Circa Circa 1986 1986 In The Beginning! Imaging in Therapy: Image Registration & Data Fusion Algorithms * * Marc L Kessler, PhD Department of Radiation Oncology The University of Michigan, Ann Arbor * * Disclosure Disclosure online online And here! Look here! Acknowledgements! Many wonderful people have contributed material for this presentation ! Outline ¾ What is image registration ? ¾ Why do we want to do it ? ¾ How do we do it ? the mechanics ! Data Handling for IGRT Imaging Studies Process Images (Actual) (Digitized) (Segment, Register) Tx Plan DRRs Radio graphs (Virtual) (Virtual) (Actual) Patient Patient (Virtual) Patient + … Dose (Actual) Data Handling for IGRT Imaging Studies Process Images (Actual) (Digitized) (Segment, Register) Tx Plan DRRs Radio graphs (Virtual) (Virtual) (Actual) Patient Patient (Virtual) Patient + … Dose (Actual) Data Handling for IGRT Volumetric data Imaging Patient acquired Studies at the treatment unit! (Actual) (Digitized) Tx Plan DRRs (Virtual) Process Images (Segment, Register) On-line and (Virtual) Patient (Virtual) Patient + … off-line Dose Radio graphs (Actual) (Actual) Data Handling for IGRT Once you look, … you will see! Once you see, … you will have to act ! maybe Data Handling for IGRT ! ? … what do we do now? Data Handling for IGRT adaptive Imaging Planning Delivery Imaging … bi-directional … flow betterbecomes bulk up on DICOM too! Data Handling for IGRT Txx Planning CT MRI cor,sag,axial cor,sag,axial NucMed TBD Tx Plan 3D Dose “Adapting” Patient Model 3D Dose Day n Cone Beam 1…n Portal Images 4D Cone Beam Marker Locations How ? To get the plethora of data to all groove together we need to know the geometric transformations F that relate the coordinates of the different imaging studies How ? 1 Compute the geometric transformation that relates the coordinate systems of two datasets registration 2 Apply the computed transformation to map information from one dataset to another data fusion How ? Reference Image Moving Image Data Fusion Image Registration Image Registration a2 ? b2 b1 a1 a3 Study A rotate, scale translate, deform ... b3 Study B XB = F ( XA , { ß }) Data Fusion XBB = F ( XAA , { ß }) tumor volumes 3D doses, ... Study A ? Study B Computer graphics and image processing Degrees of Freedom Patient dependent – rotation ( θxx , θyy , θzz ) – translation ( txx , tyy , tzz ) – distortion many DOF Machine dependent – pixel size – slice thickness – distortion ( sxx , syy , szz ) many DOF {ß} What is F ? ¾ Rigid / Affine ¾ Piecewise Rigid / Affine … limited field-of-view ¾ Full 3D / 4D Deformation Parametric models Free-form models What is F ? Affine Assumption y = mx+b … in three dimensions xB = A xA + b Otherwise rotation (3) … up to 12 DOF translation (3) scale (3) Spatially variant function shear (3) Various splines and free form models … lots of degrees of freedom! www.gnome.org www.gnome.org Affine Transformations Parallel lines stay parallel! non- Affine Transformations Parallel lines don’t stay parallel! non- Affine Transformations XB = F ( XA , { ß }) non- Affine Transformations XB = F ( XA , { ß(XA) }) non- Affine Transformations Transformation parameters to apply to a voxel depends on the location of the voxel … up to 3N parameters XB = F ( XA , { ß(XA) }) DICOM 3 Parts 3 & 17 DICOM handles only up to affine ... (and most Tx planning systems) New DICOM Objects ? OK Marc New DICOM Objects ? No Marc How ? Prospective ¾ reproduce imaging geometry exactly ¾ attach coordinate system to patient • frames / fiducials Retrospective ¾ patient intrinsic • anatomy / shape / image intensities Prospective PET / CT Hybrid GE Discovery LS F = Identity XB = XA Prospective … attach a coordinate system to the patient! ouch … Stereotactic Radiosurgery Retrospective CT ? – rotate – translate – scale – warp? MR Geometry-Based Intensity-based Interactive or Automated Automated Registration 1 Construct a metric that measures the mismatch (or similarity) between a pair of datasets 2 Apply a minimization algorithm to determine the parameters (DOF) that minimize (maximize) this metric Automated Optimization Twiddle Parameters Evaluate Metric (DOF) (Cost) Registration Metrics ? ? Geometry-based Intensity-based Retrospective ¾ Point Matching Least Squares ¾ Surface Matching Chamfer Matching ¾ Voxel Intensities Sum of squares diff Σ(X B B - XAA ) 22 Σ min