Professional Documents
Culture Documents
Download textbook Information Processing In Medical Imaging 24Th International Conference Ipmi 2015 Sabhal Mor Ostaig Isle Of Skye Uk June 28 July 3 2015 Proceedings 1St Edition Sebastien Ourselin ebook all chapter pdf
Download textbook Information Processing In Medical Imaging 24Th International Conference Ipmi 2015 Sabhal Mor Ostaig Isle Of Skye Uk June 28 July 3 2015 Proceedings 1St Edition Sebastien Ourselin ebook all chapter pdf
https://textbookfull.com/product/health-information-science-4th-
international-conference-his-2015-melbourne-australia-
may-28-30-2015-proceedings-1st-edition-xiaoxia-yin/
https://textbookfull.com/product/mathematics-of-program-
construction-12th-international-conference-mpc-2015-konigswinter-
germany-june-29-july-1-2015-proceedings-1st-edition-ralf-hinze/
https://textbookfull.com/product/ad-hoc-mobile-and-wireless-
networks-14th-international-conference-adhoc-now-2015-athens-
greece-june-29-july-1-2015-proceedings-1st-edition-symeon-
papavassiliou/
https://textbookfull.com/product/proceedings-of-second-
international-conference-on-electrical-systems-technology-and-
information-2015-icesti-2015-1st-edition-felix-pasila/
https://textbookfull.com/product/experimental-algorithms-14th-
international-symposium-sea-2015-paris-france-
june-29-july-1-2015-proceedings-1st-edition-evripidis-bampis-eds/
https://textbookfull.com/product/artificial-intelligence-in-
education-17th-international-conference-aied-2015-madrid-spain-
june-22-26-2015-proceedings-1st-edition-cristina-conati/
Sebastien Ourselin
Daniel C. Alexander
Carl-Fredrik Westin
M. Jorge Cardoso (Eds.)
LNCS 9123
Information Processing
in Medical Imaging
24th International Conference, IPMI 2015
Sabhal Mor Ostaig, Isle of Skye, UK, June 28 – July 3, 2015
Proceedings
123
Lecture Notes in Computer Science 9123
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen
Editorial Board
David Hutchison
Lancaster University, Lancaster, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Friedemann Mattern
ETH Zurich, Zürich, Switzerland
John C. Mitchell
Stanford University, Stanford, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
C. Pandu Rangan
Indian Institute of Technology, Madras, India
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbrücken, Germany
More information about this series at http://www.springer.com/series/7412
Sebastien Ourselin Daniel C. Alexander
•
Information Processing
in Medical Imaging
24th International Conference, IPMI 2015
Sabhal Mor Ostaig, Isle of Skye, UK, June 28 – July 3, 2015
Proceedings
123
Editors
Sebastien Ourselin Carl-Fredrik Westin
Centre for Medical Image Computing Dept. of Radiology
University College London Harvard Medical School Brigham
London and Women’s Hospital
UK Boston, MA
USA
Daniel C. Alexander
Centre for Medical Image Computing M. Jorge Cardoso
University College London Centre for Medical Image Computing
London University College London
UK London
UK
LNCS Sublibrary: SL6 – Image Processing, Computer Vision, Pattern Recognition, and Graphics
scientist who is the first author of a paper and a first-time IPMI oral presenter. This year
16 (of 22) oral presenters were eligible for the Erbsmann Prize.
IPMI 2015 featured a keynote talk describing the Allen Brain atlas and its impact in
neuroimaging research by Boudewijn Lelieveldt, Professor of Biomedical Imaging at
Leiden University Medical Center, and Delft University of Technology, The
Netherlands.
The conference features the traditions that past IPMI attendees have come to expect
as part of the unique character of the meeting. Most importantly, each oral presentation
has unlimited time for discussion to give the audience the opportunity to resolve all
questions regarding the methods and results. IPMI discussions can go on for hours,
involving virtually every attendee in the room! Session chairs play a key role at IPMI
as they strive to ensure that these extended discussions are productive and continuing to
add to the group’s understanding of the nature and significance of each presentation’s
original contribution.
