Presentation is loading. Please wait.

Presentation is loading. Please wait.

A 3D U LTRASOUND-BASED T RACKING S YSTEM FOR P ROSTATE B IOPSY D ISTRIBUTION Q UALITY I NSURANCE AND G UIDANCE. PhD Thesis Michael Baumann Supervisors.

Similar presentations


Presentation on theme: "A 3D U LTRASOUND-BASED T RACKING S YSTEM FOR P ROSTATE B IOPSY D ISTRIBUTION Q UALITY I NSURANCE AND G UIDANCE. PhD Thesis Michael Baumann Supervisors."— Presentation transcript:

1 A 3D U LTRASOUND-BASED T RACKING S YSTEM FOR P ROSTATE B IOPSY D ISTRIBUTION Q UALITY I NSURANCE AND G UIDANCE. PhD Thesis Michael Baumann Supervisors Jocelyne Troccaz Vincent Daanen

2 2 Context of this thesis TIMC laboratory specializing in computer-assisted medical interventions for more than twenty years now many clinical and industrial collaborations Pitié-Salpétrière hospital, urology department active support of this work and very inspiring exchanges clinical data acquisition on more than 70 patients now Koelis SA industrial partner objective: commercialize products based on prostate tracking Registration Framework Experiments and Results Discussion Outline Conclusion/Applications ¯Introduction Prostate Tissue Tracking

3 3 Introduction Prostate Prostate Cancer most frequent cancer in men - ~ new cases in US (2007) - ~ new cases in EU25 (2006) second cause of cancer death for men deaths in US (2007) deaths in EU25 (2006) slow growing disease affects mostly elder men (>50 years) Bladder Seminal Vesicles Rectum Prostate Urethra Registration Framework Experiments and Results Discussion Outline Conclusion/Applications ¯Introduction Prostate Tissue Tracking

4 4 DRE Introduction Prostate Specific Antigene (PSA) screening biological tumor marker sensitivity for 4ng/ml threshold: 68-83% (clinically significant cancer) specificity: ~30% false positives! Digital Rectal Exams (DRE) highly varying sensitivity in clinical studies reported: 18% to 68% specificity: 4% to 33% complementary to PSA screening Prostate Biopsies Sensitivity: % (clinically significant cancer) Specificity: >95% (histological analysis) invasive programmed only if DRE/PSA positive dilemma: false negatives repeated biopsies Registration Framework Experiments and Results Discussion Outline Conclusion/Applications ¯Introduction Prostate Tissue Tracking

5 5 Prostate Biopsies 2D transrectal ultrasound (TRUS) control needle guide on probe guide aligned with longitudinal plane of probe 2D TRUS probe with needle guide Introduction longitudinal cut corresponding 2D US image with needle trajectory Registration Framework Experiments and Results Discussion Outline Conclusion/Applications ¯Introduction Prostate Tissue Tracking

6 6 Introduction Biopsy targets prostate cancer is isoechogenic systematic targets McNeals 3-zone model: central zone (CZ), transition zone (TZ), peripheral zone (PZ) 68% of cancer can be found in peripheral zone Systematic 12-core protocol clinical representation in (pseudo-)coronal plane coronal plane Registration Framework Experiments and Results Discussion Outline Conclusion/Applications ¯Introduction Prostate Tissue Tracking

7 7 Prostate Motion Problem Prostate motion main challenge for any prostate tissue tracking system displacements and deformations Transrectal biopsy specific: probe-related motion end-fire probe deformations and displacements due to probe pressure Neighboring organs (diaphragm motion, rectal and bladder filling) minor impact during prostate biopsies Registration Framework Experiments and Results Discussion Outline Conclusion/Applications ¯Introduction Prostate Tissue Tracking

8 8 Prostate Motion Problem Patient motion (small) deformations displacements with respect to surrounding tissues displacements with respect to operating room (pelvis movements!) Registration Framework Experiments and Results Discussion Outline Conclusion/Applications ¯Introduction Prostate Tissue Tracking

9 9 Introduction Biopsy and Target Localization Problem only rudimentary knowledge about biopsy position at all stages of intervention! Pre-interventional stage/planning n-core protocol target definition highly approximate targets have to be mentally mapped into patient anatomy During intervention: target localization problem difficult to aim invisible target under 2D control - ultrasound: few structural information - 2D: no depth information - prostate motion finding the target : what do we aim exactly? Registration Framework Experiments and Results Discussion Outline Conclusion/Applications ¯Introduction Prostate Tissue Tracking

