 35 40

05506

06507

07508

08509

1095

120572120573 (Gy)






4 Discussion









(a) (b)

(c) (d)

(e)




5 Conclusion




Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















Planning CT

(a)

Posttreatment CT

(b)


(a) (b)






ice s

imila

rity

coeffi

cien

t (D

SC)

050 3 5 10 15 20 3025 35 40

05506

06507

07508

08509

1095

120572120573 (Gy)






4 Discussion









(a) (b)

(c) (d)

(e)




5 Conclusion




Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















ice s

imila

rity

coeffi

cien

t (D

SC)

050 3 5 10 15 20 3025 35 40

05506

06507

07508

08509

1095

120572120573 (Gy)






4 Discussion









(a) (b)

(c) (d)

(e)




5 Conclusion




Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)

(c) (d)

(e)




5 Conclusion




Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















5 Conclusion




Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of






























of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















research article methodology for registration of shrinkage ...downloads.hindawi.com › journals ›...

Documents