optimization of acquisition geometry for intra operative
play

Optimization of Acquisition Geometry for Intra-operative Tomographic - PowerPoint PPT Presentation

Optimization of Acquisition Geometry for Intra-operative Tomographic Imaging J. Vogel T. Reichl J. Gardiazabal N. Navab T. Lasser Computer-Aided Medical Procedures, Technische Universit at M unchen Institute of Biomathematics and


  1. Optimization of Acquisition Geometry for Intra-operative Tomographic Imaging J. Vogel T. Reichl J. Gardiazabal N. Navab T. Lasser Computer-Aided Medical Procedures, Technische Universit¨ at M¨ unchen Institute of Biomathematics and Biometry, HelmholtzZentrum M¨ unchen MICCAI October 4, 2012 1

  2. Motivation ◮ Enable flexible intra-operative functional imaging ◮ Identify cancer tissue during surgery using radioactive tracer 2

  3. SPECT 3

  4. SPECT 3

  5. SPECT 3

  6. SPECT 3

  7. SPECT 3

  8. SPECT 3

  9. SPECT 3

  10. Diagnostic SPECT 4

  11. Freehand SPECT SurgicEye Press Picture T. Wendler et al., Eur J Nucl Med Mol Imaging 37 (8), 2010 5

  12. Freehand SPECT Courtesy of Aslı Okur & Thomas Wendler T. Wendler et al., Eur J Nucl Med Mol Imaging 37 (8), 2010 5

  13. Intra-Op Tomographic Imaging Robotic SPECT C-Arm CT Siemens Press Picture ◮ Problem: Find optimal and reproducible trajectories! Y. Murayama et al., Neurosurgery 68, 2011 6

  14. Sneak Peek ◮ Optimize sensor trajectory for tomographic reconstruction ◮ Directly use mathematical framework ◮ Control robotic arm accordingly 7

  15. Algebraic Reconstruction: Discretized Signal f = � f : Ω ⊂ R 3 → R f ≈ � i x i · b i 8

  16. Algebraic Reconstruction: Measurement Model m j ≈ M j ( � M j : (Ω → R ) → R i x i · b i ) m j = M j ( f ) � m j = f ( x ) d x L j 9

  17. Algebraic Reconstruction: Measurement Model m j ≈ M j ( � M j : (Ω → R ) → R i x i · b i ) = � m j = M j ( f ) i x i · M j ( b i ) � �� � � =: a ji m j = f ( x ) d x L j 9

  18. Algebraic Reconstruction: Linear System � a T m j = j , x �  — a 1 —      = m    x  10

  19. Algebraic Reconstruction: Linear System � a T m j = j , x �   — a 1 —   — a 2 —   = m   x   10

  20. Algebraic Reconstruction: Linear System � a T m j = j , x �   — a 1 —   — a 2 —   = m   x  — a 3 —  . . . = A · x ◮ A contains the geometry, m the measurements ◮ Kernel and rank of A are quality indicators 10

  21. Singular Value Spectrum of a System Matrix 1 10   0 10 σ 1   V T A = U  σ 2  singular value intensity −1 10 ... � �� � −2 10 =: S � −3 10 η ( A ) = i σ i −4 10 singular value index ◮ SVD is slow T. Lasser et al., Medical Image Analysis 11 (4), 2007 P. C. Hansen et al., Johns Hopkins University Press 2012 11

  22. Pivoted QR Decomposition 1 10   0 10 r 11 r 12 · · ·   r 22 · · · P T A = Q   singular value intensity −1 10 ... � �� � −2 10 =: R � −3 10 η ( A ) = � diag ( R ) � ℓ 1 = i | r ii | −4 10 singular value index ◮ Pivoted QR is considerably faster P. C. Hansen et al., Johns Hopkins University Press 2012 T. Lasser et al., Medical Image Analysis 11 (4), 2007 12

  23. Robot Control: Overview ◮ Find trajectory maximizing cost function η ◮ Constrain motion to bounding surface possible next destinations current destination p i current location 13

  24. Robot Control: Overview ◮ Find trajectory maximizing cost function η ◮ Constrain motion to bounding surface possible next destinations currently optimal η current destination p i current location 13

  25. Robot Control: Overview ◮ Find trajectory maximizing cost function η ◮ Constrain motion to bounding surface possible next destinations currently optimal η current destination p i current location 13

  26. Robot Control: Overview ◮ Find trajectory maximizing cost function η ◮ Constrain motion to bounding surface possible next destinations currently optimal η current destination p i current location 13

  27. Robot Control: Overview ◮ Find trajectory maximizing cost function η ◮ Constrain motion to bounding surface possible next destinations currently optimal η current destination p i current location 13

  28. Robot Control: Overview ◮ Find trajectory maximizing cost function η ◮ Constrain motion to bounding surface current dest. p i +1 current location p i 13

  29. Robot Control: Exploring the Surface bounding triangle-mesh basis functions b i 14

  30. Robot Control: Exploring the Surface � � A i random pose, evaluate η — a ∗ — ◮ Predicted matrix A i after robot reaches current destination p i 14

