professor sir michael brady frs freng department of ...jmb/lectures/medimanallecture2.pdf ·...
TRANSCRIPT
Clinical applications of MRIProfessor Sir Michael Brady FRS FREng
Department of Engineering ScienceOxford University
Michaelmas 2004
MRI is used for a huge range of clinical applications
• Clinical neurology– Segmentation and classification– Measuring volumes of brain structures– Multiple sclerosis, neurodegeneracy, stroke, …
• Cardiology– Either need to image fast, or deal with heart motion!
• Cancer– Breast, colorectal, liver, prostate, …
• Soft tissue damage – Cartilage, ligaments, …
MRI is also used a great deal in basic science to study brain function and cancer growth
Segmentation and measurement of brain tissueNote the huge overlap between the Gaussian pdf for GM and that for WM
This means that there are many misclassifications of GM pixels as WM and vice versa.
Even small amounts of noise can change the classification.
The segmentation / classification of brain tissue depends upon estimating the likelihood of each class at each voxel x and then iteratively updating and propagating these estimates to the neighbours
This uses Hidden Markov Random Fieldsand will be defined in the Informatics course
MR image of a horizontal slice through the brain.
In this T1-weighted image, grey matter is lightly coloured, while white matter appears darker.
Two spots of plaque, corresponding to multiple sclerosis are visible.
MR angiography – the vasculature of the brain
MR angiography is rapidly gaining in importance; it can be fused with the complementary CT angiography
Diffusion Imaging
• Measure self diffusion of protons in every voxel.• In white matter areas – more diffusion in fibre direction.
We can measure e.g. principle fibre direction and “anisotropy” (Strength of fibre direction).
Algorithms have been developed to trace out fibre tracts – connectivity in the brain
Application of DWI Analysis:Thalamus Segmentation
Seeded in medio-dorsal nucleus in thalamus; in monkey this projects to prefrontal cortex and receives projections from temporal lobe
Work done in the Functional MRI of the Brain Laboratory by Tim Behrens, Dr. Johansson-Berg, Mike Brady, Paul Matthews, Steve Smith, Mark Jenkinson, and Mark Woolrich
Motor pathways
This example shows a tract from VL nucleus going to M1, cerebellum and brainstem
Visual pathways
This example shows a tract from LGN going to optic tract and visual cortex
Segmentation of left and right thalami, based on projections to 4 cortical zones.
Purple: MediodorsalnucleusProjects to PFC, receives from temporal lobeBlue: Ventral posterior nucleus, projects to S1/S2Orange: Ventral lateral and ventral anterior nuclei. Project to M1 and PMC/SMAYellow: Pulvinar, projects to PPC and extrastriate cortex.
Arterial Spin Labelling
This method enables clinicians to estimate cerebral blood flow, in order to manage stroke patients
Source: Daniel Gallichan, FMRIB, May 18 2004
TAGGED MRI DATA (1996)
Normal heartNormal heart
Abnormal heartAbnormal heartinferior infarctinferior infarct
Short Axis 1Short Axis 1 Short Axis 2Short Axis 2 Long AxisLong Axis
TAGGED MRI MOTION RECONSTRUCTION
Normal heartNormal heart
Short Axis 1Short Axis 1tag planestag planes
andandLV boundariesLV boundarieswith SA imageswith SA images
