Deformable registration using shape statistics with applications in - - PowerPoint PPT Presentation

โ–ถ
deformable registration using shape statistics
SMART_READER_LITE
LIVE PREVIEW

Deformable registration using shape statistics with applications in - - PowerPoint PPT Presentation

Oct 10, 2023 266 likes โ€ข947 views

Deformable registration using shape statistics with applications in sinus surgery Ayushi Sinha March 22 nd , 2018 1 In endoscopic surgery endoscope G. Scadding et al., Diagnostic tools in Rhinology EAACI position paper , Clinical and

slide-1
SLIDE 1

Deformable registration using shape statistics

with applications in sinus surgery Ayushi Sinha

March 22nd, 2018

1

slide-2
SLIDE 2

In endoscopic surgery

2

  • Restricted field of view
  • Need

ed to know surrounding and

  • ccluded structures
  • G. Scadding et al., Diagnostic tools in Rhinology EAACI position paper, Clinical and

Translational Allergy, 1(2), 2011

endoscope

slide-3
SLIDE 3

3 A Sin inha ha, et al., Automatic segmentation and statistical shape modeling of the paranasal sinuses to estimate natural variations, SPIE Medical Imaging, 2016 A Sin inha ha, et al., Simultaneous segmentation and correspondence improvement using statistical modes, SPIE Medical Imaging, 2017

slide-4
SLIDE 4

4 S Leonard, A Reiter, A Sinh nha, et al., Image-based navigation for functional endoscopic sinus surgery using structure from motion, SPIE Medical Imaging, 2016

  • S. D. Billings, A. Sin

inha ha, et al., Anatomically Constrained Video-CT Registration via the V-IMLOP Algorithm, MICCAI, 2016 S Leonard, A Sin inha ha, et al., Evaluation and Stability Analysis of Video-Based Navigation System for Functional Endoscopic Sinus Surgery on In-Vivo Clinical Data, Trans. Medical Imaging, 2018 (in submission)

slide-5
SLIDE 5

In clinic

5

  • For diagnosis and planning
  • Need

ed to know how much patient anatomy diverges from normal

  • G. Scadding et al., Diagnostic tools in Rhinology EAACI position paper, Clinical and

Translational Allergy, 1(2), 2011

endoscope

slide-6
SLIDE 6

Objectives

  • Surgical/clinical navigation without additional tools
  • Compensate for deformations in anatomy
  • Estimate anatomy in the absence of CT

6

slide-7
SLIDE 7

Accomplishments

  • Technical
  • Clinical

7

Segmentat entation ion and modeli eling ng Corr rrespondence espondence improv

  • vem

ement ent Deformab

  • rmable

le registrati istration

  • n

Anatom

  • mica

ical l variat ation ion Nasal al cyc ycle le Nasal al patency ency

slide-8
SLIDE 8

Outline

8

  • Statistical shape models
  • Iterative closest point (ICP)
  • Iterative most likely point (IMLP)

Background Deformable registration framework Results and confidence estimates

slide-9
SLIDE 9

Outline

9

  • Deformable iterative most likely point (D-IMLP)
  • Deformable iterative most likely oriented point (D-IMLOP)
  • Generalized deformable iterative most likely oriented point

(GD-IMLOP)

Background Deformable registration framework Results and confidence estimates

slide-10
SLIDE 10

Outline

10

  • Results on simulation data
  • Confidence associated with results
  • Results on in-vivo clinical data

Background Deformable registration framework Results and confidence estimates

slide-11
SLIDE 11

Statistical shape models

โ‹ฏ

๐–

1 =

v11 v12 โ‹ฎ v1๐‘œv ๐–2 = v21 v22 โ‹ฎ v2๐‘œv ๐–๐‘œs = v๐‘œs1 v๐‘œs2 โ‹ฎ v๐‘œs๐‘œv

โ‹ฏ

11

A Sin inha ha, et al., Automatic segmentation and statistical shape modeling of the paranasal sinuses to estimate natural variations, SPIE Medical Imaging, 2016 A Sin inha ha, et al., Simultaneous segmentation and correspondence improvement using statistical modes, SPIE Medical Imaging, 2017

slide-12
SLIDE 12

Statistical shape models

Mean Variance iance

12

slide-13
SLIDE 13

Statistical shape models

Variance along the principal mode for the maxillary sinus

Front view Left view

13

slide-14
SLIDE 14

Statistical shape models

Variance along the principal mode for the middle turbinates

Front view Right view

14

slide-15
SLIDE 15

Shape estimation using statistical shape models (SSMs)

Mode we weights hts Estimat mated ed shape pe

๐–โˆ— = v1

โˆ—

v2

โˆ—

โ‹ฎ v๐‘œv

โˆ—

15

slide-16
SLIDE 16

Shape estimation using statistical shape models (SSMs)

Mode we weights hts Estimat mated ed shape pe

๐–โˆ— = v1

โˆ—

v2

โˆ—

โ‹ฎ v๐‘œv

โˆ—

What if correspondences are not available?

