11

Head and Neck Lymph Head and Neck Lymph Node Region Delineation Node Region Delineation with Auto-segmentation with Auto-segmentation and Image Registrationand Image Registration

Chia-Chi TengChia-Chi Teng

Department of Electrical EngineeringDepartment of Electrical EngineeringUniversity of WashingtonUniversity of Washington

22

OutlineOutline IntroductionIntroduction Related WorkRelated Work Lymph Node Region Contouring with Lymph Node Region Contouring with

Image RegistrationImage Registration Automatic Segmentation of Landmark Automatic Segmentation of Landmark

StructuresStructures Geometrical Feature Based SimilarityGeometrical Feature Based Similarity ResultsResults ConclusionConclusion

33

ContextContext 3D Conformal Radiotherapy 3D Conformal Radiotherapy (beams (beams

are shaped to match the tumor)are shaped to match the tumor) Intensity Modulated Radiation Intensity Modulated Radiation

Therapy Therapy (controls intensity in small volumes)(controls intensity in small volumes)

44

Target VolumesTarget Volumes GTV / CTV / PTVGTV / CTV / PTV

55

MotivationMotivation Improve the process of target Improve the process of target

volume delineation for volume delineation for radiation therapy planning.radiation therapy planning.

Objective:Objective:– Auto-contour lymph node Auto-contour lymph node

regions.regions.– Initial focus on head and neck.Initial focus on head and neck.

66

ProblemProblem Where are the lymph nodes?Where are the lymph nodes? Where are the lymph node regions?Where are the lymph node regions?

none of thestructures arelymph nodes

77

SolutionSolution Create reference (canonical) models.Create reference (canonical) models. Map reference nodal regions to patients.Map reference nodal regions to patients.

88

System OverviewSystem Overview

Segmentation of Landmark Structures

Target CT Images

Projected Lymph Node Regions

Retrieve Similar Reference Models

3D Volume and Mesh of Mandible, Hyoid, ...

Image Registration

Reference Models with Lymph Node Regions

99

Image RegistrationImage Registration

Align the transformed reference image Align the transformed reference image ffRR ° ° gg to the target image to the target image ffT T . .

Find the optimal set of transformation Find the optimal set of transformation parameters parameters that maximize an image that maximize an image similarity function similarity function SS::

optimaloptimal = argmax = argmax SS(())

1010

Mattes’ MethodMattes’ Method Similarity FunctionSimilarity Function

SS(()) = mutual_information( ffRR ° ° gg , f, fT T ))

Transformation FuncitonTransformation Funcitongg((xx||) = ) = RR((x x - - xxCC) – ) – TT((x x - - xxCC) + ) + DD((xx||))

x x = [= [x, y, zx, y, z]]T T in the reference image in the reference image coordinates.coordinates.

1111

DeformableDeformable TransformationTransformation

Control points Control points (15*15*11).(15*15*11).

Each control point is Each control point is associated with a 3-associated with a 3-element deformation element deformation vector vector , describing , describing x-, y-, z-components x-, y-, z-components of the deformation.of the deformation.

1212

Project Target Lymph Project Target Lymph RegionsRegions Image registration aligns Image registration aligns

reference and target CT sets.reference and target CT sets. Apply result transformation Apply result transformation gg to to

reference lymph node regions.reference lymph node regions. Incorporate anatomical landmark Incorporate anatomical landmark

correspondences.correspondences. Use surface mesh of outer body Use surface mesh of outer body

contour, mandible, hyoid …contour, mandible, hyoid …

1313

Surface WarpingSurface Warping Shelton’s method used to find Shelton’s method used to find correspondences between surfaces.correspondences between surfaces. Energy based surface mesh warping.Energy based surface mesh warping. E(C) = Esim(C) + Estr(C) + Epri(C)

C is the function which maps points from reference surface SSRR to target surface SST T .

1414

Landmark Landmark CorrespondenceCorrespondence

The deformation The deformation at landmark points at landmark pointskk = = kk kk

kk : points from reference surface mesh : points from reference surface mesh SSRR..kk : corresponding locations on : corresponding locations on

transformed reference surface transformed reference surface SSRR ° ° CC matching the target surface mesh matching the target surface mesh SSTT..

1515

Surface SSRR Surface SSTT

SSRR ° ° CC kk = = kk kk

1616

Using Landmark Using Landmark CorrespondenceCorrespondence Deformation vectorsDeformation vectors DD((jj) ) are are

modified according to landmark modified according to landmark correspondencescorrespondences kk in the in the proximity of the control pointsproximity of the control points jj..

Landmark structures align better.Landmark structures align better. Faster convergence.Faster convergence.