10 10 Introduction Target localization problem (ctd) there exist better targets than systematic protocol high quality cancer distribution atlas available [Shen01] - simulations: biopsy sensitivity > 96% with only 6 needles (transperineal access) suspicious lesions identified on IRM repeated biopsy series - avoid already sampled tissues (negative targets) how to aim these targets? After intervention : sample localization problem where were the samples taken exactly? quality control? - are there unsampled regions? difficult to map histological cancer information back to anatomy difficult to use histological information for focal treatment planning Registration Framework Experiments and Results Discussion Outline Conclusion/Applications ¯Introduction Prostate Tissue Tracking

11 11 Magnet Resonance imaging-based approaches objective : target suspicious lesions detected on MR images biopsy under MRI control instruments calibrated with MR frame Beyersdorff [05], Musil, Krieger et al. [04,05,07], Stoianovici [07] IRM compatible biopsy acquisition instruments/robot pro: possibility to aim IRM targets con: cannot detect/compensate patient movements - would require high resolution, real-time MRI con: diagnosis: cost-benefit ratio unsatisfying - several millions of biopsies/year in US and EU Existing Solutions Registration Framework Experiments and Results Discussion Outline Conclusion/Applications ¯Introduction Prostate Tissue Tracking

12 12 Probe tracking + registration based approaches 2D transrectal ultrasound track US probe with optical or magnetic tracking system - identifies view cone motion register 2D tracking images with free-hand reference volume - identifies prostate motion Xu et al. [07]: Magnetic probe tracking + registration pro: can compensate smaller rigid prostate-movements con: free-hand volume with end-fire probe low accuracy con: rigid registration con: registration of lateral biopsy images not robust (partial gland problem) con: difficult to compensate large pelvis movements Existing Solutions Registration Framework Experiments and Results Discussion Outline Conclusion/Applications ¯Introduction Prostate Tissue Tracking

13 13 Outline of the Presentation Prostate Tissue Tracking and Guidance clinical and scientific objectives soft-tissue tracking Prostate Image Registration registration framework multi-resolution techniques image distance metric (rigid) probe movement model rigid refinement elastic registration framework forces for elastic registration Experiments and Results registration success rate accuracy biopsy maps and targeting accuracy study Discussion Conclusion and Potential Applications clinical and scientific contributions potential applications Registration Framework Experiments and Results Discussion Outline Conclusion/Applications ¯Introduction Prostate Tissue Tracking

14 14 Objectives Scientific objectives prostate tissue tracking - establish tissue correspondence - with respect to a reference space goal: establish correspondence between - biopsy site planning - reference space - needle position during intervention Clinical objectives more sophisticated targets - MRI, statistical cancer atlas, unsampled zones when repeating biopsies guide clinician to target feed-back to clinician about exact sample position - immediately and after intervention biopsy maps ¯ Prostate Tissue Tracking Introduction Registration Framework Experiments and Results Discussion Outline Conclusion/Applications

15 15 guidance: target projection into tracking volume anchor volume: acquired before intervention defines the reference space tracking volume: acquired during intervention: contains sample trajectory needle projection: projection into anchor volume projection can lead to curbed trajectories registration: establishment of correspondences for identical tissues present in both images Image-based Prostate Tracking Framework Proposed Solution 3D ultrasound-based hybrid registration - image-based - a priori model based deformation estimation no probe tracking miniminal overhead for clinician, no segmentation 3D ultrasound view cone anchor volume tracking volume biopsy map: contains projections of all samples ¯ Prostate Tissue Tracking Introduction Registration Framework Experiments and Results Discussion Outline Conclusion/Applications

16 16 Registration Framework Image Registration Optimization (minimization) problem φ = transformation model T = template/transformed image (R3 R) R = reference/fixed image D[.] = cost functional Problems registration only efficient with local minimization (downhill search) successful local minimization requires - locally unimodal cost functional - start point inside the convex region the more degrees of freedom (DOF) of φ, the more difficult to find unimodal region of D [R,T, φ ] ! Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

