Lines of Curvature for Polyp Detection in Virtual Colonoscopy

1
Lines of Curvature for Polyp Detectionin Virtual
Colonoscopy

• Presenter Lingxiao Zhao
• Coauthors Charl P. Botha, Javier O. Bescos,
• Roel Truyen, Frans M. Vos and Frits H.
  Post

2
Contents

• Introduction
• Visualization-based polyp characterization
• Improved automatic polyp detection and
  visualization using lines of curvature
• Generating lines of curvature on colonic surfaces
• Feature calculation for polyp candidate selection
• Evaluations
• Conclusions and future work

3
Introduction

• Colonic polyps are an important precursor of
  colon cancer
• Early detection and removal of polyps
  significantly decrease the incidence of colon
  cancer

4
Introduction

• Virtual colonoscopy (VC)
• Causes much less discomfort to the patient
• CT or MRI scans are processed using iso-surface
  extraction
• However, laborious visual inspection makes VC
  unattractive for large-scale screening

5
Introduction

• Computer-aided diagnosis (CAD) widely uses
  curvature
• Curvature definition a measure of surface
  curvedness
• Curvature representations principal curvature
  scalars and direction vectors, etc

Parabolic
N
T
Elliptic
T1
Hyperbolic
T2
6
Introduction

• Existing automatic polyp detection tech.
• Most are scalar curvature based
• Indicators, e.g. volumetric shape index,
  sphericity index
• Can result in a large number of false-positive
  detections
• Not all available information is used, e.g.
  principal curvature direction fields

7
Visualization-based polypcharacterization

• Potential of principal curvature direction fields
  for polyp characterization

Maximum curvature direction
Minimum curvature direction
8
Visualization-based polypcharacterization

• Characteristic patterns in principal curvature
  direction fields

• A generalization of polyp surface
• Polyp cap the most protruding part, which is
  ideally represented as a spherical or ellipsoidal
  surface
• Polyp neck the transition area from the
  background to the cap, which is typically a
  closed-ring and anticlastic surface

9
Visualization-based polypcharacterization

• Our hypothesis
• Principal curvature direction vector fields can
  be visualized by lines of curvature or curvature
  streamlines
• Specific patterns in lines of curvature, i.e.
  approximately closed streamlines, can help to
  discriminate polyps from background
• What we did is

10
Improved CAD using lines of curvature

• Procedure
• Generate curvature streamlines on the colon wall
• Curvature estimation
• Streamline tracing
• Streamline seeding and spacing
• Calculate features on curvature streamlines
• Discriminate between true- and false-positive
  detections

11
Improved CAD using lines of curvature

• Apply for the explicit surface representation
• The Normal Vector Voting method for calculating
  principal curvature values and directions on
  triangular meshes
• Improved Girshick et al.s method
• for 3D line drawings on triangular
• meshes
• Alliez et al.s curvature-adaptive
• seeding and spacing of streamlines
• on triangular meshes

12
Improved CAD using lines of curvature

• Apply for the implicit surface representation
• Van Vliet et al.s method based on the Hessian
  matrix for implicit iso-surface curvature
  estimation
• where
• Principal curvature values are calculated based
  on eigen values of the rotated H, and
  corresponding eigen vectors are principal
  directions in the local tangent plane

13
Improved CAD using lines of curvature

• Curvature streamline tracing on the implicit
  iso-surface
• A streamline is traced using stepwise integration
  in forward and reverse directions
• In each stroke, the streamline point is projected
  back onto the iso-surface
• Integration step length is adaptive
• to the local curvature magnitude

14
Improved CAD using lines of curvature

• Curvature-adaptive streamline seeding spacing
  on the implicit iso-surface
• Make curvature streamlines better fit surface
  features
• On highly curved surface area, more streamlines
  are generated
• Use principal curvature magnitude to determine
  the spacing distance between curvature streamlines

ds
15
Improved CAD using lines of curvature

• Comparison

16
Improved CAD using lines of curvature

• Calculate features on curvature streamlines
• Seek approximately closed streamlines as an
  indicator of the polyp neck
• Use a visualization-based analysis of the polyp
  neck, as an addition to the polyp cap-based
  techniques, to better characterize the polyp
  surface

17
Improved CAD using lines of curvature

• Selecting streamlines on the polyp neck
• The polyp neck is typically an anticlastic
  surface, where principal curvatures have
  different signs
• A derived quantity, namely hyperbolic percentage
  HP
• The first M (3, 5) streamlines with the largest
  nh are selected and each must have a HPgt50

18
Improved CAD using lines of curvature

• Employing fluid flow visualization techniques
• The winding angle calculation a characteristic
  curvature streamline of the polyp should have a
  winding angle of at least 2xPI
• The mean radius of the characteristic streamline
  polyps larger than 5mm in diameter are
  significant for clinical diagnosis

19
Improved CAD using lines of curvature

• Polyp candidate selection study
• We are planning to integrate our streamline
  selection and streamline-based feature
  calculation in an improved polyp detection
  protocol
• Stage 1 a polyp pre-detection phase
• Stage 2 reduce the number of false-positive
  candidates
• We assessed our methods in a large number of
  candidate areas
• 5 CT scans with 331 pre-detected polyp
  candidates, of which 50 sites (15.1) were
  classified as true polyps

20
Improved CAD using lines of curvature

• Statistical analysis
• Wilcoxon rank sum test Plt0.001
• Polyp candidate clustering

Median WA IQR
True positive 7.817 6.770-9.288
False positive 2.954 1.995-3.749
21
Improved CAD using lines of curvature

• Enhanced visualization of the shaded colon wall

22
Conclusions

• Our work explored the potential of principal
  curvature directions to characterize specific
  surface shape features, e.g. colonic polyps
• We adapted existing techniques for generating and
  rendering lines of curvature on explicit
  triangular surface meshes and developed a new
  scheme on implicit iso-surfaces
• We defined and assessed new streamline features
  that are highly correlated with true-positive
  polyp detections
• We proposed a streamline selection scheme based
  on hyperbolic points to guide the calculation of
  features

23
Future work

• Improve curvature streamline tracing, seeding and
  spacing strategies
• The streamline-enhanced visualization must be
  compared with the standard visualization in a
  clinical environment
• Most important, our polyp candidate selection
  scheme will be further tested with a large number
  of clinical CT scans
• Develop and evaluate additional streamline
  features for polyp detection

24
Acknowledgements

• This work is supported in part by Philips Medical
  Systems Netherlands BV, Best.
• Special thanks to Dr. Frans A. Gerritsen and Kees
  van Wijk

25
Finally

• Questions?