16

slide-17
SLIDE 17

Iterative closest point (ICP) algorithm

Find closest point match on ฮจ for all X Compute transformation to align matches y๐‘— โˆˆ ฮจ X = {x๐‘—}

17

slide-18
SLIDE 18

Iterative most likely point (IMLP) algorithm

Find most likely point match on ฮจ for all X Compute transformation to align matches

18

y๐‘— โˆˆ ฮจ X = {x๐‘—}

slide-19
SLIDE 19

Outline

19

  • Deformable iterative most likely point (D-IMLP)
  • Deformable iterative most likely oriented point (D-IMLOP)
  • Generalized deformable iterative most likely oriented point

(GD-IMLOP)

Background Deformable registration framework Results and confidence estimates

slide-20
SLIDE 20

Deformable iterative most likely point (D-IMLP) algorithm

Find most likely point match on ฮจ for all X Compute transformation to align matches

and deform

  • rm ๐›€ to fit to ๐˜

20

y๐‘— โˆˆ ฮจ X = {x๐‘—}

slide-21
SLIDE 21

Deformable iterative most likely point (D-IMLP) algorithm

y๐‘— โˆˆ ฮจ X = {x๐‘—}

21

๐‘”(yi, s)

slide-22
SLIDE 22

Deformable iterative most likely point (D-IMLP) algorithm

y๐‘— โˆˆ ฮจ X = {x๐‘—}

Find R, t and a such that x is best aligned with a deformed y Find s such that y deforms to fit x

22

๐‘”(yi, s)

slide-23
SLIDE 23

Deformable iterative most likely oriented point (D-IMLOP) algorithm

y๐‘— โˆˆ ฮจ X = {x๐‘—}

Find R, t and a such that x is best aligned with a deformed yโ€ฆ Find s such that y deforms to fit x and such that the normal of y aligns with that of x

23

๐‘”(yi, s),

slide-24
SLIDE 24

Generalized deformable iterative most likely

  • riented point (GD-IMLOP) algorithm

y๐‘— โˆˆ ฮจ X = {x๐‘—}

Find R, t and a such that x is best aligned with a deformed yโ€ฆ Find s such that y deforms to fit x and such that the normal of y aligns with that of x ,

24

๐‘”(yi, s),

slide-25
SLIDE 25

What is Tssm(y๐‘—)?

25

๐ฐ๐‘—

(2)

๐ฐ๐‘—

(3)

๐ฐ๐‘—

(1)

๐œˆ๐‘—

(1)

๐œˆ๐‘—

(2)

๐œˆ๐‘—

(3)

y๐‘—

slide-26
SLIDE 26

Deformable most likely point paradigm

26

slide-27
SLIDE 27

Outline

27

  • Results on simulation data
  • Confidence associated with results
  • Results on in-vivo clinical data

Background Deformable registration framework Results and confidence estimates

slide-28
SLIDE 28

Comparison betweenโ€ฆ

D-IMLP

  • Position
  • Anisotropic

Gaussian noise

28

y๐‘— โˆˆ ฮจ X = {x๐‘—} y๐‘— โˆˆ ฮจ X = {x๐‘—} y๐‘— โˆˆ ฮจ X = {x๐‘—}

D-IMLOP

  • Position
  • Anisotropic

Gaussian noise

  • Orientation
  • Isotropic Fisher

noise

GD-IMLOP

  • Position
  • Anisotropic

Gaussian noise

  • Orientation
  • Anisotropic Kent

noise

SSM estimate

  • Coherent Point

Drift (CPD)

  • Position
  • Isotropic

Gaussian noise Y = {y๐‘—} X = {x๐‘—}

slide-29
SLIDE 29

Error metrics

29

Total al shape pe error

  • r

(TSE) Total al registra stratio tion n error (TRE RE)