1717

Reference Mattes w/ Landmark Target

Compare Image Compare Image Registration ResultsRegistration Results

1818

Reference Mattes w/ Landmark Target

1919

Automatic Automatic Segmentation of Segmentation of Landmark StructuresLandmark Structures Given: Cancer radiation treatment

patient’s head and neck CT image. Find:

– Skull base & thoracic inlet.– Anatomical structures:

cervical spine (white) respiratory tract (dark green) mandible (turquoise) hyoid (yellow) thyroid cartilage internal jugular veins (pink) carotid arteries (dark yellow) sternocleidomastoid muscles (light green,

orange)

2020

MethodMethod 2D knowledge-based segmentation

– Based on Kobashi’s work– Dynamic thresholding– Progressive landmarking

Combined with 3D active contouring– Do not require successful 2D Do not require successful 2D

segmentation on every axial slicesegmentation on every axial slice– Initialize with 2D segmentation result

2121

A B C D

E

2D Segmentation 2D Segmentation ResultsResults

2222

2D/3D Iteration2D/3D Iteration

1 3 5

2 4 6

Identify objects that are easy to find, use them to find harder ones.

2323

2D/3D Iteration – cont.2D/3D Iteration – cont.

7 9 11

8 10 12

2424

Geometrical Feature- Geometrical Feature- Based SimilarityBased Similarity Given: A stored database DB of CT

scans from prototypical reference head and neck cancer patients and a single query CT scan Q from a target patient.

Find: Similarity between Q and each database image d in DB in order to find the most similar database images {ds}.

2525

StructuresStructures Outer body Outer body

contourcontour MandibleMandible HyoidHyoid Internal Internal

jugular jugular veinsveins

2626

Feature TypesFeature Types Simple Simple numeric 3D regional numeric 3D regional

properties: volume and extents. properties: volume and extents. VectorVector properties: relative properties: relative

location between structures.location between structures. ShapeShape properties: surface meshes properties: surface meshes

of structures.of structures.

2727

Features for Similarity Features for Similarity MeasureMeasure

Volume and extents Volume and extents of the overall regionof the overall region Normalized Normalized centroidcentroid of hyoid and of hyoid and

mandiblemandible 3D 3D centroid difference vector centroid difference vector between between

mandible and hyoidmandible and hyoid 2D 2D centroid difference vectorscentroid difference vectors between between

hyoid and jugular veinshyoid and jugular veins Surface meshes Surface meshes of mandible and outer of mandible and outer

body contourbody contour

2828

Mesh Feature Distance Mesh Feature Distance

Register reference mesh Register reference mesh SR and and target mesh target mesh ST with Iterative with Iterative Closest Point (ICP), result Closest Point (ICP), result T..

Hausdorff distance Hausdorff distance between between two aligned surface meshes, TSR and ST

),(max),( TSpTRh STpdSTSdR

The Hausdorff distance is the maximum distance from anypoint in the transformed reference image to the test image.

2929

Feature Vector Feature Vector DistanceDistance Given feature vectors Given feature vectors FFdd and and FFQQ

for model for model dd and query and query QQ in the in the feature vector space feature vector space RRNN..

21

2

1),(),(

iQ

N

iidiiQdF FFdwFFD

3030

EvaluationEvaluation Surface mesh distance after full Surface mesh distance after full

image registration image registration DDHH – slow. – slow. Feature vector distance Feature vector distance DDFF – fast. – fast.

DH

DF

corr_coef(corr_coef(DDHH, , DDFF) ) = 0.72= 0.72

Images with smallfeature vector distanceshould produce the bestresults after registration.

3131

Experiment ResultsExperiment Results 50 head and neck patient CT sets.50 head and neck patient CT sets. 34 subjects are segmented.34 subjects are segmented. 20 subjects with lymph node 20 subjects with lymph node

regions drawn by experts.regions drawn by experts. Image registrationImage registration

20 * (20 – 1) = 380 total cases.20 * (20 – 1) = 380 total cases.

3232

Auto-segmentation Auto-segmentation ResultsResults Correct SegmentationsCorrect Segmentations

3333

Auto-segmentation Auto-segmentation cont.cont. Incorrect SegmentationsIncorrect Segmentations

Carotid artery misidentified as Hyoid partly missing due tojugular vein due to surgery. too low inter-slice resolution.

3434

Successs Failure Incorrect % of successCervical Spine 34 0 0 100.00%Respiratory Tract 34 0 0 100.00%Mandible 34 0 0 100.00%Hyoid 34 0 0 100.00%ThyroidCartilage 33 0 1 97.06%Left Internal Jugular Vein 27 3 4 79.41%

Right Internal Jugular Vein 31 1 2 91.18%Left Carotid Artery 25 9 0 73.53%Right Carotid Artery 30 4 0 88.24%Left SCM 24 10 0 70.59%Right SCM 25 9 0 73.53%

Auto-segmentation Auto-segmentation cont.cont.

3535

Image Registration Image Registration ResultsResults

Total cases Successful Success rate (%)Mattes method 380 367 96.57%

New method using landmark correspondence 380 380 100.00%

Average Standard deviationMattes method 32 minutes 6 minutesNew method using landmark correspondence 26 minutes 5 minutes

Time of Convergence

Success/Failure

3636

Quantitative Quantitative Evaluation -Evaluation - Surface Surface Mesh DistanceMesh Distance

Projected Region SR ° ° ggColor is Color is distance to distance to truth.truth.

