-
Notifications
You must be signed in to change notification settings - Fork 0
/
paper.Rnw
359 lines (316 loc) · 19.5 KB
/
paper.Rnw
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
\documentclass{article}
\usepackage{spconf,amsmath,graphicx,color}
\usepackage[square,numbers]{natbib}
\setlength{\bibsep}{0.0pt}
\title{How many atlases do you really need for accurate multi-atlas segmentation?}
\name{ Jon Pipitone$^{1}$ Jason P. Lerch$^{4,5}$ Miriam Friedel$^{4}$
Aristotle N. Voineskos$^{1,3}$ M. Mallar Chakravarty$^{1,2,3}$ }
\address{
$^1$ Kimel Family Translational Imaging Genetics Research Laboratory, \\
Research Imaging Centre, Centre for Addiction and Mental Health, Toronto, ON, Canada \\
$^2$ Institute of Biomaterials and Biomedical Engineering, University of
Toronto, Toronto, ON, Canada \\
$^3$ Department of Psychiatry, University of Toronto, Toronto, ON, Canada \\
$^4$ Program in Neuroscience and Mental Health, The Hospital for Sick Children,
Toronto, ON, Canada \\
$^5$ Department of Medical Biophysics, University of Toronto, Toronto, ON, Canada
}
\begin{document}
\SweaveOpts{concordance=TRUE}
<<setup include=FALSE, echo=FALSE>>=
library(knitr)
opts_chunk$set(
concordance=TRUE
)
@
\maketitle
\begin{abstract}
Neuroimaging research often relies on automated anatomical segmentations of MR
images of the brain. Current multi-atlas based approaches provide accurate
segmentations of brain images by propagating manually derived segmentations of
specific neuroanatomical structures to unlabelled data. These approaches often
rely on a large number of such manually segmented atlases that take significant
time and expertise to produce. We present an algorithm for the automatic
segmentation of the hippocampus that minimizes the number of atlases needed
while still achieving similar accuracy to multi-atlas approaches. We perform
repeated random subsampling validation on the International Brain Segmentation
Repository (IBSR) dataset to compare our approach to basic multi-atlas
segmentation using the full IBSR dataset, and to single-atlas (model-based)
segmentation. Our results show that with only 8 input atlases, MAGeT brain can
achieve to within 2.0\% segmentation accuracy of the basic multi-atlas approach
using 17 input atlases (mean $\kappa = 0.775$ vs. $\kappa = 0.791$). These
results compare favourably to those of other investigators who have used the IBSR
data for validation.
\end{abstract}
\begin{keywords}
hippocampus, segmentation, multi-atlas methods, label fusion, magnetic
resonance imaging
\end{keywords}
\section{Introduction}
\label{sec:intro}
The hippocampus is of particular interest to many researchers because it is
implicated in forms of brain dysfunction such as Alzheimer's disease and
schizophrenia, and has functional significance in cognitive processes such as
learning and memory. For many research questions involving magnetic resonance
imaging (MRI) data accurate identification of the hippocampus and its
subregions is a necessary first step to better understand the individual
neuroanatomy of subjects.
Currently, the gold standard for neuroanatomical segmentation is manual
delineation by an expert human rater. This is problematic for segmentation of
the hippocampus for several reasons. First, manual segmentation takes a
significant investment of time and expertise \citep{Hammers2003} which may not
be readily available to researchers or clinicians. Second, the amount of data
produced in neuroimaging experiments increasingly exceeds the capacity for
identification of specific neuroanatomical structures by an expert manual
rater. Third, the true definition of hippocampal anatomy in MR images is
disputed \citep{Geuze2004}, as evidenced by efforts to create an unified
segmentation protocol \citep{Jack2011}.
Compounding each of these problems is the significant neuroanatomical
variability in the hippocampus throughout the course of aging, development, and
neuropsychiatric disorders \citep{Mouiha2011}. Additionally, it may be
necessary to use several different hippocampal definitions or, in fact, make
specific modifications in the course of research. For example, Poppenk et al.
\citep{Poppenk2011} found that subdividing the hippocampus into anterior and
posterior regions resulted in a predictive relationship between volume
difference of those regions and recollection memory performance. Thus, while
manual segmentation of the hippocampus is a necessary technique, to researchers
or clinicians who do not have access to the needed human expertise its use may
be infeasible.