Ground truth shape Estimated shape Registered points

slide-30
SLIDE 30

Leave-one-out experiment

  • # sample point: 1000 (uniformly)
  • Translational offset: [0, 15] mm
  • Rotational offset: [0, 9] degrees
  • Noise:
  • 1 ร— 1 ร— 1mm3
  • 20ยฐ (๐‘“ = 0.5)
  • Noise assumed:
  • 1 ร— 1 ร— 1mm3
  • 20ยฐ (๐‘“ = 0.5)

30

Middle turbinate

slide-31
SLIDE 31

Leave-one-out experiment

Total al shape pe error

  • r

31

Total al registrati stration n error

slide-32
SLIDE 32

Leave-one-out experiment

32

Time

CPD CPD

Total al shape pe error

  • r
slide-33
SLIDE 33

Leave-one-out experiment

  • # sample point: 1000 (uniformly)
  • Translational offset: [0, 15] mm
  • Rotational offset: [0, 9] degrees
  • Noise:
  • 1 ร— 1 ร— 1mm3
  • 20ยฐ (๐‘“ = 0.5)
  • Noise assumed:
  • 1 ร— 1 ร— 1mm3
  • 20ยฐ (๐‘“ = 0.5)

33

Right nostril

slide-34
SLIDE 34

Leave-one-out experiment

34

Total al shape pe error

  • r

Total al registrati stration n error

slide-35
SLIDE 35

Parameter sweep

  • # sample point: 500 (uniformly)
  • Translational offset: [0, 15] mm
  • Rotational offset: [0, 9] degrees
  • Noise:
  • 2 ร— 2 ร— 4mm3
  • 10ยฐ (๐‘“ = 0.5)
  • Noise assumed:

35

Right nostril

Position Orientation 1x1x1mm3, 1x1x2mm3, 2x2x2mm3, 2x2x3mm3, 2x2x4mm3, 3x3x3mm3, 3x3x4mm3, 3x3x5mm3, 4x4x4mm3, 4x4x5mm3 2ยฐ, 10ยฐ, 20ยฐ

slide-36
SLIDE 36

Parameter sweep

36

Actual noise in samples: 2 ร— 2 ร— 4mm3, 10ยฐ (๐‘“ = 0.5)

Assumed ed Assumed ed Assumed ed

Assumed position noise: Isotropic

(Assumed) (Assumed) (Assumed)

slide-37
SLIDE 37

Parameter sweep

37

Actual noise in samples: 2 ร— 2 ร— 4mm3, 10ยฐ (๐‘“ = 0.5)

Assumed ed Assumed ed Assumed ed

Assumed position noise: Anisotropic

(Assumed) (Assumed) (Assumed)

slide-38
SLIDE 38

Leave-one-out experiment

  • # sample point: 3000 (nasal passage)
  • Translational offset: [0, 10] mm
  • Rotational offset: [0, 10] degrees
  • Noise:
  • 0.5 ร— 0.5 ร— 0.75mm3
  • 10ยฐ (๐‘“ = 0.5)
  • Noise assumed:
  • 1 ร— 1 ร— 2mm3
  • 30ยฐ (๐‘“ = 0.5)

Right nostril

38

slide-39
SLIDE 39

Leave-one-out experiment

39

Total al shape pe error

  • r
slide-40
SLIDE 40

Leave-one-out experiment

40

Total al shape pe error

  • r

Total al registrati stration n error

slide-41
SLIDE 41

Success classification: D-IMLP

41

โ‰ˆ ๐œ“2 =

Chi-squar square e value lue Probabil abilit ity densi sity ty ๐œ“2

๐‘ž = Pr[E๐‘ž < ๐œ“2] distribution

slide-42
SLIDE 42

Success classification: D-IMLP

42

โ‰ค =

Chi-squar square e value lue Probabil abilit ity densi sity ty

๐‘ž = Pr[E๐‘ž < ๐œ“2]

slide-43
SLIDE 43

Success classification: D-IMLOP

43

โ‰ค = โ‰ˆ ๐œ“2 =

distribution

slide-44
SLIDE 44

Success classification: D-IMLOP

44

โ‰ค = โ‰ค =

slide-45
SLIDE 45

Success classification: GD-IMLOP

45

โ‰ค = โ‰ˆ ๐œ“2 =

distribution

slide-46
SLIDE 46

Success classification: GD-IMLOP

46

โ‰ค = โ‰ค =

slide-47
SLIDE 47

Success classification (leave-out analysis)

