Dushyant Kumar (PhD)

1,2
,
Prof. Dr. Jens Fiehler
1
, Prof. Dr. Heesen Christoph
3
, Dr. Susanne
Siemonsen
1,2
, Dr. Rer. Nat. Jan Sedlacik
1
,






1
Neuroradiology, Universittsklinikum Hamburg-Eppendorf (UKE),
2
Multiple Sclerosis Imaging Section (SeMSI), (UKE)
3
Institute for Neuroimmunology and Clinical MS Research, UKE,
Deutschland

Feasibility of T2 relaxometry as a clinical and
pre-clinical investigative tool in demyelinating
neurological diseases

Declaration of Relevant Financial Interests or
Relationships
Dushyant Kumar & Other Authors

We have no relevant financial interest or relationship to disclose with
regard to the subject matter of this presentation.
Outline of talk
What is the T2 relaxometry (T2R)
Importance of myelin imaging
MRI sequences for collecting T2R data
Quantitative myelin imaging using T2 relaxometry (Data Processing)
Problems with T2-relaxometry
Conventional approach
Our multi-voxel spatial regularization approach
Improvement shown using simulated data
Implementation using CPMG based sequence
3D multi echo Spin Echo (MESE) My past work
2D fast spin echo (FSE) implemented at human and animal scanner, UKE
Improvement shown using MR data
Implementation using T2 prep spiral sequence My past work
Applications
What is the T2-relaxometry
Collect T2-W image at different TEs
from two liquid phantoms:
Individually: mono-exponential
T2 decay
Both together: bi-exponential (3
unknowns: 2 T2s value, 1
volume fractions)
Non-linear model
Principle of T2-relaxometry:
Water molecules in different
microscopic environment have
different T2 Values.
In case of animal tissue, there are
more variability and
multiexponential behavior is more
suitable.


Introduction to Quantitative T2 with emphasis on Medical Imaging
ISBN 978-0-557-67403-9; Editor: Thorarin A Bjarnason
Two bottles
of liquid with
different T2
decay rates
Faster Decay
T2-W Imaging and T2 Relaxometry
Courtesy: Thanh Nyugun
T2 W Image from a multiple sclerosis (MS):
T2 distribution from selected ROI:
Free Water
Myelin
Intra/extra
cellular water
T2 Relaxometry using Multi Echo Spin Echo
Forward Model: y = Ax + c, (multi-exponentials)
A
ki
= exp(-TE
k
/T
2
(i)); x are volume fractions of sub-pools
Inverting the data:
o 5 ms to 600 ms : 50-100 discrete points on log-scale.
o What are the relative signal strength at those T2-pts
(Histogram)??
o Problem is undetermined i.e. 32 data point and
50-100 unknowns.
Many solutions: How to choose the right one?
Data Acquisition
CPMG Based Sequence
Multi Echo Spin Echo (MESE)
Fast Spin Echo (FSE)
T2-prep based method:







10
-3
10
-2
10
-1
10
0
0
20
40
60
80
100
120
140
160
180
200
Myelin
Peak
Intra-/Extra
Cellular water
Typical T2- Distribution
0 50 100 150
0
200
400
600
800
1000
1200
1400
1600
1800
2000
TE -->
E
c
h
o

A
m
p
l
i
t
u
d
e
Typical T2- Decay
Data Processing:

Forward Model (single voxel) : y = Ax + c, with A
ki
= exp(-TE
k
/T
2
(i));
collect single-voxel quantities x, y into multi-voxel vectors , and
similarly define the expanded matrix A
ex
(block diagonal matrix)
Minimization problem:

Step 1: Set
S
= 0 & solve conventional
regul. Problem:


for all choice of
Choose optimum
T
for each voxel
Approach by Whittall and MacKay
L-curve approach: Corner of L-curve a good compromise between the
data fidelity and the constraint in the solution.
Step 2: Implement spatial constraints to improve noise robustness.

| |
2 5
T
.,10 ,......... 10

e
L Curve
Step 2: Spatial Regularization
Step 2: Include the spatial term and
solve:

Our spatial filter encodes the expectation
that
T2 relaxation times and volume
fractions of underlying pools must
change smoothly with in coherent
brain regions.
Notice: No tissue edge information
has been encoded externally.
Since we use quite a weak and
local prior, the tissue edge is
preserved in our result.


D
s
: first-order finite
difference operator
W: a diagonal weighting matrix
choosing spatial regularization parameters

S
assumed to be same for all slices

Where = median(voxel-wise
T
)

A supervised trial and error strategy:
Reconstruct a selected ROI for each
Our criteria for optimal are:
a) visual inspection of smoothness without excessive blurring,
b) low mean fitting error (MFE) and
c) R, ratio of contributions from spatial to temporal terms,
close to 1.


opt
T
opt
S
=
| | 10000 3000, 1000, 300, 100, 50, 10, 1, e
opt
T

MWF maps at different spatial regul.

Improvement: MWF from Simulation
Simulated Data ( 50 x 50 voxels):
Surrounding matrix consisting of two pools
Lesions of various sizes (smallest lesion:
single voxel wide) are simulated as single
pool with no myelin.
The improvement due to proposed alg.:
Reduced spatial variation of MWF
improved MWF accuracy,
better lesion detection
Even the visibility of smallest lesion (1 voxel
wide) is improved due to reduced spatial
variation in matrix.
Symmetric Kullbeck-Leibler (SKL) score
between true dist and solution:
Lower SKL score for proposed algorithm at
various SNR levels between 50 and 500.
Lower SKL score means returned
distributions are similar to true distribution.



