extracting white matter hyperintensities in alzheimer's disease risk and aging studies using...
TRANSCRIPT
Poster Presentations: P1P262
selectively binds to PHF-tau in AD brains, but does not bind to tau lesions in
non-AD tauopathies, nor to a -synuclein in PD brains.
P1-295 A DIFFERENTIAL PATTERN OF HIPPOCAMPAL
ATROPHY IN ALZHEIMER’S DISEASE WITH
COEXISTENT SMALLVESSEL DISEASE: A
MULTIVARIATE SHAPE-BASED ANALYSIS
Sean Nestor1, Sandra Black2, 1University of Toronto, Toronto, Ontario,
Canada; 2Sunnybrook Research Institute, Toronto, Ontario, Canada.
Contact e-mail: [email protected]
Background:Hippocampal volume is a sensitive marker of Alzheimer’s dis-
ease (AD) progression. Previous studies in AD have demonstrated heteroge-
neous atrophy profiles among hippocampal subregions. Cerebral small vessel
disease (SVD) often coexists with AD, and may differentially affect atrophy
within sub-regions of the hippocampus compared to AD alone. Shape-based
studies have traditionally used univariate methods, which independently test
significance at each element of a surface/structure (vertex/voxel). Conversely,
multivariate models such as Partial Least Squares (PLS) indentify signifi-
cantly distributed patterns of atrophy across all structural elements. Thus,
we examined with PLS,whether a differential pattern of hippocampal atrophy
existed between AD, AD+SVD and normal controls (NC).Methods: Cross-
sectional data were acquired from the Sunnybrook Dementia Study: AD
(n¼144), AD+SVD (n¼47) and NC (n¼95). All subjects had 1.5 Tesla
T1-SPGR MRIs (Matrix¼256x192; TE/TR¼35ms/5ms; flip-angle¼35�, in-plane resolution¼0.85930.859x1.2-1.4mm). Participants with SVD had
evidence of white matter hyperintensities on PD/T2-MRI and/or evidence
of subcortical lacunar infarcts. A novel surface-based technique was devel-
oped to measure hippocampal shape differences. Briefly, an average template
was generated and used to align all subjects to a common space. Next, the Ad-
vanced Normalization Tools SyN algorithm was used to nonlinearly register
all subjects to the template. A mesh was generated over the template hippo-
campi, and the inverse warp vector was indexed and multiplied by the normal
vector at each mesh vertex; this provided a vertex-wise scalar value of shape
difference in relation to the template surface across all subjects (Matlab-Math-
Works). PLS software (http://www.rotman-baycrest.on.ca/index.php?
section¼84) was adapted for surface-based parametric data. Mean-centred
PLS with bootstrapping(x1000) and permutation testing(x1000) was applied
to detect significant patterns of atrophy between groups. Analyses were cor-
rected for sex and age.Results: PLS revealed a significantly different pattern
of atrophy in NC versus AD and AD+SVD involving the anteriolateral CA1,
medial subiculum and posteriorlateral hippocampus (p<0.05). A trend was
realized for less atrophy in the left anterior-medial hippocampal region for
AD+SVD versus AD, with more involvement of the left lateral subiculum
in AD and left anteriolateral region in AD+SVD (p ¼0.06). Conclusions:
These data suggest that a differential pattern of hippocampal atrophy may ex-
ist in AD+SVD versus AD and may provide insight into how these coexistent
pathologies interact in-vivo.
P1-296 CORRECTION OF SCANNER DIFFERENCES IN
MULTICENTER J-ADNI AND U.S.-ADNI PET
STUDIES
Figure 1. Example segmentation results of our methods (SC, RC and RR)
compared to LSTopt (optimal baseline): Each row corresponds to one subject.
First column are the expert indications overlayed onto bias corrected T2-
MR, second to last columns are the LSTopt outputs followed by CPEs of
SC, HC and RR respectively, overlayed on the corresponding bias corrected
T2-MR. The color map of overlay range from blue (0) to red (1).