Automated segmentation techniques overcome the need for human expertise by
performing segmentations computationally. A popular class of automated
methods, {\it multi-atlas-based segmentation}, rely on a set of expertly
labeled neuroanatomical atlases. Each atlas is warped to fit a subject's
neuroanatomy using nonlinear registration
techniques \citep{Collins1995,Klein2009}. Atlas labels are then transformed by
this warping and a {\it label fusion} technique, such as voxel-wise voting, is
used to merge the competing label definitions from each atlas into a final
segmentation for a subject.
Many descriptions of multi-atlas-based segmentation algorithms report relying
on an atlas library containing anywhere between 30 and 80 expertly labeled
brains \citep{Heckemann2011,Collins2010,Aljabar2009,Leung2010,Lotjonen2010}. As
noted, the production of an atlas library requires significant manual effort,
and is limited since the choice of atlases or segmentation protocol may not
reflect the underlying neuroanatomical variability of the population under
study or be suited to answer the research questions at hand.
In this paper we propose an automated segmentation method to address the above
issues of existing multi-atlas-based methods. Principally, our method aims to
dramatically reduce the number of manually labelled atlases needed (under 10).
This is achieved by using the small atlas library to boot-strap a much larger
"template library", which is then used to segment the subjects in a similar
fashion to basic multi-atlas segmentation. This approach has the additional
advantage of using the unique subject population on hand to initialize the
segmentation process and improve accuracy.
The essential insight of generating a template library is not new. Heckemann
\citep{Heckemann2006} compared generating a template library from a single atlas
to standard multi-atlas segmentation and found poor performance and so deemed
the approach as inviable. The LEAP algorithm \citep{Wolz2010} proceeds by
iteratively segmenting the unlabelled image most similar to the atlas library
images and then incorporating the now-labelled image into the atlas library,
but requires 30 starting atlases. The novelty of our method is to demonstrate
the possibility of producing comparable segmentation accuracy to these and
other multi-atlas-based methods while using significantly fewer manually
created atlases.
In our previous work \citep{Chakravarty2012}, we applied MAGeT brain to
segmentation of the human striatum, globus pallidus, and thalamus using a
single histologically-derived atlas. The main contribution of this paper is to
extend our approach to the human hippocampus and perform a thorough validation
over a range of atlas and template library sizes, which was not done in our
previous work. Due to the small number of atlases required, our method can
easily accommodate different hippocampal definitions. Our aim is not to improve
on segmentation accuracy beyond existing methods, but instead to provide a
method that trades off manual segmentation expertise for computational
processing time while providing sufficient accuracy for clinical and research
applications.
\section{Materials and Methods}
\subsection{The Multiple Automatically Generated Templates (MAGeT) Algorithm}
In this paper, we use the term {\it atlas} to mean any manually segmented MR
image, and the term {\it atlas library} to mean a set of such images. We use
the term {\it template} to refer to any MR image, and associated labelling,
used to segment another image, and the term {\it template library} to refer to
a set of such images. An atlas library may be used as a template library but,
as we will discuss, a template library may also be composed of images with
computer generated labellings.
The segmentation approach we propose is best understood as an extension of
basic multi-atlas segmentation \citep{Collins2010}. In multi-atlas segmentation,
an atlas library and unlabelled MR images are given as input. Every atlas
image is nonlinearly registered to each unlabelled image, and then each atlas'
labels are propagated via the resulting transformations. These labels are then
fused to produce a single, definitive segmentation by some label fusion method
(e.g. voxel-wise majority vote).
Our extension adds a preliminary stage in which a template library is
constructed from input images, and used in place of an atlas library in the
standard multi-atlas-based method. To create the template library, labels from
each atlas image are propagated to each template library image via the
transformation resulting from a non-linear registration between pair of images.
As a result, each template library image has a label from each atlas.
Basic multi-atlas segmentation is then used to produce segmentations for the
entire set of unlabelled images (including those images used in the template
library).
Label fusion is performed by cross-correlation weighted voting, a strategy
weighted towards an optimal combination of subjects from the template library
which has been previously shown to improve segmentation
accuracy \citep{Aljabar2009,Collins2010}. In this method, each template library
image is ranked in similarity to each unlabelled image by the normalized
cross-correlation of image intensities after linear registration in a region of
interest (ROI) generously encompassing the hippocampus. Only the top ranked
template library image labels are used in a voxel-wise majority vote. The ROI
is heuristically defined as the extent of all atlas labels after linear
registration to the template, dilated by three voxels \citep{Chakravarty2012}.