This year, also for the first time, the conference featured an elective oral session.
Meeting attendees voted for the poster presentations they would most like to see
presented orally. The IPMI board selected several poster presenters from those most
popular among the attendees. The selected presenters gave their oral presentation
during the closing session of the conference providing extra exposure to work of key
interest to the community.
IPMI is traditionally held in a small and sometimes remote location. This year’s
venue, the Sabhal Mor Ostaig on the Isle of Skye, Scotland, is a unique venue, situated
above the shoreline of the picturesque Sound of Sleat in Scotland, known worldwide as
the University for Gaelic Language. Sabhal Mor Ostaig is spread over two campuses,
Arainn Ostaig and Arainn Chaluim Chille. The two are interlinked by a well-lit
pathway and are only a short walk apart. The venue itself is conveniently located just
two miles north-east from the main ferry terminal at Armadale village. The venue has
great facilities: The Talla Dhonaidh Caimbeul lecture theater can seat up to 190 people,
there are two additional board rooms that can accommodate up to 25 people each, and
the restaurant consists of a glass-fronted dining room that offers magnificent views over
the Sound of Skye and the nearby Knoydart peninsula.
Conference delegates relaxed and enjoyed various group excursions on the after-
noon of day 3 of the conference. The excursions made the most of the unique location
and culture ranging from whale-spotting in the Sound of Skye, through axe-throwing
competitions, to visiting the Talisker distillery (the most famous of Skye’s malt
whiskies). The afternoon of day 4 of the conference saw the traditional IPMI soccer
match, pitting the American team against “the rest of the world.” After the disap-
pointing 0-0 draw in 2013, “the rest of the world” are still looking to avenge their
first-ever defeat in the history of IPMI in 2011 – an embarrassment that still scars deep.
These proceedings contain the IPMI 2015 papers in the order presented at the
meeting. We hope that this volume serves as a valuable source of information for the
participants, as well as a reminder of the great conference experience. For those who
were not able to attend the conference, we hope that these proceedings provide you
with an excellent summary of the latest research contributions to the medical imaging
field. We look forward to the next IPMI.
Preface VII
The organization of the 24th IPMI conference was only possible through the efforts
and contributions of several organizations and many individuals. First of all, the IPMI
2015 co-chairs would like to thank the members of the Scientific Review Committee
for providing so many high-quality reviews within a very limited time frame; because
of these reviews, we were able to make a fair selection of the best papers for the final
program. We also express our gratitude to the Paper Selection Committee members,
who each read many papers and their reviews and travelled to London for a marathon
organizational meeting that resulted in an outstanding final program.
Thanks also to Microsoft Corporation for assistance with their excellent CMT
conference management system, which we used for the automation of submissions and
review of manuscripts. We also thank Jerry Prince for providing the template for this
front matter. Finally, we thank Ron Gaston and Dominique Drai for their invaluable
administrative support, without which the conference would most certainly not have
happened, and the many local organizers at Sabhal Mor Ostaig for their dedicated
support.
Conference Chairs
Sebastien Ourselin University College London, UK
Daniel C. Alexander University College London, UK
Carl-Fredrik Westin Harvard University, USA
Organizing Committee
Ron Gaston University College London, UK
Dominique Drai University College London, UK
Committee
Simon Arridge University College London, UK
John Ashburner University College London, UK
Suyash Awate University of Utah, USA
Christian Barillot IRISA/CNRS, France
Pierre-Louis Bazin Max Planck Gesellschaft, Germany
Ismail Ben Ayed Western University, USA
Sylvain Bouix BWH Harvard, USA
Djamal Boukerroui Mirada Medical, UK
Michael Brady University of Oxford, UK
Owen Carmichael University of California - Davis, USA
Gary Christensen University of Iowa, USA
Albert Chung HKSUT, HK
Ela Claridge The University of Birmingham, UK
Olivier Commowick Inria Rennes, France
Tim Cootes Manchester University, UK
Sune Darkner DIKU, Denmark
Benoit Dawant Vanderbilt University, USA
Rachid Deriche Inria Sophia Antipolis, France
X Organization
IPMI Board
MRI Reconstruction
Clustering
Statistical Methods
Longitudinal Analysis
Microstructure Imaging
Shape Analysis
Multi-atlas Fusion
Deformation Models
Poster Papers
Functional Nonlinear Mixed Effects Models for Longitudinal Image Data . . . 794
Xinchao Luo, Lixing Zhu, Linglong Kong, and Hongtu Zhu
1 Introduction
The degree of codependence in the spatial distributions of two proteins can reveal
details about their dynamic interaction and help understand several biochemi-
cal processes [4,8]. Such studies are crucial in understanding the mechanisms of
several debilitating diseases, including cancer.
