2015 ANNUAL MAC-AAPM CONFERENCE: Purely Data-Driven Respiratory Motion Compensation Methods for 4D-CBCT Image Registration and Reconstruction M J Riblett 1, E Weiss 1, G E Christensen 2, and G D Hugo 1 1, 2 University of Iowa Baltimore, MD October 2 nd 2015
SUPPORT AND DISCLOSURES This work was supported by the National Cancer Institute of the National Institutes of Health under award number R01-CA-166119. The authors have no potential conflicts of interest to disclose for this study. 2 of 20
Rationale for Motion Compensation Motion Blurring Without Projection Binning (3D-CBCT) Streaking (View Aliasing) With Projection Binning (4D-CBCT) 3 of 20
Motion Compensation Methods Backproject-Deform Deform-Backproject Example: Li, 2006 Modified General Example: Template Rit, 2009 1. Reconstruct 4D-CBCT frames from a subset of the projection dataset binned according to a signal (i.e. respiration). 2. Compute an estimate of motion in each reconstructed frame and deform image. 1. Motion model is known upfront or computed prior to 4D-CBCT image reconstruction. 2. Full projection dataset is deformed based on motion model during reconstruction of each frame. 4 of 20
Motion Compensation Methods Backproject-Deform Example: Li, 2006 Matthew J. Riblett Example: Medical Physics Rit, 2009 Advantages: + Motion model can be created directly from 4D-CBCT dataset. + Day of treatment modeling. Modified General Advantages: Template Deform-Backproject + Uses full projection dataset for every frame reconstruction. + View aliasing artifact is reduced. Disadvantages: - Projection binning results in view aliasing artifact. - Registration (motion modeling) is challenging due to poor image quality. Disadvantages: - Requires an a priori motion model prior to reconstruction. - May fail to accommodate large variations in patient anatomy or motion over the course of treatment. 5 of 20
Purpose of Research To develop purely data-driven 4D-CBCT workflows combining both Matthew motion J. Riblett compensation Medical Physics methods to enhance image quality. Improved CBCT Image 1. Backproject-Deform Registration of 4D-CBCT 2. Deform-Backproject Projection-Warped Reconstruction Motion Model 6 of 20
Study Contributions Combination of Both Motion Compensation Methods Backproject-Deform: Build motion model (DVF) from groupwise registration of respiratory phase-correlated 4D-CBCT reconstruction. Deform-Backproject: Apply motion model to warp full projection data during subsequent motion-compensated 4D-CBCT reconstruction. Application of Groupwise Registration to 4D-CBCT Similar methods have demonstrated registration advantages for fanbeam CT, MR, and US. 7 of 20
Study Contributions Purely Data-Driven Methods Data-driven methods offer solutions robust to variations in patient anatomy Modified and motion General over Template the course of treatment. A priori motion modeling may be unable to handle large differences in patient anatomy or motion during treatment. Week 2 4D-CT Week 7 4D-CT 8 of 20
Groupwise Registration Conventional Registration Groupwise Registration I S2 T 2 T 3 ~ I T I S3 I S2 T G,2 T G,3 ~ I T I S3 I S1 T 1 T 4 I S4 I S1 T G,1 T G,4 I S4 Registrations between source and target frames occur independently, permitting frame-to-frame bias to manifest in the 4D transform. Registration to the target frame occurs simultaneously for all source frames mitigating frame-to-frame bias in the resulting 4D transform. 9 of 20
Developed Workflows Workflows can be subdivided by inclusion of one or both motion-compensation methods: Registration Only 1 Registration to Preselected Frame 2 Registration to Mean Frame Registration with Projection-Warping Reconstruction 3 Registration with Reconstruction of Preselected Frame 4 Registration with Reconstruction of Mean Frame 10 of 20
Data Sources and Implementation Eight Clinical Patient Datasets Long CBCT acquisitions (single rotation) 2200-3500 projections per patient set. Respiratory Signal Extraction Amsterdam shroud as implemented in RTK*. (Zijp, 2004) Used for projection sorting and reconstruction. Registration Elastix Toolkit 4.7 (Klein, 2010; Shamonin, 2014) Insight Toolkit (ITK) 4.7.0 (Yoo, 2002) Reconstruction RTK 1.0 (Rit, 2014; openrtk.org) 11 of 20
Free-Breathing 4D-CBCT Qualitative Results Registered to Preselected Frame Registered to Preselected Frame and MC-Reconstructed Free-Breathing Mean Registered to Mean Frame Registered to Mean Frame and MC-Reconstructed 12 of 20