Ken Fujiwara1, Takashi Kato1, Kengo Ito1, Michio Senda1, Kenji Ishii1,
Kazunari Ishii2, Takeshi Iwatsubo3, Japanese Alzheimer’s Disease
Neuroimaging Initiative4, 1J-ADNI PETcore, Obu, Japan; 2J-ADNI
PETcore, Osakasayama, Japan; 3The University of Tokyo, Tokyo, Japan;4J-ADNI, Tokyo, Japan. Contact e-mail: [email protected]
Background: It has been reported that systematic differences of cerebral
FDG-PET image among PET cameras can be obstacles to statistical analy-
ses of PET image. The purpose of this study was to evaluate and correct the
difference between scanners with brain phantom scans. Methods: FDG-
PET images were obtained in 95 normal subjects (NL) with major seven
models of PET scanner at the base line in J-ADNI study. The images of other
scanner models were excluded because of small number of scans. The im-
ages were classified into seven groups as to the scanner model. The PET im-
ages of Hoffman phantom were smoothed with Gaussian filter of 12mm in
FWHM and spatially normalized to PET template of which cerebellum was
masked in order to avoid mis-registration. The normalized PET images were
averaged in each scanner model and in overall. These spatial processes were
done using SPM8.Seven correction filter images were calculated as ratio of
the averaged overall scanner models to averaged each scanner model on
pixel by pixel basis. Difference-corrected images were created by multiply-
ing each image with the correction filter image of its own scanner model.
T-test group comparisons were made between scanner models with and
without the difference-correction on pixel basis. The images of 78 subjects
by six scanner models in US-ADNI were processed and analyzed using the
same methods as above.Results: The degree of difference between scanner
models became smaller after the difference correction. The differences of
PET images between J-ADNI and US-ADNI were reduced with the correc-
tion method. Conclusions: Systematic image difference owing to scanner
models can bemade to be smaller by the correction using the Hoffman phan-
tom data. T his correction method may be effective to obtain more accurate
statistical results in multi-center studies.
P1-297 EXTRACTING WHITE MATTER
HYPERINTENSITIES IN ALZHEIMER’S DISEASE
RISK AND AGING STUDIES USING SUPERVISED
SEGMENTATION METHODS
Vamsi Ithapu1, Vikas Singh2, Benjamin Austin1, Chris Hinrichs3,
Cynthia Carlsson4, Barbara Bendlin1, Sterling Johnson4, 1University of
Wisconsin-Madison, Madison, Wisconsin, United States; 2University of
Wisconsin-Madison, Madison, Wisconsin, United States; 3Univesity of
Wisconsin-Madison, Madison, Wisconsin, United States; 4VA GRECC,
Madison, Wisconsin, United States. Contact e-mail: [email protected]
Background: White matter hyperintensities (WMH) observed in T2-
weighted Fluid Attenuated Inversion Recovery (FLAIR) Magnetic Reso-
nance Images (MRI) may reflect comorbid neural injury or cerebral vascular
Table 1
Segmentation performance measures of LSTopt (optimal baseline) compared
to our methods, SC, RC and RR: TP - True positives, FP - False positives, TN
- Trne negatives, FN - False negatives. F-measure (or Dice Coe_cient) is the
maximumof the ratio of 2TP to 2TP+FP+FN. Average Pression (AP) (which
is the same as area under Precision Recall curve) and Break Even Point
(BEP) summarize the e_ectivity of each method in minimizing both FP and
FN simultaneously. F0:5 and F2 penalize FP over FN and FN over FP
respectively. Note that the second column represents the underlying learning
model of each of the above methods.
Method Model F AP BEP F0:5 F2
LSTopt MRF 0:406 0:310 0:394 0:402 0:474
SC SVM 0:520 0:525 0:504 0:518 0:586
RC RF 0:642 0:749 0:632 0:647 0:708
RR RF 0:649 0:762 0:638 0:645 0:723
Poster Presentations: P1 P263
disease burden. Accurate segmentation and quantification of these WMH is
of substantial interest in aging, and age related neurological disorders such
as Alzheimer’s Disease (AD). Several existing automated lesion detection
algorithms, that use unsupervised learning models, result in unreliable de-
tections. This is especially apparent in older population where WMH are
small, diffuse and irregular in shape, and sufficiently heterogeneous within
and across subjects.Methods:We pose the detection of hyperintensities as
a supervised inference problem and adapt two learning models, specifically,
Support Vector Machines (SVM) and Random Forests (RF). Structural in-
formation that characterizes a hyperintensity is incorporated through texture
features engineered using texton based filter banks, and training is employed
to supervise the models. Interpreting the segmentation as both regression
and classification, we provide a suite of effective segmentation methods.
The detections are further quantified as a scalar accumulation parameter, re-
ferred to as Effective WMH Volume (EV). Extensive evaluations are con-
Table 2
Regression analysis of the hyperintensity accumulation (EV ) scores from
two of our methods (RC and RR) with six vascular factors, CL - Cholesterol,
HDL - Highdensity Lipoprotein, SBP - Systolic Blood Pressure, DBP -
Diastolic Blood Pressure, BMI - Body Mass Index and CV Risk - 10–year
cardiovascular risk: EV –RC (EV – RR resp.) denotes the EV s calculated
using RC (RR resp.) method. Significant p–values (and corresponding t–
values) at 95% confidence level and higher are bold-faced.