  31. Robot Control: Exploring the Surface high need for measurement highest energy next destination low need 14

  32. Real-Time Implementation ◮ Small voxel basis (10 × 10 × 7 = 700 basis functions) for optimization (finer for actual reconstruction) ◮ Mesh of 200–450 triangles, evaluated in parallel ◮ In-place decomposition using LAPACK’s SGEQP3 15

  33. Experiments: Simulation grid expert random robot−1 robot−2 deviation of major hotspot 0 500 1000 1500 2000 2500 3000 3500 4000 number of measurements used for reconstruction Alexander Hartl contributed. 16

  34. Experiments: Simulation grid expert random robot−1 robot−2 deviation of major hotspot 0 500 1000 1500 2000 2500 3000 3500 4000 number of measurements used for reconstruction Alexander Hartl contributed. 16

  35. Experiments: Simulation grid expert random robot−1 robot−2 deviation of major hotspot 0 500 1000 1500 2000 2500 3000 3500 4000 number of measurements used for reconstruction Alexander Hartl contributed. 16

  36. Experiments: Simulation grid expert random robot−1 robot−2 deviation of major hotspot 0 500 1000 1500 2000 2500 3000 3500 4000 number of measurements used for reconstruction Alexander Hartl contributed. 16

  37. Experiments: Real World Setup Result 17

  38. Conclusion ◮ Sensor trajectory optimization for tomographic reconstruction ◮ General approach directly using the mathematical framework ◮ Real-time implementation available 18

  39. Acknowledgements ◮ DFG SFB 824 ◮ DFG Cluster of Excellence MAP ◮ European Union FP7 grant N o 25698 ◮ Looking forward to meet you at our poster ( Th-2-AG-01 ), today 3:00 – 4:30 pm 19

  40. Singular Value Spectrum: Sample Cases 1 10 0 10 singular value intensity −1 10 −2 10 −3 10 −4 10 singular value index 700 basis functions, 522 measurements, case id 94 21

  41. Singular Value Spectrum: Sample Cases 1 10 0 10 singular value intensity −1 10 −2 10 −3 10 −4 10 singular value index 700 basis functions, 522 measurements, case id 171 21

  42. Singular Value Spectrum: Sample Cases 1 10 0 10 singular value intensity −1 10 −2 10 −3 10 −4 10 singular value index 700 basis functions, 853 measurements, case id 190 21

  43. Singular Value Spectrum: Sample Cases 1 10 0 10 singular value intensity −1 10 −2 10 −3 10 −4 10 singular value index 700 basis functions, 853 measurements, case id 249 21

  44. Singular Value Spectrum: Sample Cases 1 10 0 10 singular value intensity −1 10 −2 10 −3 10 −4 10 singular value index 700 basis functions, 939 measurements, case id 167 21

  45. Singular Value Spectrum: Sample Cases 1 10 0 10 singular value intensity −1 10 −2 10 −3 10 −4 10 singular value index 700 basis functions, 939 measurements, case id 233 21

  46. SVD vs. QR Spectra: Sample Cases 1 10 0 10 singular value intensity −1 10 −2 10 −3 10 −4 10 singular value index 700 basis functions, 522 measurements, case id 94 23

  47. SVD vs. QR Spectra: Sample Cases 1 10 0 10 singular value intensity −1 10 −2 10 −3 10 −4 10 singular value index 700 basis functions, 522 measurements, case id 94 23

  48. SVD vs. QR Spectra: Sample Cases 1 10 0 10 singular value intensity −1 10 −2 10 −3 10 −4 10 singular value index 700 basis functions, 853 measurements, case id 190 23

  49. SVD vs. QR Spectra: Sample Cases 1 10 0 10 singular value intensity −1 10 −2 10 −3 10 −4 10 singular value index 700 basis functions, 853 measurements, case id 190 23

  50. SVD vs. QR Spectra: Sample Cases 1 10 0 10 singular value intensity −1 10 −2 10 −3 10 −4 10 singular value index 700 basis functions, 939 measurements, case id 167 23

  51. SVD vs. QR Spectra: Sample Cases 1 10 0 10 singular value intensity −1 10 −2 10 −3 10 −4 10 singular value index 700 basis functions, 939 measurements, case id 167 23

  52. SVD vs. QR Spectra: Random Paths SVD QR 0.25 0.2 energy 0.15 0.1 0.05 0 0 10 20 30 40 50 60 70 80 90 100 iterations 25

  53. SVD vs. QR Spectra: Random Paths SVD QR 0.25 0.2 energy 0.15 0.1 0.05 0 0 10 20 30 40 50 60 70 80 90 100 iterations 25

  54. SVD vs. QR Spectra: Random Paths SVD QR 0.25 0.2 energy 0.15 0.1 0.05 0 0 10 20 30 40 50 60 70 80 90 100 iterations 25

  55. SVD vs. QR Spectra: Random Paths SVD QR 0.25 0.2 energy 0.15 0.1 0.05 0 0 10 20 30 40 50 60 70 80 90 100 iterations 25

  56. Robot Control: Correlation Path – Energy 27

  57. Robot Control: Correlation Path – Energy 27

Recommend


More recommend