Free-Breathing 4D-CBCT Qualitative Results Registered to Preselected Frame Registered to Preselected Frame and MC-Reconstructed Free-Breathing Mean Registered to Mean Frame Registered to Mean Frame and MC-Reconstructed 13 of 20
Free-Breathing 4D-CBCT Qualitative Results Registered to Preselected Frame Registered to Preselected Frame and MC-Reconstructed Free-Breathing Mean Registered to Mean Frame Registered to Mean Frame and MC-Reconstructed 14 of 20
Quantitative Results Statistical noise reduction relative to 4D-CBCT Initial CBCT Air Aorta Soft Tissue Air Free-Breathing Mean 64% (σ=13%) 54% (σ=20%) 41% (σ=16%) Mean Air Aorta Soft Tissue Reg. Only 63% (σ=12%) 51% (σ=20%) 34% (σ=11%) Reg./Recon. 68% (σ=15%) 55% (σ=22%) 36% (σ=13%) Preselected Air Aorta Soft Tissue Reg. Only 62% (σ=16%) 50% (σ=24%) 32% (σ=13%) Aorta Reg./Recon. 67% (σ=15%) 43% (σ=21%) 36% (σ=13%) 15 of 20
Normalized CBCT Intensities -1000-500 0 Quantitative Results Increase in edge sharpness (TIS) relative to 4D-CBCT Initial CBCT Free-Breathing Mean TIS Increase -3% (σ=56%) Mean Target Frame Reg. Only TIS Increase* 75% (σ=98%) Diaphragm Dome Profile Reg./Recon. 52% (σ=54%) Preselected Target Frame Reg. Only Reg./Recon. TIS Increase* 65% (σ=51%) 49% (σ=35%) Mean Initial 4D-CBCT Mean Frame, Reg+Recon 45 55 65 75 85 Z-axis Coordinate [mm] 16 of 20
Existing Challenges Respiratory Signal Virginia Commonwealth A University Signal acquired from either RPM or Amsterdam Modified shroud General Template for projection sorting and/or reconstruction. Choice of parameters for Amsterdam shroud impact ability to extract signal. B Noise and Artifacts Deleterious image elements cause errors in registration: latches on to erroneous signal and guides transform. Projection-warped reconstruction using: A. accurate signal B. erroneous signal 17 of 20
Conclusions Mean v. Preselected Virginia Frame Commonwealth University Image quality improvement is similar for both methods. Current implementation offers computational advantage with preselected. Reconstruction Advantage Registration improves edge sharpness and noise. MC reconstruction improves edge sharpness and image noise while also mitigating appearance of some artifacts. Free-Breathing 4D-CBCT Registration + Reconstruction 18 of 20
Conclusions Respiratory Signal Critical Correct acquisition and interpretation of respiratory signal greatly impacts initial Modified and motion-compensated General Template reconstruction. A 1 0.5 0 0 200 400 600 B 1 0.5 0 0 200 400 600 Projection-Warped Reconstructions Data-driven Respiratory Signals 19 of 20
Future Directions Improve Signal Virginia Acquisition Commonwealth University Shroud generation, signal extraction, projection sorting, etc. Additional Iterations Currently single pass Multiple iterations may continue to improve. Refine Workflow Parameters B-spline grid spacing reduction, iterations, etc. Additional Patients Near-term: 10-20 patients 64mm B-Spline Grid 16mm B-Spline Grid 20 of 20
Highlighted References Klein S, Staring M, Murphy K, Viergever MA, Pluim JPW. elastix: a toolbox for intensity based medical image registration. IEEE Transactions on Medical Imaging, 29 (1): 196 205; January 2010. Metz CT, Klein S, Schaap M, van Walsum T, Niessen WJ. Nonrigid registration of dynamic medical imaging data using nd + t b-splines and a groupwise optimization approach. Medical Image Analysis, 15 (2): 238 49, April 2011. Li T, Schreibmann E, Yang Y, Xing L. Motion correction for improved target localization with on-board cone-beam computed tomography. Physics in Medicine and Biology, 51(2): 253, 2006 Rit S, Wolthaus JW, van Herk M, Sonke JJ. On-the-fly motion-compensated conebeam CT using an a priori model of the respiratory motion. Medical Physics, 36 (6): 2283-96; June 2009. Shamonin DP, Bron EE, Lelieveldt BPF, Smits M, Klein S, Staring M. Fast parallel image registration on CPU and GPU for diagnostic classification of Alzheimer s disease. Frontiers in Neuroinformatics, 7 (50): 1-15; January 2014. Yoo TS, Ackerman MJ, Lorensen WE, Schroeder W, Chalana V, Aylward S, Metaxas D, Whitaker R. Engineering and algorithm design for an image processing API: a technical report on ITK the insight toolkit. Proc. of Medicine Meets Virtual Reality, Westwood J, ed., IOS Press Amsterdam: 586-592; 2002. Zijp L, Sonke JJ, van Herk M. Extraction of the respiratory signal from sequential thorax cone-beam X-ray images. International Conference on the Use of Computers in Radiation Therapy (ICCR). Seoul, Republic of Korea: Jeong Publishing: 507-509; 2004.
