Title: Style B 36 by 54 wide
1 P2-013 Superior
Alzheimer's diagnosis by combination of
partial-volume effects correction and
multivariate analysis Christian Habeck, Ajna
Borogovac, Truman Brown, Iris Asllani, Yaakov
Stern Taub Institute and Hatch MRI Research
Center, Department of Radiology, Columbia
University Medical Center, NY, NY 10032
Multivariate versus univariate analysis. For
the last part of the poster we would like to
illustrate the differences between multivariate
and univariate analysis with a simple split-half
analysis. (The univariate results are discussed
at length in detail in the adjacent poster
IC-P3-166.) For the split-half analysis, both AD
patients and healthy controls are divided into
bipartitions (derivation and replication samples)
10,000 times. Each time, our analysis discussed
before is performed in the derivation sample,
i.e. a covariance pattern is derived with maximal
discriminatory power, while the replicability is
checked in the replication sample. T-statistics
and ROC-areas are computed in both derivation and
replication sample. For the univariate analog of
this analysis, we just picked the voxel with the
highest T-contrast in the derivation sample.
T-statistic as well ROC-area were also recorded
in both derivation and replication samples for
this disease marker. The panels
above show the results of the split half
analysis. On the left side, ROC-areas and
T-statistics are plotted, for both markers (UV
and MV) and both samples (DER and REP). One
can appreciate the better replicability of the
multivariate results for both statistics. Most
striking are the high T-values achieved for the
univariate marker in the derivation sample, which
are not followed up with equal goodness in the
replication sample. On the right side, we plotted
a subset of 1,000 T-statistics from both
derivation and replication samples against one
another to convey the sample-to-sample
variability in our performance metrics.
Interestingly, for the multivariate marker (blue)
there is a negative correlation between
performance in both samples better performance
in the derivation sample implies worse
performance in the replication sample. We see
this as an example of bias-variance trade-off
better group discrimination in the derivation
sample might be achieved through overfitting
noise, which hampers the performance in the
replication sample. For the univariate marker
(red), this relationship is much weaker, but
still present. The general cautionary tale
illustrated here is that strong measures of
association do not guarantee good replication of
results.
Introduction and Methods
Multivariate analysis on Cerebral Blood Flow
(CBF) measured by Arterial Spin Labeling (ASL)
has shown promise as an early marker of
Alzheimers disease1 (AD). Further improvement on
this preliminary result should be possible by
adjusting for partial-volume effects in the ASL
data. Since net total CBF at any brain voxel has
different contributions from different tissue
types according to the voxels tissue
composition, it follows that (1) structural brain
changes can impact net total CBF, even though CBF
per unit volume of tissue type is unchanged. The
converse might be true as well, i.e. (2) CBF per
unit volume of tissue type can change, even when
the overall structural composition of a voxel
stays constant. Both effects could be utilized in
the diagnosis of AD with respect to Gray Matter
(GM) during the course of AD progressive atrophy
results in drastic lowering of GM density in the
brain, which will manifest as lower total net CBF
(1). We want to investigate whether relative CBF,
i.e. CBF per unit volume of GM is also affected
(2). This may be particularly important in the
prodromal disease stages, prior to any obvious
structural changes and Gray Matter atrophy in the
brain if Gray Matter CBF is already lowered in
this very early stage, this might provide
valuable diagnostic and prognostic information.
1) Data acquisition ASL perfusion images were
acquired on 10 patients with moderate-to-severe
probable AD and 30 age-matched healthy controls.
Pre-processing comprised of various stages
described previously1, with the additional step
of partial-volume correction inserted prior to
the spatial normalization to the MNI template3.
2) Partial-volume effects correction a method of
computing pure-tissue contributions to regional
CBF 2,3 was used prior to our group analysis to
isolate the gray-matter contributions to CBF in
particular. 3) Group-level data analysis
Principal components analysis (PCA) and linear
discriminant analysis was performed on all scans
to derive a pattern that distinguished patients
from controls (AD1/Controls0), similar to our
previous report1. Bootstrap resampling was
performed to produce Z-maps to identify the most
robust areas in the covariance pattern. This was
done for both total CBF scans as well as pure
Gray-Matter CBF scans. Several robustness
computations were performed for both scan types
for comparison purposes (described in Results
section). We derived a discriminant
pattern in both net CBF scans as well as pure GM
scans. Classification accuracy was very high for
both with almost identical results.
The plots above show subject expression for the
first Principal Components of both pure GM and
net CBF analyses. Diagnostic classification was
very good, the ROC area was 0.96 for both
patterns (left panel) and the expression values
highly correlated (right).
Pure GM vs net CBF. The panels above show
significant negative voxel loadings (uncorrected
plt0.001) indicating decreased CBF in the AD
patients, superimposed on a probabilistic
Gray-Matter mask. (No positive loadings were
obtained.) On the left side, significant voxel
loadings that were common to both covariance
patterns are displayed. Large areas are
participating in the covariance patterns,
manifesting a large whole-brain flow difference
between both groups. On the right side, an XOR
plot shows voxels with significant loadings that
are unique to each pattern (BLUE pure GM
pattern, RED net CBF). The results are complex
and do not permit easy characterization however,
one general observation is true the unique
voxels of the GM pattern are generally more
anterior than the unique net-CBF pattern voxels,
this is particularly obvious for ventromedial
prefrontal areas that are only showing up in the
GM pattern. On the other hand, there are no
parietal areas in Brodmann Areas 7, 40 that are
unique to the GM pattern, only the net-CBF
pattern. Next, we want to investigate further
whether the pure GM flow can supply useful and
robust information beyond net CBF. Adjusting for
partial volume effects should pay off
particularly in regions with mixed tissue types,
since Gray Matter atrophy will have attenuated
effects in the regions due to the noise of other
tissue types, making these regions less useful as
a diagnostic marker. To simulate this scenario
realistically, we repeated our analysis on
various voxel subsets using the probabilistic
Gray-Matter mask supplied by SPM5, we picked all
voxels whose Gray Matter probabilities were less
than a range of thresholds chosen from 1 to 99.
We expected that for high threshold both GM and
net CBF pattern would give similar performance,
while diverging for low thresholds.
Results
This, of course, will only work out if the CBF
per unit GM volume is really also decreased. The
figure to the right shows the classification
performance of net CBF and GM patterns derived
from restricted set of mixed-tissue voxels as
anticipated, the worse the admixture of non-GM
tissue, the more useful the partial-volume
correction becomes in ensuring good diagnostic
classification.
Conclusion and References
- - Both net CBF and GM-CBF images achieve very
good diagnostic classification with multivariate
analysis - More robust classification achieved by
multivariate analysis than univariate analysis - Future plans Prospective application on MCI
population for comparison of prognostic power - Asllani et al., JCBF M, Oct24 (2007)
- Donahue et al., MRM X (2006)
- Asllani et al. (Poster IC-P3-166, ICAD 2008,
Chicago).