The semantics of colocalization has two aspects: (i) co-occurrence, i.e., the
presence of the two entities within a spatial neighborhood, and (ii) correlation or
codependence, where the two entities co-occur and their local quantities exhibit
a functional relationship. This paper focuses on the latter, more general, aspect
and proposes a framework to automatically estimate the nature and strength of
such colocalization from image data.
In microscopy, the measurement of colocalization typically relies on fluores-
cence imaging that involves simultaneous or successive detection of multiple
fluorescent chemicals in the same biological specimen, where each fluorescent
chemical is designed to bind to exactly one of the biological entities. Thus, the
imaging produces multiple images, lying in the same physical coordinate space,
where the intensity at each pixel in an image is indicative of the presence/absence
of the corresponding biological entity at that pixel location. This paper proposes
to estimate a quantitative measure of colocalization directly from noisy image
data involving two fluorescence microscopy images.
The literature describes several different strategies for formulating and esti-
mating quantitative measures of colocalization [10,11,20,23]. However, none of
these measure colocalization directly from image data, but only from a post-
processed version of the data. This post-processing is typically performed man-
ually and in an adhoc manner, i.e., via an adhoc sequence of algorithms with
several free parameters tuned visually. This leads to loss of reproducibility of the
results. In contrast, this paper proposes a brand new framework for colocaliza-
tion estimation directly from unprocessed data, by solving a single optimization
problem with data-driven parameter optimization for all model parameters. This
framework includes (i) a novel quantitative measure of colocalization, incorpo-
rated as a model parameter, and (ii) a novel scheme for statistical inference.
2 Related Work
3 Methods
Let L := {L1 , L2 } denote a random field modeling the pixel labels of the two
objects. Specifically, let L1 be a binary label image with I pixels {L1i }Ii=1 with
possible label values ∈ {+1, −1}, where li1 = +1 indicates the object’s presence
at pixel i in l1 ; similarly, li1 = −1 indicates the object’s absence at pixel i in l1 .
We follow similar notations and semantics for L2 . We assume that the spatial
distributions of objects are (i) smooth within each image and (ii) codependent
between images. To model these properties, each pixel has intra-image and inter-
image neighbors.
We introduce a neighborhood system N := {Niself , Niother }Ii=1 , where Niself
and Niother denote sets of neighbors of pixel i in the same image (to which pixel
i belongs) and the other image, respectively. To have isotropic neighborhoods,
we employ Gaussian-weighted masks w where the ij-th element wij := exp(−
xj − xi 2 /α2 ), where xi and xj are physical coordinates of pixels i and j,
respectively. We use two such masks, i.e., wself (for intra-image smoothness)
and wother (for inter-image colocalization) with underlying parameters αself and
αother , respectively. We set αself := 2/3 pixels to restrict the neighborhood to 2
neighbors along each dimension (we restrict the neighborhood so that wij > 10−3
for neighbors i, j); we found that this level of smoothness suffices for the data
used in this paper. We set αother to be large enough to be able to model the
longest-range direct spatial interactions between the biological entities; this is
application dependent and informed by prior knowledge, but the results are
fairly robust to the choice of this parameter. In this paper, αother := 2 pixels.
We model the prior probability mass function (PMF) of the label image pair
L as
⎛ ⎞
I I
1 11
P (l) := exp ⎝ β li lj + li2 lj2 wij
self
+ other ⎠
ρli1 lj2 wij ,
Z(ρ, β) i=1 self i=1 other
j∈Ni j∈Ni
(1)
where t1i maps the label value li1 to the object number, i.e., t1i := 1 if li1 = −1
and t1i := 2 if li1 = +1; similarly for t2i .