Special Thanks Dr. Geoffrey Hugo Advisor Dr. Elisabeth Weiss Collaborator Nicky Mahon Labmate Dr. Gary Christensen Collaborator Chris Guy Collaborator Eric Laugeman Labmate
Thanks for listening. ANY COMMENTS, QUESTIONS, OR SUGGESTIONS? Matthew J. Riblett: riblettmj@vcu.edu
Appendix I: PRACTICAL EXAMPLES OF CBCT COMPLICATIONS
Factors Affecting 4D CBCT Quality Undersampling and Reconstruction Artifacts Cupping and Virginia Streaking Commonwealth University View aliasing caused by frame-binning of projections and inherent undersampling in each frame.* Motion Related Degradation Averaging motion in 3D results in blurred boundaries and structures. Variable Gantry Motion and Flexing Variable image centroid Image blurring Increased X-ray Scattering Over CT ( SPR) Decreased voxel noise in individual projections Decreased contrast (CNR) Incorrect CT numbers (~30% error in MV CBCT)
MOTION EFFECTS Image: Delmon et al. (2011) Blurring of Masses Blurring of Vessels and Tissue Blurring of Diaphragm
UNDERSAMPLING Streaking (View Aliasing) Axial Sagittal
Practical Examples Example CT acquisition Projection Undersampling In CBCT
Practical Examples Example CT acquisition Motion-Averaged Blurring in CBCT
Other Examples Cupping Streaking Self-attenuation at center Variable gantry trajectory and scatter out of plane and CBCT flexing
Appendix II: EXISTING MOTION COMPENSATION METHODS
Sampling of Existing Motion Compensation Methods Authors Method Matthew J. Riblett Medical Findings Physics Limits Rit et al. (2009) A priori motion modeling with projection warping reconstruction Develops motion model from respiratory signal. DVF compensates for motion in 4D CBCT reconstruction. Requires a planning CT and an a priori motion model. Delmon et al. (2011) Sliding lung mask registration with mutual information metric Masks limit registration to sliding lung anatomy. Registration of frames results in DVF for projection warping during reconstruction. Requires masking of the lung anatomy which may require manual intervention. Metz et al. (2011) Groupwise-cyclic registration with temporal variance metric Implementation can register multiple temporal frames to reference and average frames. Has been applied to CT, MR, and US imaging. Not yet applied to CBCT. Images are transformed; not projection warping.
A priori Motion Model Method Rit et al. (2009) Acquires a 4D planning CT with respiratory signal. Offline model correlates signal to organ motion: forms 4D DVF. CBCT projections are acquired and respiratory signal extracted. 3D CBCT image is reconstructed using 4D DVF to warp projections.
Sliding Mask Method Delmon et al. (2011) Applies mutual information metric with series of sliding masks DVF is applied during CBCT reconstruction to correct projections. Results in an image with sharper vessels and tumor boundaries. Requires masks.