  • # sample point: 3000 (nasal passage)
  • Translational offset: [0, 10] mm
  • Rotational offset: [0, 10] degrees
  • Noise:
  • 0.5 ร— 0.5 ร— 0.75mm3
  • 10ยฐ (๐‘“ = 0.5)
  • Noise assumed:
  • 1 ร— 1 ร— 2mm3
  • 30ยฐ (๐‘“ = 0.5)

Right nostril

47

slide-48
SLIDE 48

Success classification: D-IMLP

48

No orientation entation inform

  • rmation

ation

slide-49
SLIDE 49

Success classification: D-IMLOP

49

slide-50
SLIDE 50

Success classification: GD-IMLOP

50

slide-51
SLIDE 51

Success classification: GD-IMLOP

51

slide-52
SLIDE 52

Success classification: GD-IMLOP

52

๐‘ž = 0.95 (ve very confident) t)

TR TRE = 0.34 (ยฑ 0.03)mm mm

slide-53
SLIDE 53

Success classification: GD-IMLOP

53

๐‘ž = 0.9975 (confid ident) t)

TR TRE = 0.62 (ยฑ 0.03)mm mm

slide-54
SLIDE 54

Success classification: GD-IMLOP

54

๐‘ž = 0.9999 (some mewhat hat confident) t)

TR TRE = 0.78 (ยฑ 0.04)mm mm

slide-55
SLIDE 55

Success classification: GD-IMLOP

55

๐‘ž = 0.999999 (low confidence) ce)

TR TRE = 0.80 (ยฑ 0.05)mm mm

slide-56
SLIDE 56

Success classification: GD-IMLOP

56

remain inin ing (n (no

  • confidence)

e)

TR TRE = 1.31 (ยฑ 0.85)mm mm

slide-57
SLIDE 57
  • 5 clinical sequences
  • Noise assumed:
  • 1 ร— 1 ร— 2mm3
  • 30ยฐ (๐‘“ = 0.5)

In-vivo data: GD-IMLOP

57

Right nostril Dense reconstruction from video

slide-58
SLIDE 58

In-vivo data: GD-IMLOP

58

Residual error (mm) Max error (mm) Min error (mm)

All registrations

1.09 (ยฑ1.03) 4.74 0.50

Registrations that pass Ep test

0.76 (ยฑ0.14) 0.99 0.50

Registrations that pass Ep and Eo tests

0.78 (ยฑ0.07) 0.94 0.72

slide-59
SLIDE 59

Right nostril

Shape inference (leave-out analysis)

  • # sample point: 3000 (nasal passage)
  • Translational offset: [0, 10] mm
  • Rotational offset: [0, 10] degrees
  • Noise:
  • 0.5 ร— 0.5 ร— 0.75mm3
  • 10ยฐ (๐‘“ = 0.5)
  • Noise assumed:
  • 1 ร— 1 ร— 2mm3
  • 30ยฐ (๐‘“ = 0.5)

59

slide-60
SLIDE 60

Shape inference

60

Compute area within boundary

slide-61
SLIDE 61

Shape inference: GD-IMLOP

61

Mean cross sectional area of the external nasal valve in our sample: 112.47 ยฑ26.5 mm2

slide-62
SLIDE 62

Conclusion

  • Framework for deformable registration using statistical shape

models and features with uncertainty

  • 3 algorithms built within this framework using different features

and noise models

  • Method to assign confidence to the computed registration

62

slide-63
SLIDE 63

Conclusion

  • Surgical/clinical navigation without additional tools
  • Compensate for deformations in anatomy
  • Estimate anatomy in the absence of CT

63

slide-64
SLIDE 64

Future work

  • Larger dataset to build statistical shape models
  • More clinical evaluations
  • Explore additional features like contours
  • Automatically initialize registration
  • Explore applications in expression and pose

estimation

64

slide-65
SLIDE 65

Thank you!

  • Advisors
  • Russ Taylor
  • Greg Hager
  • Misha Kazhdan
  • Family

65

  • Mentors
  • Masaru Ishii
  • Austin Reiter
  • Undergrad advisors

and mentors

slide-66
SLIDE 66

Thank you!

66

slide-67
SLIDE 67

67

Questions?