Source: Jérôme Declerck, PhD thesis, 1997
Heartbeat 2.0T cardiac MR, taken by Professor Stefan Neubauer, JR hospital, June 2003
Images courtesy of subject: Mike Brady
MRI and cancer
• Breast cancer– Contrast enhanced MRI and angiogenesis– T1 estimation
• Colorectal cancer– Bias field correction– Anatomical frame of reference– Lymph node detection & classification– Pre- and post-chemotherapy image registration
Limitation of anatomical MR imaging
1. Take one or more images prior to injection of contrast agent
2. Inject contrast agent, typically Gd-DTPA
3. Take images as fast as possible afterwards (note: all of both breasts imaged)
The tumour cannot be differentiated by its T1; conventional MRI gives information about anatomy not physiology
There is a massive adenocarcinoma there –with secondary spread
Contrast agent take-up
Inside the tumour, the enhancement is over 100%
Contrast agent take-up
.. Normal tissue enhances less
Contrast agent take-up
.. Whereas fat barely enhances at all
Unfortunately, some benign tissue can enhance more than malignant, and the amount of uptake is highly variable, making quantitative analysis difficult
Reason: contrast agent take-up is non-linearly related to intensity change
0 1 2 3 4 5 6 7 8 9 10
FAT
NORMAL
BENIGN
MALIGNANT
0
0.2
0.4
0.6
0.8
1
1.2
TIME (mins)
Sign
al E
nhan
cem
ent
Benign & Malignant in CEMRI
No absolute scale for ∆S increase for malignant
Benign processes such as fibroadenoma can enhance as much as tumour
T1 mapping
• Signal enhancement is non-linearly related to Gd concentration, which is directly proportional to local vascularity
• Working from a model of signal enhancement, we have developed a method to estimate the T1 at each time point, hence the change in T1
0 1 2 3 4 5 6 7 8 9 100
0.2
0.4
0.6
0.8
1
1.2
TIME (mins)
SIG
NA
L E
NH
AN
CE
ME
NT
FAT
NORMAL
BENIGN
MALIGNANT
Contrast Agent Uptake Profiles
0 1 2 3 4 5 6 7 8 9 100
0.2
0.4
0.6
0.8
1
1.2
TIME (mins)
CO
NC
EN
TRA
TIO
N (m
M)
FAT
NORMAL
BENIGN
MALIGNANT
• Malignant to benign distinction is massively improved using concentration based analysis.
Gd
Con
cent
ratio
n
Sign
al E
nhan
cem
ent
Gradient Echo Signal Model• Use Bloch equation to describe signal for a gradient echo pulse
sequence (for example)
• Add effects of contrast agent (T1 & T2 alteration).
1
1*2
/
//
cos11
sinTTR
TTRTTE
ee
egS−
−−
−
−=
ααρ
( )⎟⎠⎞⎜
⎝⎛ +−
⎟⎠⎞⎜
⎝⎛ +−
⎟⎟⎠
⎞⎜⎜⎝
⎛+−
−
−=
tT
tTtT
CRTR
CRTRCRTE
t
e
eegCS
111
111
2*2
1
cos1
1sin
α
αρ
Signal Enhancement vs. Concentration
( ) ( )( )
⎟⎟⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜⎜⎜
⎝
⎛
⎟⎟
⎠
⎞
⎜⎜
⎝
⎛−−−
⎟⎟
⎠
⎞
⎜⎜
⎝
⎛−−−
==⎟⎠⎞⎜
⎝⎛ +−⎟
⎠⎞⎜
⎝⎛ +−−
⎟⎠⎞⎜
⎝⎛ +−−⎟
⎠⎞⎜
⎝⎛ +−
−
tTtT
tTtT
t
CRTRCRTRTTR
CRTRTTRCRTR
CTERtt
eee
eee
eS
CSCE
112
111
1
112
111
1
2
cos1
cos1
0/
/
α
α
0.0 0 .1 0 .2 0 .30 .0
0 .2
0 .4
0 .6
0 .8
T 1 = 200 m s
T 1 = 500 m s
T 1 = 800 m s
T 1 = 1200 m s
Sign
al E
nhan
cem
ent
C oncen tration (m M )
T1 must be measured
Nonlinear variation with T1
Optimum Flip Angle Combinationsn = 2 - 5
16 ms17 °16 °10 °4 °3 °
18 ms15 °10 °4 °3 °
21 ms17 °10 °3 °
29 ms10 °3°
ET1α5α4α3α2α1
Armitage, Behrenbruch & Brady 2001, 2002
Two-Compartment Pharmacokinetic Model
• Models interaction between a blood pool and lesion leakage space (EES).