Groupwise Cyclic Method Metz et al. (2011): Computes cost metric as variance in the temporal dimension. Registers to an average phase instead of a reference phase. Imposes smooth cyclic motion constraint. Applied to CT, MR and US imaging. Input CT Image not to CBCT. Registered CT Image
Appendix III: PROPOSED WORKFLOW DIAGRAMS
Developed Workflows Initial 4D image Workflow parameters Registration(s) to target frame(s) DVF generation Reconstruction(s) of target frame(s) Registration Method Registration-Only Reg. and Reconstruction Mean Target Frame Preselected Target Frame Mean image of groupwise registered frame Mean image of groupwise registered frame Reconstructed image at mean target frame Reconstructed image at preselected target frame
Mean Frame Registration with(out) Reconstruction Methods Render an improved image of the patient at the 3D Modified mean General Template Goal frame of the respiratory cycle. Method Implement the groupwise registration with elastix VarianceOverLastDimension metric (VOLDM), and the reconstruction with RTK. Considerations Registers to automatically defined average temporal frame with no respiratory cycle weighting. Initial Average Frame FDK Motion Compensated
Preselected Frame Registration with(out) Reconstruction Methods Render an improved image of the patient at each Modified of the General Template Goal original frames of 4D image. Method Implement a series of groupwise 4D registrations with elastix mean squared differences (MSD) metric, and the reconstruction with RTK Considerations Registers original image to a set of pseudo-4d frames: 10 frames = 10 registrations. Initial Frame 0 FDK Motion Compensated
Hierarchical 4D Registration to 3D Frame - Registration to - Mean Frame - Initial Initial 4D Virginia Commonwealth Registration University Image Parameters Hierarchical Registration VOLDM and TBEP: Elastix and Transformix Registration with Adjusted Metric Parameters: Elastix & Transformix Acceptance Criteria 4D Transform to Average Phase Image A priori Parameters and Metrics Return Image and Transform Yes Accept Result No Adjust Registration Parameters
Registration with 3D MC Reconstruction - Reconstruction - of Mean Frame - Initial 4D Image Matthew Initial J. Riblett Medical Physics Projection Phase Registration Virginia Commonwealth Data Signal University Parameters Hierarchical Registration VOLDM and TBEP: Elastix and Transformix 3D Motion Compensated Reconstructions RTK or Simple RTK 3D 3D DVFs DVFs to to 4D DVFs Phase Phase to Phase [0 N] [0 N] [0 N] 3D Average Frame Recon. Acceptance Criteria Registration with Adjusted Metric Parameters: Elastix & Transformix Adjust Registration Parameters No Accept Image Return Image Yes A priori Parameters and Metrics
Registration with 4D MC Reconstruction - Reconstruction of 3D Frames - [0,N] and 4D Stacking - Initial 4D Image Initial Registration Parameters Projection Phase Virginia Data Commonwealth Signal University Hierarchical Registration Hierarchical Registrations MSD and TBEP: MSD and TBEP: Elastix and Transformix Elastix and Transformix 3D Motion Compensated Reconstructions RTK or Simple RTK 4D Stacking of Phase Images ribpy or Matlab 3D 3D DVFs DVFs to to 4D DVFs Phase Phase to Phase [0 N] [0 N] [0 N] Reconst. 3D Phase Images Stacked 4D- MC Image Acceptance Criteria Registration with Adjusted Metric Parameters: Elastix & Transformix Adjust Registration Parameters No Accept Image Yes Return Image A priori Parameters and Metrics
Development Steps Implementation Deliverable Component 1. VOLDM methods based on the 1. Python framework (ribpy) for work of Metz et al., and MSD image generation, manipulation, methods with Python Modified backend. General Template basic masking, and sampling. 2. Tested registration settings with clinical images parametrically. 3. Improved the methods performance with phantom model studies. 4. Reconstruct images with motion compensation: projection warping according to DVF 2. Parametric study tool for automatic review of registrations. 3. Geometric phantom generator for thorax modeling and known deformations. 4. Added HNC file I/O and flood field correction to in-house RTK deployment.
Observed Challenges Driving Data Quality of initial and motioncompensated images Modified are General Template subject to quality of acquired data (respiratory signal, projections, flood field, etc.) Static Anatomy Close proximity of static and mobile anatomy introduces challenges in registration. Computational Cost Registration and additional reconstruction carry nontrivial computational expense. 64mm B-Spline Grid 16mm B-Spline Grid
Appendix IV: PHANTOM MODELS
Simple Phantom Model
Simple Phantom Model
Geometric Anatomical Phantom
Appendix V: RESPIRATORY SIGNAL EXTRACTION
Respiratory Signal Extraction A 1 0.8 0.6 0.4 0.2 0 0 100 200 300 400 500 600 B 1 0.8 0.6 0.4 0.2 0 0 100 200 300 400 500 600 Projection-Warped Reconstructions (Motion-compensated per DVF) Data-driven Respiratory Signals (Amsterdam shroud-type signal)