distance 22 Σ ( IBB - IAA ) 22 Mono-modality Retrospective ¾ Point Matching Least Squares ¾ Surface Matching Chamfer Matching Σ(X B B ) 22 Σ min distance 22 ¾ Voxel Intensities Mutual Information - XAA - Σ p(A,B) log p(A,B) p(A) p(B) Multi-modality Catallo Catallo // UM UM Geometry-based … using an extracted anatomic surfaces a ? b c Σ di 2 objects misaligned compute mismatch mismatch minimized Intensity-based … using an information theory-based approach CT MR ? H(ICT) H(IMR) Individual Information Content H(ICT,IMR) Joint Information Content Information Theory H(IA,IB) = H(IA) + H(IB) - MI(IA,IB) Joint Entropy MI(IA,IB) = Individual Entropies Mutual Information Σ p(I , I ) log A A B B 2 p(IAA, IBB) p(IAA) p(IBB) These are just intensity histograms! Information Theory H(IA,IB) = H(IA) + H(IB) - MI(IA,IB) The mutual information of two image datasets is a maximum when they are geometrically registered … … MI can be used as a metric ‘48 Shannon - Bell Labs / ‘95 Viola - MIT Mutual Information I reformatted I CT CT MI = .99 M R Aligned! p(ICT, IMR) original MR 2D joint intensity histogram Mutual Information p(ICT, IMR) I reformatted I CT CT MI = .62 M R Not so Aligned! original MR 2D joint intensity histogram Automated Optimization Rotate, Translate, Evaluate MI Deform? Metric Automated Registration MR CT Maximize Mutual Information In the Brain Multimodality image registration in the cranium is a solved problem! CT / MR PET / CT MR / PET Outside the Brain ? There is a need to better handle data from other regions of the body! Prostate Liver Spine Outside the Brain ? Rotate-translate and MI can still be used effectively, … over a limited field-of-view geometric anatomic piecewise Region Region 11 Region Region 22 Prostate Example MR data discarded curved curved flat flat MR CT radioactive seeds Prostate Example Can you tell what is different in the 2 images? Prostate Example Bones aligned, prostate region not aligned No Cropping The “answer” depends on the region defined! Prostate Example Bones ignored, prostate region aligned Cropping The “answer” depends on the region defined! Liver Example Dawson Dawson // PMH PMH Most of the motion of the liver seems to be rigid or affine! … some deformation does occur though. Liver Example anatomic-based cropping ignore anatomy outside liver Liver Example anatomic-based cropping ignore anatomy outside liver Liver Example anatomic-based cropping ignore anatomy outside liver Limited Field-of-View Rigid assumption used for regional registration Pick Your Battles Wisely! Limited Field-of-View Oops! What about Deformations ? Image Registration 201 4 credit course 3 credits - science 1 credit - art non- Affine Transformations Displacement Function Displacement Vectors non- Affine Transformations XB = F ( XA , { ß }) non- Affine Transformations XB = F ( XA , { ß(XA) }) non- Affine Transformations Transformation parameters to apply to a voxel depends on the location of the voxel … up to 3N parameters XB = F ( XA , { ß(XA) }) What about Deformations ? Spatially variant transformations ¾ B-Splines ¾ Thin-Plate splines parametric ¾ Finite element models ¾ Intensity flow models free-form What about Deformations ? Each have some distinct properties ¾ B-Splines … local ¾ Thin-Plate splines … global ¾ Finite element … bio-mechanical ¾ Intensity flow … image forces ( mono-modality ) What about Deformations ? Each have some distinct properties ¾ B-Splines X’ = X + ΔX = X + Σ wi·Β(X-ki) weights ( … parameters! ) basis function What about Deformations ? Each have some distinct properties ¾ B-Splines Before 1-D Example voxel voxel 11 voxel voxel 22 voxel voxel 33 voxel voxel 44 voxel voxel 55 voxel voxel 66 voxel voxel 77 voxel voxel 33 After voxel voxel 11 voxel voxel 22 voxel voxel 44 voxel voxel 55 voxel voxel 77 voxel voxel 66 What about Deformations ? Each have some distinct properties ¾ B-Splines Before After 1-D Example voxel voxel 11 voxel voxel 22 voxel voxel 33 voxel voxel 44 voxel voxel 55 voxel voxel 66 voxel voxel 77 What about Deformations ? Each have some distinct properties ¾ B-Splines 1-D Example ΔX w w77 knots kk11 kk22 kk33 kk44 kk55 kk66 kk77 kk88 kk99 kk10 k 11 10 k11 X X’ = Xa +lot ΔXlike = Xintensity + Σ wii·Βmodulation! (X-kii) This seems B-Spline Transformation Model “Knots” 3-D Grid of Control Points Multiresolution Deformations Divide and Conquer 