th 2008 Tuesday, July 29th th 2008 Tuesday, July 29th 8:30 AM – 9:25 AM 8:30 AM – 9:25 AM By 9:25 AM, you will be able to … Deformable Image Registration: Methods and Endpoints Marc L Kessler, PhD The University of Michigan 9 Describe the basic mechanics (recipes ) of image registration techniques used in commercial & research planning and delivery systems 9 Describe the different techniques used to combine, display & interact with multimodality and 4D image data 9 Understand the clinical use & application of these techniques for treatment planning, treatment delivery & plan adaptation th 2008 Tuesday, July 29th The Recipe 8:30 AM – 9:25 AM Recipe for Image Registration 2 Image Volumes What’s the recipe for registering 1 Transformation Model and fusing 3D & 4D image data? 1 Registration Metric 1 Optimization Engine What’s the cake ? n Evaluation Tools AAPM ~ 50 ~ Recipe for Recipe for Image ImageRegistration Registration 2 Image Volumes Volume 1 1 Transformation Model 1 Registration Metric 1 Optimization Engine n Evaluation Tools Recipe for Recipe for Image ImageRegistration Registration stationary Registration Metric Volume 2’ transformed Volume 2 moving Geometric Transformation # Optimization Engine {ß} Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 1 Recipe for Optimization Recipe for Plan Image Registration Criteria Plan Quality Metric The Cake ! CT # Dose Distribution MR Optimization Engine NM Physics Dose Calculation Model {ß} US The Cake ! Tx Plan 3D Dose Portal Images Generalized Adaptable Patient Model CBCT 3D Dose Day n 4D CBCT 1…n 1…n US The Cake ! PET PET • anatomic • physiologic MR • geometric • density CT MR CT CT+ MR + NM + Dose (τ) (τ) Scalar Data 4-D Vectors Scalar Data 4-D Vectors Pouliot Pouliot // UCSF UCSF The Cake ! The Cake ! Map structure outlines from one image volume to another Map and combine doses from multiply treatment fractions Fraction 1 = Fraction 2 Δ Dose Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 2 Recipe for Recipe for Image ImageRegistration Registration The Cake ! 2 Image Volumes Dose – Function Analysis Chose 2 from the plethora of imaging volumes we now have available in radiation therapy! Simulation Simulation Planning CT SPECT Imaging Gregoire Gregoire // St-Luc St-Luc Available Image Volumes Balter Balter // UM UM Available Image Volumes ? ? Physics X-ray CT Anatomy MRI Physiology Nuc Med Motion Information …multiple modalities …4-D imaging Available Image Volumes Normal Normal Dawson Dawson // PMH PMH Available Image Volumes MR ! Target Target ? Varian OBI™ ? …repeat imaging during therapy Elekta Synergy™ Siemens PRIMATOM™ TomoTherapy Hi-Art™ ViewRay Renaissance™ Resonant Restitu™ …at the treatment unit too! Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 3 Recipe Recipefor forImage ImageRegistration Registration 2 Image Volumes Prostate Example Can you tell what is different in the 2 volumes? Depending on the application, choose either entire volumes or the relevant sub-volumes Volume 1 Prostate Example Bones aligned, prostate region not aligned Entire Volume Volume 2 Prostate Example Bones ignored, prostate region aligned Sub volumes Prostate Example Liver Example MR data discarded ? curved curved flat flat MR CT radioactive seeds anatomic-based cropping ignore anatomy outside liver Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 4 Liver Example The Past / Present Registration Segmentation ignore anatomy outside liver anatomic-based cropping The Present / Future Treatment Delivery Example Registration Segmentation Rigid assumption used for regional registration Pick Your Battles Wisely! Treatment Delivery Example Sonke Sonke // NKI NKI Registration at Delivery Regional registration of serial CBCT scans and TP CT Oops! 6 DOF Choose your battles wisely! Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 5 Recipe Recipefor forImage ImageRegistration Registration Transformation Model 1 Transformation Model X2 X1 Depending on the input data T and clinical situation, chose a transformation model Volume 2 Recipe Recipefor forImage ImageRegistration Registration 1 Transformation Model X1 = Volume 1 Degrees of Freedom { PET/CT MR - CT ß} 4D CT T ( X2 , { ß }) Choose the flavor of T and the set of parameters { ß} Available Flavors of T 0? 