VTree-3D - National Alliance for Medical Image Computing

System Challenges in Image Analysis
for Radiation Therapy
Stephen M. Pizer
Kenan Professor
Medical Image Display & Analysis Group
University of North Carolina
Co-authors: Edward L. Chaney,
Julian G. Rosenman
Credits to many others in UNC MIDAG
MIDAG@UNC
Objective: Segmentation in
Radiation Treatment Planning & Delivery
 Target
problems, in everyday RT
 Planning
radiotherapy
 Segmentation
of normal organs to be spared
New patients
 Kidney, liver, head and neck

 Segmentation
of regions implied by
segmented organs: lymph levels
 Adaptive
radiotherapy, incl. IGRT
 Segmentation
of organs to be spared and
target organ
Day to day changes within a patient
 Male pelvic organs: bladder, prostate, rectum

bladder, protsate,
rectum
MIDAG@UNC
Goal: Segmentation for
Radiation Treatment Planning & Delivery

Our objective: A whole
new level of
segmentation ability
 On tough images
 As good as humans
in most cases

The principle
Bladder
CTs
Prostate
 Make
use of probability distribution
of geometric variation: p(m)
 Make use of probability distribution of
geometry-relative intensity patterns p(I | m)
 Posterior optimization

arg max m p(m | I) =
arg max m [log p(I | m) + log p(m) ]
MIDAG@UNC
To achieve segmentation via p(m) and p(I | m):
Obtaining Training Data

Fitting m to training binary images



arg minm f(m | binary I) =
arg minm [image match penalty
+ geometric penalty]
Tight fit of geometric model to binary is critical
Extracting object-relative intensity patterns from
corresponding CT images

The tight fit of geometric model to binary makes
regional intensity patterns more informative
MIDAG@UNC
Representing m and I | m
Principle: representation should support PCA
 Representing object geometry m


 M-rep:
sheet of medial atoms
Captures local twisting, bending,
magnification of interior
 Unfamiliar to physicians

Representing image pattern relative to geometry
I
| m = = I relative to m
= RIQF(interior), RIQF(exterior)
RIQF: regional intensity quantile function
 Unfamiliar to physicians

MIDAG@UNC
Segmentation Program
 Initialize
pose of mean according to
image landmarks
 Conjugate gradient optimization of
log p(I | m) + log p(m)
over coefficients of 9 principal
geodesics of p(m)
 Objects
are thereby restricted to credible shapes
MIDAG@UNC
System Challenges
 Challenges
of doing the
research within a clinical setting
 Challenges of getting the
research results evaluated in a
clinical context
 Challenges of clinical adoption
of the research results
MIDAG@UNC
System Challenges in the Research
 Acquisition
of image data of adequate quality
 Meet
HIPAA regulations: anonymization
 Homogeneous, high resolution data sets
 Full volume of interest
 As artifact free as possible, or with
typical artifacts
 Training cases from target population (with
cancer) and from patients with normal anatomy
 Conversion
of images and segmentations
images and segmentations in RT planning
and delivery system into research system
MIDAG@UNC
System Challenges in the Research, cont.
 High
quality manual segmentations
in RT planning and delivery system
 Consistency
across training cases
 In
adaptive RT: consistency with MD’s
planning day
 Planning of RT: Include multi-expert variation
 Improved
manual segmentation
tools were developed
MIDAG@UNC
System Challenges in the Research, cont.
 Need
research-tolerant and interested
physicians on the team
 Need physician input all along, without too
heavily disappointing them with early results
 New
approach was expected to, and did, take a
decade to develop
 Obtaining a large number of
careful, manual segmentations
MIDAG@UNC
Evaluation experiments
 Retrospective
 Our
data
 Other hospital’s data
 Prospective
 Within
clinical practice, but not interfering with it
 “Jeopardy” that they might use the good results
clinically during the test
 Complete
their own segmentations first
MIDAG@UNC
System Challenges in the Evaluation
Access to software that is used clinically
 Adding objects with geometry, not just image
slice contours (in one orientation), to RT
system and its philosophy

 Providing
segmentations with clinically useful
measures of tolerance
Hiding the image analysis details from the
clinical user, while allowing access to them by
the image analysis researcher
 Software’s robustness, reproducibility, user
independence, speed (also for clinical use)

MIDAG@UNC
System Challenges in the Evaluation, cont.

Need consensus performance
standards
 No
gold standard with real
clinical material
Case index
Comparisons of computer vs.human
Cf. human-to-human prostate
differences against inter-human
agreement: 1.9mm average
or intra-human differences
surface distance
 Means of generating synthetic but
realistic cases with known truth
 Community-wide test case
collections

MIDAG@UNC
System Challenges of the Clinical Context
 Adding
indications of
non-credibility to
segmentations; they may
believe too readily
 Adding editing capability
to computer generated
segmentations
 Software continuing to
change, during clinical
tests and clinical use
MIDAG@UNC
Conclusion
 Research
on IGT methods has not
only technical and clinical challenges
but also significant software system
challenges
MIDAG@UNC
MIDAG@UNC