60 60 xx 60 60 xx 48 48 mm mm 4 4 xx 4 4 xx 3 3 mm mm Coarse Fine Knot Spacing ABC Breath hold Segmental 4D CT Balter Balter // UM UM Multiresolution Deformations ABC CT Example Exhale State Inhale State Multiresolution Deformations ABC CT Example Exhale State Inhale State deformed B-Spline Deformations ABC CT Example Inhale Inhale deformed to match exhale B-Spline Deformations MR - CT Example B-Spline Deformations MR - CT Example CT CT MR MR CT CT MR MR CT CT MR MR Coming to a DICOM server near you soon? X X deformation deformation Y Y deformation deformation Z Z deformation deformation magnitude magnitude Multiresolution Deformations MR-CT Example Reference Dataset Homologous Dataset Multiresolution Deformations MR-CT Example Reference Dataset Homologous Dataset Multiresolution Deformations MR-CT Example Split screen Image Switch We are not really Splines! Tissue Sliding Rigid + Deformation Brock Brock // UM UM Finite Element Modeling Exhale Exhale Inhale Inhale Take into account physical tissue properties Brock Brock // PMH PMH Finite Element Modeling … thorough segmentation is necessary Data Fusion Image Mapping / Fusion resample one image series to match the scale and orientation of another Structure Mapping / Transfer transfer contours defined in one image series (MR) to another (CT) Image Mapping / Fusion MR Study A Study A’ CT ... resample one study to match scale and orientation of the other Image Mapping / Fusion CT CT’ MR CT’ ... resample one study to match scale and orientation of the other Structure Mapping ? ... map structure outlines defined on one study on the other Structure Mapping Stack the Outlines A Create a 3D model B Structure Mapping Transform and Slice C Apply Outlines to CT D How … do we know these algorithms work? ¾ build phantoms and test them we can know the truth! ¾ provide tools to examine results we don’t know the truth! Multimodality Phantoms CT MR 1986 1986 van van Herk Herk // NKI NKI Multimodality Phantoms Validation Visualization Tools ¾ Color gel or wash overlay ¾ Split /dual screen displays ¾ Anatomic boundary overlay! Validation Split Screen Display CT CT MR MR Validation Image Switch Validation Structure Overlay Auto-segmented boundaries from CT mapped to PET Validation Threshold + Compositing MR CT bone window Validation Threshold + Compositing MR CT bone window Validation Colorwash MR + PET overlay CT + PET overlay Validation + CT GE GE Web Web Site Site = PET CT + PET Validation Numerical Tools Point 1 2 3 4 5 6 Description 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 A Exhale Y 0.98 2.12 2.77 3.37 -1.95 2.47 * Z -3.42 -5.42 -8.42 -9.92 -4.42 0.58 X -4.62 A A -5.40 -6.24 -8.12 -7.67 Study -10.78B (x , y , z ) All numbers in centimeters all values in cm. Inhale Y -0.22 A A 0.74A A 0.80 1.40 -3.20 1.16 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 ΔX -0.09 0.05 B B -0.03 -0.07 0.40 -0.07 Z -2.92 -5.92 -9.42 -11.42 -3.92 1.08 * 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 (x , y , z ) 0.19 0.29 0.26 Examples Treatment Planning using Magnetic Resonance Imaging Treatment Delivery using volumetric information Target Volume Definition … draw on MR Axial “Target” Coronal “Target” + Optic Structures Target Volume Definition … map to CT Boolean OR OR ? Target Volume Definition % Volume … optimize as usual ! Dose-Volume Histogram % Dose Dose Visualization CT-based dose Displayed on MR Dose Mapping Inhale CT Exhale CT Rosu Rosu // UM UM Munro Munro // Varian Varian Registration @ Delivery CT CBCT CT CBCT Munro Munro // Varian Varian Registration @ Delivery Multi-res Gustavo Gustavo // Tomo Tomo Registration @ Delivery Original planning CT Reference CT Daily CT Daily CT mapped to Reference CT Gustavo Gustavo // Tomo Tomo Registration @ Delivery Original planning CT Reference CT Daily CT Daily CT mapped to Reference CT Gustavo Gustavo // Tomo Tomo Registration @ Delivery Map contours too! Summary Taxonomy of Registration Process Geometry Intensity Interactive Automated Affine Warping Summary ¾ Techniques are now available to register 3D/4D image data from different modalities ¾ Registered data can be fused to create more complete models of the patients ¾ Accuracy on the order of image resolutions reported for “well behaved” situations Take Home Quiz! www.ITNonline.net Thank you for your time ! There’s more than one way to scan a cat