3 to 6 3xN None ? Few Many Available Flavors of T Affine Assumption ¾ Rigid / Affine … or regional rigid / affine ¾ Full 3D / 4D Deformation Parametric transformations Free-form transformations y = mx+b … in three dimensions x11 = A x22 + b Otherwise rotation (3) … up to 12 DOF translation (3) scale (3) Spatially variant transformations shear parametric (3) Various & free form models … up to 3 x N ! … lots of degrees of freedom! Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 6 www.gnome.org www.gnome.org Affine Transformations Volume 2 www.gnome.org www.gnome.org Affine Transformations Volume 2 A A Square Square Volume 1 6 DOF A A Square Square Volume 1 Rotation Rotation Translation Translation Scaling Scaling Shearing Shearing Rotation Rotation Translation Translation Scaling Scaling Shearing Shearing 3 3 3 3 3 3 3 3 Parallel lines stay parallel! Parallel lines stay parallel! Munro Munro // Varian Varian Affine Transformations www.gnome.org www.gnome.org Affine Transformations 3 or 4 DOF Rotation Rotation CT Translation Translation Scaling Scaling Shearing Shearing CT CBCT CBCT Parallel lines stay parallel! non- Affine Transformations Volume 1 Volume 2 Parallel lines don’t stay parallel! non- Affine Transformations Volume 1 Volume 2 X1 = T ( X2 , { ß }) Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 7 non- Affine Transformations Study A Study B Available Flavors of T Parametric Models Freeform Models Transformation parameters to apply to a particular voxel depends on the location of the voxel (spatially variant) ! X1 = T ( X2 , { ß(X2) }) Available Flavors of T Spatially variant transformations ¾ Parametric Brock / PMH Warp Space / … Drag Objects Warp Objects / … Drag Space Available Flavors of T Spatially variant transformations ¾ Free Form Thin-plate Splines Intensity Flow Methods Global Deformations Fluid Mechanics-like B-Splines Finite Element Models Local Deformation Account for Physical Properties Thin-Plate Splines Liver Liver Bladder Bladder B-Spline Deformations Each have some distinct properties ¾ B-Splines X2 = X1 + ΔX = X1 + Σ wi·Β(X-ki) weights nn T ( P ) = a00 + a xx x + a yy y + a zz z + ∑ wiiU ( P − Pii ) basis function ( … parameters! ) ii==11 affine warping Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 8 B-Spline Deformations B-Spline Deformations Construct overall function T from a weighted sum of local deformations Each have some distinct properties ¾ B-Splines 1-D Example ΔX w w77 ΔX w w77 knots knots knots kk11 kk22 kk33 kk44 kk55 kk66 kk77 kk88 kk99 kk10 k 11 10 k11 k k 11 k11 k k33 k k55 k k77 k k99 k k10 k22 k k44 k k66 k k88 k 10 k11 X22 = X11 + ΔX = X11 + Σ wii·Β(X-kii) X X22 = X11 + ΔX = X11 + Σ wii·Β(X-kii) parameters! B-Spline Deformations B-Splines Construct overall function T from a weighted sum of local deformations Construct overall function from a weighted sum of local deformations ΔX Deformlets? ΔX w w77 knots w w77 k k 11 k11 k k33 k k55 k k77 k k99 k k10 k22 k k44 k k66 k k88 k 10 k11 X X22 = X11 + ΔX = X11 + Σ wii·Β(X-kii) parameters! B-Spline Transformation Model 3 weights / knots ! knots k k 11 k11 k k33 k k55 k k77 k k99 k k10 k22 k k44 k k66 k k88 k 10 k11 X X22 = X11 + ΔX = X11 + Σ wii·Β(X-kii) This seems a lot like intensity modulation! Multiresolution Deformations Divide and Conquer 60 60 xx 60 60 xx 48 48 mm mm 3-D Grid of Knots X 4 4 xx 4 4 xx 3 3 mm mm Coarse Fine Knot Spacing Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 9 Recipe Recipefor forImage ImageRegistration Registration 1 Registration Metric Registration Metrics ? Choose a registration metric ? depending on the type of imaging data involved Geometry-based Geometry-Based Metrics ¾ Point Matching Least Squares Σ( X 22 Point Matching - X’11 ) 2 Σ( X 22 – X’11 ¾ Surface Matching Chamfer Matching Intensity-based Σ min distance 2 )2 Volume 2 Volume 1 Define corresponding anatomic points Surface Matching a ? b Catallo Catallo // UM UM Sonke Sonke // NKI NKI Registration at Delivery c Regional registration of serial CBCT scans and TP CT dii Σ di 2 objects misaligned compute mismatch mismatch minimized Define corresponding anatomic surfaces 6 DOF Choose your battles wisely! Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 10 Intensity-Based Metrics Sum of Squared Differences … subtract one image from the other ¾ Mono-modality Sum of Squared Diff. ¾ Multimodality Data Mutual Information Σ(I 22 - I11 ) 2 CT 22 p(A,B) p(A,B) p(A,B) log log Σ p(A,B) p(A) p(A) p(B) p(B) - I CT 22 CT 11 I CT 11 = Difference Σ (I CT 1 1 -I CT 22 )22 Sum Sum of of the the Squares Squares of of the the Differences Differences Individual Individual Intensity Intensity Distributions Distributions Sum of Squared Differences Sum of Squared Differences … subtract one image from the other … subtract one image from the other = I CT 22 I CT 11 Individual Individual Intensity Intensity Distributions Distributions Difference - CT I CT Zero MR I MR Individual Individual Intensity Intensity Distributions Distributions Registered Mutual Information = Difference Not Zero This This doesn’t doesn’t usually usually make make much much sense sense Mutual Information … using an information theory-based approach H(I1,I2) = H(I1) + H(I2) - MI(I1,I2) CT MR H(ICT) + H(IMR) = H(ICT,IMR) Individual Information Contents The mutual information is the CT, MR Joint Information Content information that is common to ? both image volumes! ‘48 Shannon - Bell Labs / ‘95 Viola - MIT Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 11 Mutual Information Mutual Information MI(I1,I2) = H(I1) + H(I2) - H(I1,I2) MI(I1,I2) = Σ p(I , I ) log 1 1 2 2 2 p(I11, I22) p(I11) p(I22) The mutual information is the The mutual information of two image information that is common to volumes is a maximum when they are both image volumes! geometrically registered … ‘48 Shannon - Bell Labs / ‘95 Viola - MIT ‘48 Shannon - Bell Labs / ‘95 Viola - MIT Mutual Information Mutual Information original MR How About An Example? Rotate - Translate CT MI = .62 R I I CT 2D joint intensity histogram Transformation p(ICT, IMR) reformatted CT I CT R I reformatted Not so Aligned! MI = .99 M p(ICT, IMR) M Aligned! original MR 2D joint intensity histogram How About An Example? PET PET Registration Metric CT Mutual Information CT Optimizer Simplex Algorithm Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 12 How About An Example? How About An Example? PET PET CT CT Recipe Recipefor forImage ImageRegistration Registration 2 Image Volumes Volume 1 stationary 1 Transformation Model transformed 1 Optimization Engine Volume 2 n Evaluation Tools moving Recipe for Optimization Recipe for Plan Image Registration Plan Quality Metric Dose Distribution Physics Model Dose Calculation Registration Metric Volume 2’ 1 Registration Metric Criteria Recipe Recipefor forImage ImageRegistration Registration Geometric Transformation # Optimization Engine {ß} Interactive Registration # 9 Translate 9 Rotate Optimization Engine ? Deform {ß} Provide tools to transform and visualize! Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 13 Automated Registration Multiresolution Deformations Divide and Conquer 60 60 xx 60 60 xx 48 48 mm mm Rigid first , ... then B-Spline deformation Multiresolution Deformations Coarse Fine Knot Spacing Multiresolution B-Splines Registration Metric vs. Iteration ABC CT Example 2.5 2.5 Registration Metric 4 4 xx 4 4 xx 3 3 mm mm Change Change in in knot knot spacing spacing 2.4 2.4 Low Res 2.3 2.3 2.2 2.2 2.1 2.1 High Res 2.0 2.0 1.9 1.9 Exhale State 1.8 1.8 0 0 20 20 40 40 60 60 80 80 100 100 120 120 140 140 160 160 Inhale State 180 180 Iteration Iteration Number Number Multiresolution B-Splines Multiresolution B-Splines ABC CT Example Multiphasic CT Data Exhale State Inhale State deformed Exhale State Inhale State deformed Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 14 Ruan Ruan // UM UM Multiresolution B-Splines We Are Not Really Splines ! Multiphasic CT Data 9 9 No “stiffness” information Exhale State Inhale State deformed Add Some Physics? Extracted Ribcage Exhale Deform Inhale Spatially Variant Stiffness Etotal = Esimilarity + α Estiffness intensity similarity metric tissue-dependent regularization Evol = ∫ wc(x) |det JT(x) – 1|2 dx Ruan Ruan // UM UM “Stiffness” Weighting Using “Prior” Information wc(x) using “stiffness” information Extracted Ribcage Exhale Deform Inhale Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 15 Balter Balter // UM UM Tissue Sliding Tissue Sliding Balter Balter // UM UM Deal with different organs individually? Deal with different organs individually? Segmentation + Registration Registration / Segmentation No masking Masking Registration Segmentation Ribs driven by large lung deformations Ribs not affected by lung registration Brock Brock // UM UM Finite Element Modeling Brock Brock // PMH PMH Finite Element Modeling Exhale Exhale Inhale Inhale Take into account