Let the set of parameters be θ := θP ∪ θL , where the prior parameters are
θP := {ρ, β} and the likelihood parameters are θL := {μ11 , σ 11 , μ12 , σ 12 , μ21 ,
σ 21 , μ22 , σ 22 }. We propose to formulate colocalization estimation as the following
maximum-a-posteriori Bayesian estimation problem over all parameters θ:
arg max max P (z|θ) , where P (z|θ) = P (z, l|θ) = P (z|l, θL )P (l|θP ).
ρ θ−{ρ}
l l
(3)
A novelty in our approach is that we estimate ρ (as well as β) automatically
from the data. We observe that the optimization of the parameters ρ and β is
non-trivial because of their involvement in the partition function Z(ρ, β) that is
intractable to evaluate. Furthermore, we optimize all parameters underlying the
model directly from the data.
where we optimize the per-pixel factors, i.e., PMFs, Fi1 (L1i ), Fi2 (L2i ) to best fit
the posterior. This factorized approximation for the posterior makes the Q(θ; θn )
function more tractable. We describe this strategy in the next section.
The M step subsequently maximizes the Q(θ; θn ) function over parameters θ.
The E and M steps are repeated until convergence; we observe that a few itera-
tions are suffice.
where (i) the first term on the right hand side is an expectation of the log
complete-data probability log P (z, l) over all the factors in F (L) except Fj1 (·)
and (ii) constA is the normalization factor for P (z, L1j ). We see that L(F (L)) is
the negative KL divergence between the distributions Fj1 (L1j ) and log P(z, L1j ).
Thus, fixing the set of functions {Fi1=j (·), Fi2 (·)}Ii=1 , the optimal function Fj1 (·)
that maximizes L(F (L)) is
⎛ ⎞
1
Fj1 (lj1 ) := P(z, lj1 ) = exp ⎝ log P (z, l) Fi1 (li1 ) Fi2 (li2 )⎠, (10)
η i
1 2 li:i=j l i:i=j
we designed P (z, l) to be in the exponential class, log P (z, l) gives us the terms
in its exponent. Second, because we are estimating a PMF solely on the label lj1 ,
we only need to consider terms involving lj1 . Third, by our design, all terms in
log P (z, l) involving the label lj1 are linear functions of other labels li:i
1
=j and li .
2
Thus, the optimal Fj (·) is similar to the conditional PMF of label lj given (i) its
1 1
neighboring labels in both images and (ii) data zj1 , but a crucial difference is that
the contributions of the neighboring labels are through their expectations under
the corresponding factors, instead of their values themselves. This variational
inference scheme is also called mean-field approximation [15].
Within the n-th EM iteration, we propose the following algorithm for varia-
tional optimization of the factors:
1. Input: Observed data z and the current parameter estimates θn .
2. Initialize the expectations of all labels {li1 , li2 }Ii=1 . This can be taken to be
the (i) maximum-a-posteriori estimate given θn and z or (ii) empirically-
computed expectation using Gibbs sampling of the labels from the posterior
PMF. We have found the former strategy to be effective as well as computa-
tionally efficient.
3. Iterate over all pixels (order column by column) in both the label images. At
each pixel, perform the next step.
4. Without loss of generality, consider pixel j in the first image. Analytically
evaluate the optimal PMF Fj1 (·), given the expected values of the interacting
labels and data; as described previously in this section. Analytically evaluate
the expectation of lj1 over this optimal PMF Fj1 (·) and update it.
5. Repeat the last two steps, until convergence of the expected label images and
thereby the factorized approximation F (L|z, θn ).
6. Output: The factorized posterior F (L|z, θn ) in terms of the factors
{Fi1 (·), Fi2 (·)}Ii=1 .
After obtaining the factorized posterior F (L|z, θn ), we evaluate the Q(θ; θn )
function efficiently as follows.
1. Input: Observed data z, the current parameter estimates θn , and the optimal
factorized posterior F (L|z, θn ).