17 17 3 DOF6 DOF~ DOF voxel intensity based image distance metrics a priori models multi-resolution approaches optimization techniques Proposed approach Registration Framework Probe kinematics based rigid presearch Refinement of rigid estimate Elastic estimation multivariate correlation coefficient parametric systematic search SSD with local intensity shift loss-containing multi-resolution techniques endorectal probe kinematics bio-mechanical probe insertion parametric local optimization variational optimization inverse consistency linear elasticity Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

18 18 Multi-Resolution Multi-resolution approach Gaussian pyramid registration performed on different resolution levels Coarse resolutions and information loss probe kinematics rigid refinement elastic 50% of fine-grid voxel mask coarse-grid voxel else use average of available voxel attention: introduces, however, local information shifts US-specific: complex image masks problematic when computing level n+1 from level n level n level n+1 Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

19 19 Multi-resolution 50-percent rule use it for pyramid construction for interpolation for every other computation on multiple voxels - gradient computation (image distance metrics!) - Gaussian smoothing Conclusion Makes high-speed volume to volume registration possible - reliable registration on very coarse levels Disadvantage - introduces small local information shifts probe kinematics rigid refinement elastic Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework 50% rule level 5 standard level 5 standard level 5 50% rule level 5 level 1

20 20 Distance metric (rigid) Image distance metric (Rigid Registration) correlation coefficient (CC) based - well-proven for monomodal registration multivariate application - intensity image + gradient magnitude image - more robust results on coarse levels probe kinematics rigid refinement elastic Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework raw image gradient magnitude

21 21 Challenge probe used to guide needle view cone motion adds up to prostate motion motion too large for capture range of image distance metric - direct downhill/local registration only ~30-40% success rate Observations probe head always in contact with rectal wall in front of prostate - if not, no prostate image or needle trajectory outside prostate anal sphincter heavily constrains probe motion - fix point for probe motion most important rotations occur around probe axis (when switching lobe) Probe kinematics probe kinematics rigid refinement elastic Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

22 22 Probe kinematics Model of endorectal probe kinematics: approximate prostate capsule with ellipsoid from bounding box estimate rectal probe fix point admit only positions for which - the probe axis lies on the fix point - the probe origin lies on the membrane 3 degrees of freedom only - can be exhaustively explored in reasonable time! Advantages Makes solution independent of external tracking system! Solves patient motion problem! probe kinematics rigid refinement elastic Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

23 23 Rigid Refinement Refinement of rigid estimate rigid registration of 5 best transformations provided by probe kinematics high quality registration of best result probe kinematics rigid refinement elastic high quality local search: from coarse to fine high speed: optimize on coarsest level classical local/downhill search algorithm: Powell-Brent Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

24 24 Elastic Registration Prostate deformations relatively small (several millimeters) strongest near probe head difficult to estimate: - few image information near probe head Transformation model: displacement field Framework : linear elastic potential – regularizes/smoothes displacement field – minimal when no deformation strong regularizer : SSD variant to measure image distance : bio-mechanical simulation of probe insertion : inverse consistency constraints probe kinematics rigid refinement elastic Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

25 25 Elastic Registration Elastic regularization: solution scheme variational approach necessary condition for solver u* of cost function: - Gâteaux derivative at u* vanishes for all perturbations ψ Euler-Lagrange equations for linear elastic regularization: trick: separate force computation and regularization 1. accumulate forces 2. solve Euler-Lagrange equations then we get an elliptic boundary value problem of the form trick: introduce artificial time to obtain iterative gradient descent scheme probe kinematics rigid refinement elastic gradient of linear elastic potential gradients of distance metrics Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

26 26 Force terms Image based forces correlation coefficient: statistically not robust when locally computed SSD assumes identity between R and T û - does not correspond to reality! - changes in ultrasound gain, probe pressure and ultrasound direction Local intensity shift model additive model: b estimated with Gaussian convolutions on R and T resulting force term probe kinematics rigid refinement elastic Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

27 27 Force terms Bio-mechanical probe insertion model model of probe-related tissue displacements Interpret displacement differences as forces in the estimation process probe kinematics rigid refinement elastic Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

28 28 Force terms Inverse consistency forces Observation: forward and backward estimation u and v not symmetric: Zhangs approach [06] - estimate u and v simultaneously - enforce inverse consistency by minimizing - alternating optimization process: resulting force term probe kinematics rigid refinement elastic Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

