Predicting breathing motion for 4D radiotherapy

1

• Quantifying the properties and accuracy of a
  deformable image registration algorithm for 4D
  treatment planning
• Krishni Wijesooriya, PhD - University of
  Pittsburgh Medical Center Cancer Centers/
  Virginia Commonwealth University
• Paul Keall, PhD - Stanford University/ Virginia
  Commonwealth University
• Radhe Mohan, PhD UT M.D. Anderson Cancer Center
• Lei Dong, PhD UT M.D. Anderson Cancer Center
• Sarang Joshi, PhD - University of North Carolina
• Elizabeth Weiss, MD - Virginia Commonwealth
  University
• Vaughn Dill - Virginia Commonwealth University

• Rationale
• 4D planning requires automation due to the 10x
  increase in input CT data!
• Contour drawing has to be consistent
• Contour drawing has to be efficient

Fractional volume summary
Manual vs. auto contouring results
Axial, sagittal and coronal views from Pinnacle
7.7 Red contours are for the inhale phase (Phase
1) Color wash contours are for the manual exhale
phase (Phase 6) Auto contours from inhale to
exhale are Black GTV Yellow Cord, Heart Pink
Esophagus White -lungs

• Methods
• Thirteen patient 4D CT datasets acquired each
  data set consists of 10 CT image sets,
  corresponding to different respiratory phases
• Anatomy was delineated on 6 structures on the
  peak-inhale CT image GTV, Cord, Heart,
  Esophagus, Right Lung, Left Lung Converted in
  to mesh format using Pinnacle 7.7
• Determined anatomical transformation using a
  deformable image registration model a
  deformation vector field is obtained from the
  peakinhale CT image to each respiratory phase CT
  image

Exhale
Inhale

• GTV motion inhale vs. exhale

3-D creation of the volumes in meshes from IDL
Left view shows the volume shifts in 3-D from
inhale to exhale Right view shows the agreement
of manually drawn contours at exhale with
automated contours from inhale to exhale

• Fractional volumes are offset by 0.5 per each
  patient for clarity

COM displacement summary

GTV
Cord
Heart
Esophagus
Lungs

• Deformable image registration model used
• Large deformation diffeomorphic image warping UNC
  (S. Joshi et al)
• Christensen et al PMB 1994 39 209-618
• Christensen et al IEEE Trans Med Imag
  1997 16864-877
• Miller et al in Toga (Ed.) Brain
  Warping, Academic Press

• Practical guide to good auto contouring
• Image artifacts affect auto contouring ?
• (discrepancy in patient 1 lung)
• Need a data set with full thorax to better
  reproduce the diaphragm
• (discrepancy in many patients right lung)
• Deficiencies in manual contouring at minimum auto
  contouring
• could be used as a QA check ?
• (discrepancy is found in about 1.6 ) Heart
  drawn incorrectly
• Auto contouring less variation leads to
  predictability ?
• COM displacements are asymmetric wrt phase. Need
  more that two phases ?
• Cord (bony anatomy) is very well reproduced ?

• An Iterative optimization process
• Coordinate space of one (peak inhale) CT image is
  transformed to another CT image
• Accommodate large deformations by assuming a
  viscous fluid, h is determined by integrating
  velocity fields forward in time
• Velocity field is computed at each iteration with
  the help of force function and regularization
  operator, L
• CT number is assumed to be a constant
• Tissue deformation is corrected using a variation
  of the image dissimilarity, force function

Corresponding author address Krishni
Wijesooriya, UPMC West, Suite C, 1600 Coraopolis
Heights Road, Moon Township, PA 15108, USA
email wijesooriyak_at_upmc.edu
2

• Quantifying the properties and accuracy of a
  deformable image registration algorithm for 4D
  treatment planning
• Krishni Wijesooriya, PhD - University of
  Pittsburgh Medical Center Cancer Centers/
  Virginia Commonwealth University
• Paul Keall, PhD - Stanford University/ Virginia
  Commonwealth University
• Radhe Mohan, PhD UT M.D. Anderson Cancer Center
• Lei Dong, PhD UT M.D. Anderson Cancer Center
• Sarang Joshi, PhD - University of North Carolina
• Elizabeth Weiss, MD - Virginia Commonwealth
  University
• Vaughn Dill - Virginia Commonwealth University

Surface congruence analysis
Surface congruence analysis difficulties
Pinnacle 7.7 meshes
IDL software
Surface Congruence analysis summary
Read in triangles and vertices
Convert to spherical polar coordinates COM of
manual volume as center
Points interpolated to a non-equi-spaced grid
of Q and F, quadratic interpolation
Per each grid point,Q and F, R_diff r_manual
r_automatic

• Surface congruence results

Multiple surfaces, not single valued in r for a
given q , j bin
Multiple sub volumes

Plot error bars are STDEV
Sphere
Cylinder
Systematic uncertainty (0.25 mm) of the full
procedure was determined using a perfect sphere
of known diameter
GTV

• A first step towards 4-D treatment planning, auto
  contouring is
• established with Pinnacle 7.7 treatment planning
  system.
• A comparison toolkit was developed using IDL
  which includes
• both visualization and quantification tools
• This system has been tested with a deformable
  image registration
• program with 13 patient 4-D CT datasets
• GTV is very well reproduced In auto contouring
• Auto-contouring shows less variability than
  manual contouring
• At a minimum, auto contouring yields
  discrepancies in manual
• contouring and could be a useful QA tool
• Image artifacts and incompletely scanned organs
  challenge
• auto-contouring
• Auto-contouring based on deformable image
  registration
• is efficient, reliable and consistent, and can
  be used
• (with caution!) for 4D treatment planning

Summary
Automatic
Manual
Cord
Manual
Automatic
Plot error bars are statistical
Heart
Manual
Automatic
Corresponding author address Krishni
Wijesooriya, UPMC West, Suite C, 1600 Coraopolis
Heights Road, Moon Township, PA 15108, USA
email wijesooriyak_at_upmc.edu