Blood PlasmaCp(t)
ExtravascularExtracellular Space
Ce(t)
Gd-DTPA
kpe
kout
C(0)
kep Whole BodyExtracellular
SpaceCx(t)
Differential Equations• Conservation of mass leads to :
• Solution gives the following:
( ) inppoutpeeeepp
p MCVkkCVkdt
dCV ++−=
eeeppppee
e CVkCVkdt
dCV −=
( ) ( ) ),,( baAfeeba
AtC atbt
e =−−
= −−
Tumour vascularity & angiogenesis
In order to fuel its growth, a tumour grows a network of millions of micro-vessels that tap in (like shunts) into supply arteries
Millions of micro blood vessels are grown – this new blood is angiogenesis
The blood vessels are both small (microns) and leaky, so Gd molecules
stick around the vicinity of a tumour
Parametric T1 mapping for analysing ce-MRI
Conventional analysis based on intensity change
Estimating change in T1 and its visualisation
Armitage, Behrenbruch & Brady, Medical Image Analysis, 2005
Ketsetzis and Brady, IEEE Trans. Med. Im., 2005
Multiple acquisitions prior to injection of Gd is well-known. We have developed a method that minimises the error in the estimation of T1
Measuring effect of chemotherapy
Pre- and post-chemotherapyPercentage increase in intensity at right
Pre- and post-chemotherapy ∆T1 at left
Armitage, Brady and Behrenbruch, Medical Image Analysis (2005)
(non-rigid) registration and pre- and post-chemotherapy, from ∆T1
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
Simultaneous segmentation and Registration of ceMRI
Probabilistic labelling of dataset from ceMRI
The need for non-rigid registration.
In this case, relaxation of the pectoral muscles causes severely non-rigid motion...
Registration Results.
The computed non-rigid motion field.
registration and image re-sampling
pre-contrast image post-contrast image subtraction image
motion field corrected post-contrast image corrected subtraction
Axial T2 MR images
Colorectal MRI – as used at Churchill Hospital
The brightness of the data is unbalanced due to RF field inhomogeneities, creating a bias field effect. This makes the images difficult to analyse.
The bias field (B1 inhomogeneity)• a low frequency distortion of the “idealised” intensities in an MRI image
• effect of these bias fields is highly disruptive when using surface coils, affecting intensities with a contribution of up to 60% of the maximum image intensity
• Acquisitions using body coils, such as the ones commonly performed for brain studies, create more uniform magnetic fields, thus reducing intensity variations to the range of 10%-20% of image amplitudes
Estimating the Bias FieldThe brightness of the data is unbalanced due to RF field inhomogeneities, creating a bias field effect. This makes the images difficult to analyse.
( ) ( ) * ( ) ( )s x o x b x n x= +
The bias can be modelled as a multiplicative field:
s(x) is the signal that is received and is seen as the resultant image;
o(x) is the original or ‘ideal’ image that we are trying to extract;
b(x) is the bias field;
n(x) is noise, which is initially assumed negligible.
It can then be converted to logarithmic form:
log( ( )) log( ( )) log( ( ))s x o x b x= +
Legendre polynomials and an ‘Evolutionary Strategy’ optimisation method (Styner et al.) are used to estimate b(x):
Removed Bias Field
Original image Bias field removed by Sarah Bond & Michael Brady
Active Contours
Using an active contour to track through the series of axial images
3D Visualisation of Colorectum
Fusion of CT and MRI requires constrained non-rigid transformation.
Segmenting the Mesorectal Fascia1. Use ‘easy to find’ shapes in order to
create a coordinate frame or reference.
2. Make initial estimate as to position of Mesorectal fascia
3. 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.
Mesorectal fascia (lace) and colorectum(solid)