2. Sample a large number S of label image pairs {ls := (ls1 , ls2 )}Ss=1 from
F (L|z, θn ) by independent sampling at each pixel from the label distribu-
tion. This is straightforward (unlike Gibbs sampling) and computationally
cheap because sampling at each pixel is sampling from a binomial (factor)
distribution. Thus, we can easily generate a large sample size to ensure a
high-quality approximation.
3. Closely approximate the Q(θ; θn ) function as follows:
S
θn ) := 1
Q(θ; θn ) ≈ Q(θ; log P (z, ls |θ) (11)
S s=1
θn ).
4. Output: The function Q(θ;
θn ).
The next section describes the M step for optimizing parameters θ using Q(θ;
Colocalization Estimation Using Graphical Modeling and VBEM 11
These updates, for the means and variances of the object-specific and
background-specific Gaussians, are obtained in closed form.
The optimal prior parameters θP maximize the label prior probability:
S
S
arg max log P (z, ls |θ) = arg max log P (ls |θP ). (14)
θP θP
s=1 s=1
The optimization of the prior parameters θP is more difficult because they appear
as part of the intractable partition function Z(θP ). Nevertheless, we have found
an effective strategy for estimating θP , which relies on the principles underly-
ing maximum pseudo-likelihood (MPL) estimation [15] that obtains parameter
estimates by maximizing a pseudo-likelihood function instead of the true likeli-
hood function (Note: in the context of MPL, the “likelihood” is the probability
of observing a label image given the MRF model and underlying parameters).
The key idea is that the pseudo-likelihood function (i) is simpler to optimize
because it eliminates the partition function and, (ii) for sufficiently large data,
mimics the true likelihood function well.
The classic MPL strategy for estimation of MRF parameters relies on a single
observed (label) image, where the MPL estimator is known to be consistent [15].
In our case, we have a set of sampled label images {ls }Ss=1 . This image set can be
considered similar to data obtained by cutting out images from a much larger
image. In a statistical sense, observing one large image drawn from a MRF
model is equivalent to observing many smaller images drawn from the same
model (i.e., with same potential functions and parameters). Thus, we propose to
apply the MPL strategy to the entire set of images by replacing each P (ls |θP )
by its pseudo-likelihood analog to optimize θP as:
S
max log P (ls |θP ) ≈
θP
s=1
S
I I
max log P (lis1 |lN
s1 s2
self , lN other , θP ) P (lis2 |lN
s2 s1
self , lN other , θP ) =
θP i i i i
s=1 i=1 i=1
Another random document with
no related content on Scribd:
The Project Gutenberg eBook of History of the
United States of America, Volume 2 (of 9)
This ebook is for the use of anyone anywhere in the United States
and most other parts of the world at no cost and with almost no
restrictions whatsoever. You may copy it, give it away or re-use it
under the terms of the Project Gutenberg License included with this
ebook or online at www.gutenberg.org. If you are not located in the
United States, you will have to check the laws of the country where
you are located before using this eBook.
Language: English
THOMAS JEFFERSON
1801–1805
HISTORY
OF THE
THOMAS JEFFERSON
By HENRY ADAMS
Vol. II.
NEW YORK
CHARLES SCRIBNER’S SONS
1889
Copyright, 1889,
By Charles Scribner’s Sons.
University Press:
John Wilson and Son, Cambridge.
CONTENTS OF VOL. II.
CHAPTER PAGE
I. Rupture of the Peace of Amiens 1
II. The Louisiana Treaty 25
III. Claim to West Florida 51
IV. Constitutional Difficulties 74
V. The Louisiana Debate 94
VI. Louisiana Legislation 116
VII. Impeachments 135
VIII. Conspiracy 160
IX. The Yazoo Claims 192
X. Trial of Justice Chase 218
XI. Quarrel with Yrujo 245
XII. Pinckney’s Diplomacy 264
XIII. Monroe and Talleyrand 288
XIV. Relations with England 316
XV. Cordiality with England 342
XVI. Anthony Merry 360
XVII. Jefferson’s Enemies 389
XVIII. England and Tripoli 410