Variables Statistics
Dependent Explanatory t–value p–value
EV –RC
CL 1:582 0:117
HDL �2:128 < 0:05
SBP 4:174 < 0:001
DBP 3:977 < 0:001
BMI 1:096 0:276
EV –RC
CV Risk 2:922 < 0:01
Variables Statistics
Dependent Explanatory t–value p–value
EV –RR
CL 1:581 0:117
HDL �2:094 < 0:05
SBP 4:166 < 0:001
DBP 3:991 < 0:001
BMI 1:067 0:289
EV –RR
CV Risk 2:932 < 0:01
ducted on 140 subjects (male: 55, female: 85). This included 107
healthy (median age: 61.16), 21 declining (median age: 60.14) and 12
mild cognitively impaired (MCI) (median age: 65.16) subjects, who
vary in cardiovascular and AD risk factors. A multivariate regression anal-
ysis is done to evaluate the association of this hyperintensity accumulation
to vascular risk factors. Results: Segmentation results, and related perfor-
mance measures (like Precision-Recall curves, Dice Coefficients etc.)
show that our methods outperform an existing state-of-the-art unsuper-
vised lesion detection algorithm. Secondary statistical analysis indicate
the association of hyperintensity accumulation to vascular risk factors,
in particular, systolic and diastolic blood pressure and Framingham risk
score. Conclusions: Our evaluations highlighted the importance of user
supervision in the form of expert indications, for accurate and reliable seg-
mentation of hyperintensities in older population. We described a new
summary measure of hyperintensity accumulation, and validated its effi-
cacy using cardiovascular risk factors. These experiments are accompa-
nied with an open source library (interfaced with existing neuroimaging
tools) that can be adapted for segmentation problems in other neuroimag-
ing studies.
P1-298 CORTICAL SURFACE PROJECTION OF PET
IMAGES FROM 11C- AND 18F-LABELED
RADIOTRACERS WITHOUT MAGNETIC
RESONANCE IMAGING
Vincent Dor�e1, Pierrick Bourgeat2, Luping Zhou3, Jurgen Fripp4,
Ralph Martins5, Lance Macaulay6, Colin Masters7, David Ames8,
Belinda Brown9, Christopher Rowe10, Olivier Salvado11,
Victor Villemagne12, 1CSIRO, Melbourne, Australia; 2CSIRO, Herston,
Australia; 3CSIRO, Herston, Australia; 4AeHRC, Herston, Australia; 5Edith
Cowan University, Perth, Australia; 6CSIRO, Parkville, Australia;7University of Melbourne, Melbourne, Australia; 8National Ageing
Research Institute Inc. (NARI), Parkville, Australia; 9Edith Cowan
University, Perth, Australia; 10Austin Hospital, Melbourne, Australia;11CSIRO, Melbourne, Australia; 12Austin Health, Melbourne, Australia.
Contact e-mail: [email protected]
Background: Clinical PET imaging in Alzheimer’s disease relies in the
visualisation of Ab deposition and glucose metabolism in the cortical
gray matter. Due to the limited structural information in PET images,
automatic tissue-specific assessment is usually performed with the aid
of MR images, which may not always be available. We evaluated a novel
MR-less method to locally estimate and project the cortical tracer reten-
tion on a common surface template. Methods: Several subjects were
scanned with different radiotracers: 18 F-Flutemetamol, 18 F-Florbeta-
pir, 18 F-Florbetaben, 11 C-PIB, 18 F-NAV4694 and 18 F-FDG. First,
each individual PET image was normalised in the MNI space and
SUVR scaled with a common cortical cerebellum mask. Radiotracer re-
tention was then estimated within several GM prior atlases. On scans
acquired on a PET/CT scanner, CT was used to estimate the GM priors.
Atlas selection and Bayesian fusion were then used for generating esti-
mated surface values, reflecting the pattern of either high (A ligands) or
low (FDG) tracer retention. Surface projections and native transaxial,
sagittal and coronal PET images were visually graded by clinicians
blinded to clinical diagnosis. Images were read separately and graded
as normal, possible AD or probable AD. For sensitivity and specificity
calculations, "possible" and "probable AD" were combined. Results: In
the visual readouts, surface projection images provided higher inter-rater
reliability and much greater reader confidence than native PET images. Vi-
sual assessment of surface projections were both very sensitive and spe-
cific for AD and performed better than visual readouts of native PET
images. Conclusions: The proposed method demonstrated accurate esti-
mations of radiotracer retention in the cortex for various 11 C and 18 F la-
beled radiotracers. Our approach provides a practical and efficient clinical
inspection tool for PET.