physical tissue properties (directly) … thorough segmentation is necessary Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 16 The Future ? Are We Already There? Several independent products are there or almost there! Family of Generalized, Customizable, Patient Models II have have no no commercial commercial interest interest in in any any of of these these company company olivier.commowick.org olivier.commowick.org Meyer/UM Meyer/UM …from Atlas to Individual thesis_work/img1.png without without with with 9 Segment /register / average www.mimvista.com www.mimvista.com …from Atlas to Individual II have have no no commercial commercial interest interest in in this this company! company! Segment using atlas Thompson Thompson // UCLA UCLA …from Individuals to Atlas Brain Brain Mapping: Mapping: The The Disorders Disorders,, Academic Academic Press, Press, 1999 1999 Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 17 Recipe for Recipe for Image ImageRegistration Registration Let’s Recap 9 The geometric transformation T Affine / not - Affine Volume 1 stationary Registration Metric Volume 2’ 9 The Registration Metric transformed Geometry / Intensity 9 Optimization Engine Volume 2 Interactive / Automated moving Recipe for Recipe for Image ImageRegistration Registration 2 Image Volumes # Optimization Engine Geometric Transformation {ß} Evaluation Tools How do we know how well these registration methods perform? 1 Transformation Model ¾ build phantoms and test them 1 Registration Metric we can know the truth! 1 Optimization Engine ¾ provide tools to examine results we don’t know the truth! n Evaluation Tools Circa Circa 1986 1986 Multimodality Phantoms 4D Phantoms Kashani Kashani // UM UM CT MR Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 18 Evaluation Tools Visualization Tools Evaluation Tools Numerical Tools ¾ Color gel or wash overlay ¾ Split /dual screen displays Point Description 1 2 3 4 5 6 2nd branch of bronchial tree 3rd branch of bronchial tree 4th branch of bronchial tree Vessel bifurcation 1 Vessel bifurcation 2 Vessel bifurcation 3 X -5.37 -5.73 -6.50 -8.12 -8.06 -10.69 Study Study A A Exhale Y 0.98 2.12 2.77 3.37 -1.95 2.47 * Inhale X Y -4.62 -0.22 A A A A 0.74A A -5.40 -6.24 0.80 -8.12 1.40 -7.67 -3.20 Study Study B 1.16 -10.78B Z -3.42 -5.42 -8.42 -9.92 -4.42 0.58 (x , y , z ) All numbers in centimeters all values in cm. ¾ Anatomic boundary overlay! Exhale' ( w/ TPS alignment ) -4.71 -0.47 -3.36 -5.35 0.58 -5.83 -6.27 0.69 -9.51 -8.19 0.91 -11.60 -7.27 -2.83 -3.63 -10.85 0.87 1.24 σ AAPM Task Group 132 Use of Image Registration and Data Fusion Algorithms and Techniques in Radiotherapy ΔX -0.09 0.05 B B -0.03 -0.07 0.40 -0.07 0.26 2 Image Volumes 1 Transformation Model 1 Optimization Engine Caution: The Full Recipe? 0.29 Recipe for Recipe for Image ImageRegistration Registration ¾ Issues related to acceptance testing and quality assurance for image registration and fusion Exhale' - Inhale ΔY ΔZ -0.25 -0.44 0.09 B B -0.16B B -0.11 -0.09 -0.49 -0.18 0.37 0.29 -0.29 0.16 0.19 1 Registration Metric C o o n! So * (x , y , z ) ¾ Methods to assess the accuracy g of in image registration andmfusion o Z -2.92 -5.92 -9.42 -11.42 -3.92 1.08 n Evaluation Tools Dose Mapping? Dealing with volume elements that may: 9 deformation 9 weight loss 9 resection … not just deformations! 9 shrinkage 9 Δ vascular change shape / appear / disappear … need proper spatial re-sampling don’t necessarily add in a linear fashion … need some sort of radiobiology exist in homogenous intensity regions … hard to evaluate registration Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 19 Opportunities & Challenges What Now ? Flair Flair T T22 T T11 More than just mechanics! Gd Gd Diff Diff MR volumes mapped to CT study Summary Product Comparison ¾ Tools are now available to register and integrate image, anatomy & dose for both Tx planning and Tx delivery www.ITNonline.net ¾ These tools can be used to help build better models of the patient and to help customize and adapt therapy ¾ Work towards more standard and robust tools and validations methods (for non-rigid) situations continues th Wednesday , July 30th Don’t try this at home! 4:00 PM – 5:30 PM Joint Imaging -Therapy Symposium Image Processing and Analysis for Radiotherapy Guidance www.itk.org Deformable Image Registration Methods and Clinical Endpoints Marc L Kessler, PhD 20