Ground Truth: Expert Drawn Target Region ST

DH(SR ° ° gg, ST, n) : Hausdorff distancen : lymph node region

3737

Average Standard deviationMattes method 2.85 1.44New method using landmark correspondence 2.12 0.64

DH(SR ° ° gg, ST, 1B) for all SR, ST.

Measurement in centimeter.

Mattes distance larger thanlandmark distance.

3838

Average Standard deviationMattes method 1.02 0.51New method using landmark correspondence 0.59 0.21

Mean_distance(SR ° ° gg, ST, 1B) for all SR, ST.

Measurement in centimeter.

3939

Similarity EvaluationSimilarity Evaluation RH(i, Q) : the ith ranked

reference subject for target Q based on the image registration results, DH.

RF(i, Q) : the ith ranked reference subject based on geometrical features, DF.P(P(RRFF(1(1, Q, Q)) ==RRHH((11, Q, Q)) = 80%)) = 80%

P(P(RRFF(1(1, Q, Q)) ==RRHH((22, Q, Q)) = 10%)) = 10%

P(P(RRFF(1(1, Q, Q)) ==RRHH((33, Q, Q)) = 4%)) = 4%

Q

di

HausdorffDistance

FeatureDistance

RH(1, Q)

RH(2, Q)

RF(1, Q)

RF(2, Q)

4040

Similarity Evaluation Similarity Evaluation ExamplesExamples

DH DH

DF DF

corr_coef(corr_coef(DDHH, , DDFF) ) = 0.74= 0.74

corr_coef(corr_coef(DDHH, , DDFF) ) = 0.68= 0.68

4141

Similarity Evaluation – Similarity Evaluation – Surface Mesh DistanceSurface Mesh Distance

AveragAveragee

Standard Standard deviatiodeviationn

DDH H for the for the closest closest reference subject reference subject to each target to each target based on feature based on feature distancedistance 1.28 1.28 0.310.31DDH H for all reference for all reference and target subjectsand target subjects 2.592.59 0.900.90Measurement in centimeter. So its better to find the closest subject.

4242

Qualitative Evaluation – Qualitative Evaluation – 1.11.1

Mattes Expert w/ Landmark Drawn

Clinically acceptable target projection.Clinically acceptable target projection.

4343

Mattes Expert w/ Landmark Drawn

Qualitative Evaluation – Qualitative Evaluation – 1.21.2 Clinically acceptable target projection.Clinically acceptable target projection.

4444

Mattes Expert w/ Landmark Drawn

Qualitative Evaluation – Qualitative Evaluation – 1.31.3 Clinically acceptable target projection.Clinically acceptable target projection.

4545

Mattes Expert w/ Landmark Drawn

Qualitative Evaluation – Qualitative Evaluation – 22 Clinically Clinically unacceptableunacceptable target target

projection.projection.

4646

ConclusionConclusion Inter-subject image registration Inter-subject image registration

technique shows promise for lymph technique shows promise for lymph node region auto-contouring.node region auto-contouring.

Knowledge-based auto-segmentation Knowledge-based auto-segmentation is useful for head and neck CT.is useful for head and neck CT.

Fast similar subject search is Fast similar subject search is possible and critical as reference possible and critical as reference database grows.database grows.

4747

Future WorkFuture Work Integrate and evaluated in a clinical Integrate and evaluated in a clinical

environment.environment. Generalize to other types of cancer.Generalize to other types of cancer. Regional lymphatic involvement prediction.Regional lymphatic involvement prediction. Improve image registration results.Improve image registration results. Improve auto-segmentation results.Improve auto-segmentation results.

– Validation logicValidation logic– Knowledge-based 3D active contour constraintsKnowledge-based 3D active contour constraints

4848

AcknowledgementAcknowledgement Linda ShapiroLinda Shapiro Ira KaletIra Kalet Jim BrinkleyJim Brinkley David HaynorDavid Haynor David MattesDavid Mattes Mark WhippleMark Whipple Jerry BarkerJerry Barker Carolyn RutterCarolyn Rutter Rizwan Rizwan NuraniNurani

4949

ContributionsContributions The first auto target contouring tool for The first auto target contouring tool for

radiation therapy. radiation therapy. (AMIA 2002)(AMIA 2002) An auto-segmentation method combining 2D An auto-segmentation method combining 2D

dynamic thresholding and 3D active dynamic thresholding and 3D active contouring. contouring. (IEEE CBMS 2006)(IEEE CBMS 2006)

An image registration method using landmark An image registration method using landmark correspondences in conjunction with mutual correspondences in conjunction with mutual information optimization. information optimization. (IEEE ISBI 2006)(IEEE ISBI 2006)

A patient similarity measurement using 3D A patient similarity measurement using 3D geometrical features of anatomical structures. geometrical features of anatomical structures. (IEEE ISBI 2007)(IEEE ISBI 2007)