Medical Image Analysis and the NCRI Demonstrator Project

1
Medical Image Analysis and the NCRI Demonstrator
Project
Professor Sir Michael Brady FRS FREng Department
of Engineering Science Oxford University
2
We can watch the body functioning in a whole
range of ways the brain thinking, degradation
in white matter, and the pulsing of the heart
Over the past 20 years, we have developed new
ways to image anatomy, new ways to see inside the
body, non-invasively
We can watch the body in action, as it responds
to the injection of a drug or contrast agent, to
highlight aberrant physiology
Now we are beginning to image cellular and
molecular processes the convergence of
molecular biology and image analysis
3
Trends in imaging image analysis

• Continual improvements in imaging
• parallel MRI
• high resolution PET (clinical pre-clinical)
• Bioluminescence, FMT
• 3D ultrasound
• Image fusion

4
The left column shows the image pre-treatment,
the centre column indicates initial response to
treatment, while the right shows that there has
been regression
5
Why more than one image?

• Often, one image does not suffice
• enormous variability of anatomy/physiology often
  make diagnosis patient management decisions
  difficult
• image quality and variations in appearance
• use a previous image as basis for comparison with
  a new one
• Different kinds of image give complementary
  information

6
Cancer has many facets
No one imaging method gives information about all
of these facets, and each person is different ?
fuse the results of several tests into a common
framework!!
7
Case history necessitating image fusion
CC View
MLOView
Mammograms (columns) of same woman, aged 23, 24,
25, and 26 Age 26, 3 clusters of
microcalcifications detected ? an MRI taken
8
X-ray mammography MRI data fusion
Project the 3D MRI data to appear like a mammogram
Project it in the CC and in the MLO directions
9
The subtle enhancement is characteristic of DCIS,
confirmed by biopsy
10
Trends in biomedical image analysis

• Continual improvements in imaging
• parallel MRI
• high resolution PET
• Optical and luminescence
• 3D ultrasound
• Image fusion
• Quantitative analysis for disease staging and
  progression

11
QuantiSPECT

• software wrapper for a specific drug
• Measures caudate, putamen, and striatum
• Database of normals patients at different
  stages of disease progression
• FindOneLikeIt

12
Simultaneous Segmentation Registration
Transformation parameters
Segmentation performance index
13
Parametric T1 mapping for analysing ce-MRI
Conventional analysis based on intensity change
Estimating change in T1 and its visualisation
Multiple acquisitions prior to injection of Gd is
well-known. We have developed a method that
minimises the error in the estimation of T1
Armitage, Behrenbruch Brady, Medical Image
Analysis, 2005 Ketsetzis and Brady, IEEE Trans.
Med. Im., 2005
14
Measuring effect of chemotherapy
Pre- and post-chemotherapy Percentage increase in
intensity at right
(non-rigid) registration and pre- and
post-chemotherapy, from ?T1
Armitage, Brady and Behrenbruch, Medical Image
Analysis (2005)
15
Simultaneous segmentation and Registration of
ceMRI
Results for four patients Left the pre-contrast
10 degree image Right the segmentations Blue
fat Green normal Orange benign lesion Red
brown malignant lesion
Compare to hand segmentation by a pair of
experienced radiologists Limitation of that
validation is that they disregard the partial
volume effect where much of the change occurs
Probabilistic labelling of dataset from ceMRI
16
Monitoring the uptake of a drug
We are beginning to image and model the dynamics
of drug activity and relate these to cellular
and molecular processes
Monitoring unacceptable build-up in organs other
than that of primary interest
17
Modelling the take up of candidate drugs
Segmentation tool
Organs of interest segmented
Miradas Research MVS
18
Model-based segmentation
PET noise model (inside and outside tumour) PK
model of 18FDG take up
Level set method (optimisation to solve
regularised differential equation)
(hexokinase)
(Glucose transporter)
Catherine White and Michael Brady, Soc. Nuc.
Med., Toronto, June 2005
19
Generation of a realistic simulation of dynamic
PET brain data
MNI probabilistic brain atlas
FDG model
GM, WM time activity curves
SORTEO simulator
Schottlander, Brady, et. al. SPIE 2005
(forthcoming)
Fusion of dynamic MR dynamic PET ? parametric
(PK) maps cancer typing
Ecat Exact HR
20
Image analysis drug development
For every 10,000 lead compounds, one makes it to
a product. It costs 1Bn to develop a drug.
Every month of delay costs 1.5M. Conventional
drug development is broken. Can image analysis
help? First in human.
?
?
21
Imaging transgene expression noninvasively