29 29 Proposed approach Registration Framework Probe kinematics based rigid presearch Refinement of rigid estimate Elastic estimation multivariate correlation coefficient parametric systematic search SSD with local intensity shift loss-containing multi-resolution techniques endorectal probe kinematics bio-mechanical probe insertion parametric local optimization variational optimization inverse consistency linear elasticity Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

30 30 Experiments and results

31 31 Experiments and Results Experiments on real patient data Pitie-Salpétrière Hospital, Paris, urology department - P. Mozer, G. Chevreau, S. Bart, J.-C. Bousquet 3D ultrasound images (GE Voluson, RIC5-9 probe) - acquired before biopsies and after each sample acquisition - targeting carried out under 2D US control Registration example Introduction Prostate Biopsy Tracking Registration Framework Discussion Outline Conclusion/Applications ¯ Experiments and Results

32 32 Experiments and Results Rigid Registration Algorithm tested on 785 image pairs from 47 patients 27 mis-registrations (success-rate 96.5 %) Conclusion probe movement model works fine! ultrasound depth ultrasound quality partial contact Introduction Prostate Biopsy Tracking Registration Framework Discussion Outline Conclusion/Applications ¯ Experiments and Results

33 33 Experiments and Results Accuracy study 208 registrations on data from 14 patients manual point fiducial segmentation (calcifications, dark spots) error computed on Euclidean distances of corresponding fiducials Registration accuracy rigid optimization performed on resolution levels 5 to 3 elastic optimization performed on resolution levels 6 to 3 Conclusion accuracy sufficient for many clinical applications Introduction Prostate Biopsy Tracking Registration Framework Discussion Outline Conclusion/Applications ¯ Experiments and Results

34 34 Experiments and Results First Application: Biopsy maps show targeting difficulties P. Mozer, M. Baumann, G. Chevreau, A. Moreau-Gaudry [Mozer08] apex and base targets more difficult to reach than central gland operator learning curve measured Introduction Prostate Biopsy Tracking Registration Framework Discussion Outline Conclusion/Applications ¯ Experiments and Results

35 35 Discussion

36 36 Discussion Automatic registration validation visual validation time-consuming and operator-dependent open issue: automatic detection of failures! necessary for guidance! Registration and real-time requires 5 – 15 seconds stream parallelization: - algorithm mainly consists of image convolutions - can be parallelized on a voxel per voxel basis - well suited for latest graphic card architectures (stream processors) registration times of 1 second or less should be feasible Similarity measures Good performance for intra-series registration Still to be evaluated for inter-series registration - only one patient with two biopsy series for instance Intensity shift model - depends strongly on parameter σ of Gaussian convolution Introduction Prostate Biopsy Tracking Registration Framework Outline Conclusion/Applications ¯Discussion Experiments and Results

37 37 Discussion Probe movement model (rigid registration) very good success rate - no probe tracking necessary less hardware in OR! Simpler workflow and logistics! improvements with model to data fitting possible should further improve success rate Bio-mechanical probe insertion model (elastic registration) for about 50% image pairs, the model improves elastic registration but: sometimes inadequate model of reality Introduction Prostate Biopsy Tracking Registration Framework Outline Conclusion/Applications ¯Discussion Experiments and Results

38 38 Discussion Clinical acceptability only slight modification of classical acquisition protocol - bounding box placement - registration validation (probably post-op step) no additional instruments/hardware in operation room cost effective: cost similar to current procedure Introduction Prostate Biopsy Tracking Registration Framework Outline Conclusion/Applications ¯Discussion Experiments and Results

39 39 Conclusion and Applications

40 40 ¯Conclusion/Application Conclusion Scientific contributions probe movement model - robust - no probe tracking hardware required - completely solves patient movement problem - reusable for many endocavitary US interventions! loss-containing multi-resolution filtering and interpolation - robust optimization on very sparse resolution levels hybrid model- and image-based elastic deformation framework novel voxel similarity measure for elastic registration - remarkably robust - simple proof of concept on large set of patient data Medical contributions biopsy accuracy study on biopsy maps - more difficult to reach apex/base than mid-gland operator learning curve proven Introduction Prostate Biopsy Tracking Registration Framework Outline Discussion Experiments and Results