A) MWF maps
B) Averaged SKL score
C) Mean square error in MWF
from simulated data at different SNR
Definition of Symmetric Kullbeck-Leibler Score

|
|
.
|

\
|
(

+
(

=
i
div
i f
i f
i f
i f
i f
i f f f SKL
) (
) (
log ) (
) (
) (
log ) ( ) || (
1
2
2
2
1
1 2 1
The SKL divergence between f1 and
f2:



CPMG based T2-relaxometry: Mechanism
Spins are flipped to XY plane from scanner axis (z- axis). by 90deg
excitation pulse.
Two types of relaxation mechanism after excitation pulse:
Growth along Z-axis (T1 relaxation)
Decay along XY plane (T2 Relaxation)
Apply multiple Refocusing pulse of 180 to observe T2-decay






Multi Echo Spin Echo: one k-space line per excitation; gold standard.
Fast Spin Echo (FSE): multiple k-space lines per excitation.
Improvement: MWFs from Healthy Volunteers
Sequence: 3D MESE (256 x 256)
MWF Maps
Conventional: High spatial variation
even within WM tract.
The proposed algorithm leads to
higher spatial coherence and better
depiction of the WM structures
Geometric mean T2 map for myelin
Conventional Alg: values vary
from 5 to 50 ms even in myelinated
WM tract regions
Proposed Alg: values in myelinated
WM tract region is mostly in the
range of 2030 ms, spatially
consistent with our expectation
Individual T2 Distributions from
3x3 ROI shown by red arrow
proposed algorithm imposes the
similarity of T2-distributions in a
local neighborhood.


Upper Row: MWF maps;
Bottom Row: GMT2 maps
using conventional and
proposed algorithm.

=
=
=
=
=
f
0
f
0
T2 T2
T2 T2
T2 T2
T2 T2
e
Pool
S(T2)
(T2)S(T2) log
GMT2
T2
0
= 5 ms & T2
f
= 50 ms for GMT2
Myelin

Definition
Lower Panel: T2 distributions
from 3x3 ROI using
BLUE: conventional alg.
GREEN: proposed alg.
Fig: Kumar et al (MRM 2012)
Healthy Volunteers at 1.5 T
Notice higher spatial coherence and
better depiction of the WM structures
achieved across all subjects for
proposed algorithm.
Coefficient of Variance:



Measure of spatial variation
Lower COV means more spatial
coherence
Notice that COV has come down
considerably for proposed
algorithm.
Our values of COV is larger than
reported by others possibly due to
large ROI selected.


MWF maps at different spatial regul.
structures Conv. algorithm Proposed P value
WM (entire)
0.72 0.07 0.62 0.10 0.04
GM (entire)
0.90 0.11 0. 84 0.10 0.14
Genu of CC
0.34 0.04 0.20 0.05 0.02
Splenium CC
0.36 0.09 0.18 0.03 0.02
Int. capsule
0.31 0.05 0.18 0.04 <0.001
Table: COV of MWF maps calculated within
various WM and GM regions (N = 5)
ROI inside MWF of Mean
ROI inside MWF Std.dev.of
COV =
Conventional Method Proposed Method


0
0.05
0.1
0.15
0.2
0.25


0
0.05
0.1
0.15
0.2
0.25
MWF map produced at UKE
2D Fast spin Echo (128 x 128)
Averaging of 2.
~ 8 minutes for 7 slices (1/4
brain)
Notice
High spatial variation even
within WM tract for conventional
reconstruction.
The proposed algorithm leads to
higher spatial coherence and
better depiction of the WM
structures.
We believe that a better lesion
quantification would be possible
for multiple sclerosis patients.

Movie: T2W images at various TEs
MWF map produced at UKE
Our Current Planning
3D FSE (better slice profile and no slice gap).
Current filter:
bad scalability and only suitable for accounting N = 9 neighbours in a slice.
The sparse NNLS solver can not solver for one slice at time; so, a part is
usually chosen (say 6 x 6 for 2D filter; 6 x 6 x6 for 3D filter)
X = snnls(Yexp, A);
2D Filter: size of A = (6
2
) x 9x nT2, (6
2
) x 9x (nT2+1) = 16524, 16848
3D filter: size of A = (6
3
) x 27x nT2, (6
3
) x 27x (nT2+1) = 297432, 303264
SNNLS solver can not handle 3D filter
A New 3D filter with better scaling property; no averaging needed; Scan time
halved (entire brain ~15-20 minutes).
Future Planning: not going to discuss in detail.
Would exploit the tempo-spatial correlation in T2R data to cut down the scan
time by > 4.
Improving resolution of myelin imaging using imaging modalities.
Preclinical Stroke Study-1
MWF maps: Stoke model in mice
Mice scan: Challenging due to
requirement of high resolution along
with high SNR .

MWF Maps from stroke mice
undergoing ischemic stroke
Conventional: High spatial
variation even within WM tract.
The proposed algorithm leads to
higher spatial coherence and better
depiction of the WM structures
Also, stroke core and its structure is
better visible.

Experimental autoimmune encephalomyelitis (EAE) model
Non-focal pathologies; gradual changes in normal appearing WM.
Pathologies not visible on T2W image.
Unfortunately, results from myelin water imaging was inconclusive as well;
Further investigation and optimization is needed.
Preclinical Stroke Study-Longitudinal
Variation
variations in myelin (Columns 1, 2):
Between day 1
st
- day 3
rd,
considerable areas in brains of
affected mice have been
remyelinated.
Noisy GMT2 map

Edema (Column 3): GMT2-remaining >
125 ms have significant edema.
Decrease in edema between day 1-
day 3
In future data up to TE ~ 300 ms
would be acquired to accurately
detect edema.

Comparison of myelin water fractions
(columns 1) and respective geometric mean
T2 (GMT2) maps (columns 2-3) for scan
performed on day 1 and day 3.
End of Talk

Thank you