• can distinguish Renilla luciferase and firefly
  luciferase
• Can monitor tumour growth/response and metastasis
  in real time
• Can measure virus infectivity

22
NCRI colorectal cancer demonstrator

• Linking (macro) pathology images to MRI
• Information integration
• Multidimensional Team (MDT) Suite
• Internet and Grid-based delivery

23
Colon Cancer patient journeys
Patient discomfort, bleeding,
reassurance
GP consultation
concern
Endoscopy in surgical OP dept
Suspicion of cancer
Surgery generally no intraop images
CT (3D xray) image
Palliative care
Resected mesorectum
Chemo/radio therapy
Macroimage slices
Microscopic images
24
MRI image enhancement
Left image shows the substantial bias field,
mostly corrected in the right hand image second
release of algorithm currently finishing
development
25
Active Contours
Using an active contour to track through the
series of axial images
26
3D Visualisation of Colorectum
27
Segmenting the Mesorectal Fascia

• Use easy to find shapes in order to create a
  coordinate frame or reference.
• Make initial estimate as to position of
  Mesorectal fascia
• Refine estimate using active shape models (Staib
  and Duncan), and training set of images (bias
  removed). Optimise using gradient descent
  algorithm.

Shape of Mesorectal fascia is described using
spherical harmonics.
28
Mesorectal fascia (lace) and colorectum (solid)
29
Finding nodes/blood vessels
The circles around the possible nodes are
coloured according to their probability Blue
Unlikely to be a related node Red Likely to be
a related node
30
Registering images pre and post treatment

• Clinicians are keenly aware of the effects that
  down-staging chemo/radiotherapy is likely to
  have. We would like to quantify these effects.
• Expected changes in the images are
• Reduction in tumour size and adenosis
• Reduction in lymph node size
• Change in shape of mesorectum and colorectum
• This is currently an unsolved problem that we are
  working on. It requires mobilisation of
  knowledge of anatomy and expected physiological
  changes in order to guarantee convergence of
  non-rigid registration to correct transform

31
A key challenge for information integration
MRI images
Resected specimen
(stained) microscopic slices
Macro slices
32
Histology aligned with MRI
Histology/MR fusion is difficult because (a)
specimens deform during fixing and slicing (b)
specimen slices may tear (below). Reconstruction
of a 3D volume from macro slices is in general
unsolved. Initially, some user intervention will
be required.
33
Arrows indicate (user specified) corresponding
points on the circumferential resection margin
(CRM) of the mesorectum for the MRI (left) and
path slide (right). Here, margins are clear.
MERCURY project (Pelican foundation)
Courtesy Dr. Gina Brown, Royal Marsden Hospital,
Prof Phil Quirk, Digital Pathology, Leeds
University
34
Arrows indicate the corresponding points on the
margins of a lymph node, showing that there is
insufficient clearance at the Circumferential
Resection Margin, significantly increasing the
possibility of local reoccurrence of cancer
(typically pelvis or liver), with very poor
prognosis.
428 cases of high resolution MRI
pathology Available for research
Courtesy Dr. Gina Brown, Royal Marsden Hospital,
Prof Phil Quirk, Digital Pathology, Leeds
University
35
Aligning pathology MRI
Once we have the transformation from the
reconstructed volume to the MRI, we can reslice
the MRI volume to extract the MRI data that
correspond to each particular macro slice.
Set of macro slices
Phil Quirke, Mike Brady, Gina Brown, David
Gavaghan, Andy Simpson NCRI supported project
36
MultiDimensional Team meeting

• Radiologists pathologist surgeon ( others)
  discuss patient management options in depth
  2-3 minutes
• Images provide key information
• Dialogue of the hard of hearing
• Surgical plan based on pre-operative (MRI)
  images, but generally performed without benefit
  of intra-operative images

37
Software support for the MDT

• Information integration
• Image analysis, image fusion (and signals)
• Efficient access to that information
• Integration with patient information
• Integration with genetics information (eg
  microarray)
• More informed decision making
• Remembering what was decided, and why

• More information patient management decisions
• Deepening appreciation by pathologists and
  radiologists of each others images
• Closing the loop for the surgeon
• circumferential resection margin (CRM)
  for successful outcome
• Basis for a teaching tool for the MDM

38
Comments on IT infrastructure

• Building in knowledge ( ontologies)
• Image formation
• Disease processes
• Tumour DNA and image signs
• Huge range of spatial and temporal scales
• Grid computing
• Federated databases (eDiaMoND)
• Dynamic atlas (iXi)
• Patient self-management (eSan)
• Image based CROs