41 41 Future work/Prospects Potential Applications : Biopsies biopsy maps - immediate feed-back, post-interventional quality control cancer maps - map histological results on 3D biopsy map guidance - assist clinician during targeting - requires automatic registration validation and real-time registration guidance MRI target mapping - reach MRI targets under ultrasound control - requires – MRI to ultrasound registration guidance repeated biopsy series - avoid multiple sampling - visualize already sampled tissues guidance cancer atlas targets - define targets with cancer probability atlas (Shen01) - map them onto anchor volume – requires atlas to ultrasound volume registration Introduction Prostate Biopsy Tracking Registration Framework Outline Discussion Experiments and Results ¯Conclusion/Applications

42 42 Future work/Prospects Potential applications : Therapy improve accuracy of ultrasound-guided therapy - brachytherapy, HIFU, cryotherapy, … focal therapy? - currently: two unknowns after positive biopsy findings 1. shape of the tumor 2. exact location of the biopsy - not accurate enough for focal therapy - we solve 2! - sufficient for focal therapy? – in combination with statistical tumor atlas? Introduction Prostate Biopsy Tracking Registration Framework Outline Discussion Experiments and Results ¯Conclusion/Applications

43 43 Publications and References Publications [Baumann07] M. Baumann, P. Mozer, V. Daanen, J. Troccaz. Towards 3D Ultrasound Image Based Soft Tissue Tracking: a Transrectal Ultrasound Prostate Image Alignment System. MICCAI'07, Brisbane, Australia, Springer LNCS [Mozer07] P. Mozer, M.Baumann, G. Chevreau, J. Troccaz. Fusion dimages : application au contrôle de la distribution des biopsies prostatiques, Progrès en Urologie (les Cahiers de la Formation Continue), vol. 18 (1), 2008 [Baumann08] M. Baumann, P. Mozer, V. Daanen, J. Troccaz. Fast and robust elastic registration of endorectal 3D ultrasound prostate volumes for transrectal prostate needle puncture tracking, In proceedings of CARS08, Barcelona, 2008 References [Shen04] D. Shen, Z. Lao, J. Zeng, W. Zhang, I. A. Sesterhenn, L. Sun, J. W. Moul, E. H. Herskovits, G. Fichtinger, and C. Davatzikos. Optimization of biopsy strategy by a statistical atlas of prostate cancer distribution, Medical Image Analysis, vol. 8, no. 2, pp. 139–150, [Zhang05] Z. Zhang, Y. Jiang, and H. Tsui. Consistent multi-modal non-rigid registration based on a variational approach, Pattern Recognition Letters, pp. 715–725, 2006.

44 44 Acknowledgements Urology department Pitié-Salpétrière Pierre Mozer, Grégoire Chevreau, Stéphane Bart Koelis SA Antoine Leroy Vincent Daanen TIMC GMCAO group Jocelyne Troccaz and everyone else who supported this project during the last three years! Funding: : Programme Hospitalier de Recherche Clinique - Prostate-Echo, French ministry of research : Surgétique Minimalement Invasive (SMI), Agence Nationale de Recherche (ANR) : Association Nationale de la Recherche Technique, bourse CIFRE

45 45 Inadequate probe model possible explanation

46 46 Separate elastic estimation first step: - estimate deformations caused by probe forces second step: - estimate deformations caused by image forces - start optimization with probe deformation as initial guess

47 47 Elastic Regularization Elastic regularization: solution scheme (ctd) von Neumann stability analysis of numerical scheme yields Stability criterion and elasticity parameters forces in our framework are not physical - derived from distances - how to calibrate them with the elastic forces? Youngs modulus E has no physical meaning - interpret it as free parameter - control elasticity parameters with Poissons coefficient v and t seek best balance between smoothness and convergence rate balance elastic smoothing and maximally admitted deformation probe kinematics rigid refinement elastic Youngs modulus Poissons coefficient Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

48 48 Elastic Registration Elastic regularization: solution scheme (ctd) solved with Gauss-Seidel and full multigrid strategy Boundary conditions bending side-walls, fixed edges good model for probe insertion probe kinematics rigid refinement elastic Introduction Prostate Tissue Tracking Experiments and Results Discussion Outline Conclusion/Applications ¯ Registration Framework

49 49 Biopsy acquisition patient in dorsal or lateral position local anesthesia 12 acquisitions Introduction


Download ppt "A 3D U LTRASOUND-BASED T RACKING S YSTEM FOR P ROSTATE B IOPSY D ISTRIBUTION Q UALITY I NSURANCE AND G UIDANCE. PhD Thesis Michael Baumann Supervisors."

Similar presentations


Ads by Google