\subsection{MRI dataset evaluated}
For evaluation purposes we used the publicly available IBSR dataset. This
dataset consists of T1-weighted MR image volumes from 18 subjects (4 females,
14 males) with ages between 7 and 71 years. Image dimensions for all MR volumes
are normalized to $256 \times 256 \times 128$ voxels, with the voxel size
ranging from $0.8 \times 0.8 \times 1.5 mm$ to $1.0 \times 1.0 \times 1.5 mm$.
The images come 'positionally normalized' into the Talairach orientation
(rotation only), and processed by the CMA 'autoseg' biasfield correction
routines. The MR brain data sets and their manual segmentations are publicly
available and were provided by the Center for Morphometric Analysis at
Massachusetts General Hospital and are available at
{\tt http://www.cma.mgh.harvard.edu/ibsr/}.
\subsection{Image Processing and Registration Method}
The N3 algorithm \citep{Sled1998} is first used to minimize the intensity
nonuniformity in each image. Image registration is carried out in two phases.
In the first, a 12-parameter linear transformation (3 translations, rotations,
scales, shears) is estimated between images using an algorithm that maximizes
the correlation between blurred MR intensities and gradient magnitude over the
whole brain \citep{Collins}. In the second phase, nonlinear registration is
completed using the ANIMAL algorithm \citep{Collins1995}: an iterative procedure
that estimates a 3D deformation field between two MR images. At first, large
deformations are estimated using blurred version of the input data. These
larger deformations are then input to subsequent steps where the fit is refined
by estimating smaller deformations on data blurred with a Gaussian kernel with
a smaller FWHM. The final transformation is a set of local translations defined
on a bed of equally spaced nodes that were estimated through the optimization
of the correlation coefficient. For the purposes of this work we used the
regularization parameters optimized in Robbins et al. \citep{Robbins2004}. It
should be noted that the MAGeT brain algorithm is not dependent on this, or
any, particular choice of registration method \citep{Chakravarty2012}.
\subsection{Experiments}
We explored how varying the size of the atlas library and the template library
effects labeling accuracy. For each parameter setting we conducted 30 rounds
of random subsampling cross-validation using the 18 manual segmented templates
from the IBSR dataset as input. In each round, atlases were randomly chosen
from the IBSR dataset and the remaining images are used both as template
library images and unlabeled subjects to be labeled using the MAGeT brain
algorithm. We varied the size of the atlas library from 3 to 8, and used
cross-correlation weighted label fusion to select the top $n$ candidate
templates from the remaining images. $n$ was varied in the range $[3, 18-a]$,
where $a$ is the size of the atlas library.
\subsection{Evaluation}
\subsubsection{Goodness-of-fit}
Automatically produced segmentations are evaluated against IBSR manual
segmentations dataset using the Dice Kappa ($\kappa$) overlap metric, $\kappa =
{2a}/{(2a+b+c)}$, where $a$ is the number of voxels common to both
segmentations and $b+c$ is the sum of the voxels uniquely identified in either
segmentation.
\subsubsection{Comparison Approaches}
The resulting segmentations from each of our experiments are compared to those
produced from two alternative segmentation approaches. The {\it single-atlas}
approach uses one atlas to segment a unlabelled subject by directly propagating
labels from the atlas by way of nonlinear registration. We computed a
single-atlas segmentation for each image in the IBSR dataset from each of the
other 17 labelled images. Similarly, we computed a segmentation for each image
using the {\it basic multi-atlas} approach, described above, using the
other 17 images as the atlas library. Additionally, we also varied the number
of atlas images used in the label fusion step by employing cross-correlation
weighted voting. In total we evaluated approximately $52,000$ segmentations
for the work presented in this manuscript.
\section{Results}
Sample segmentations from a single IBSR subject compared with the gold-standard
segmentation are in Fig. \ref{montage}. Qualitatively, as the size of the
template library is increased, the number of false negatives in the hippocampal
tail region is reduced, and segmentation accuracy increases.
\begin{figure}[h]
\begin{minipage}[b]{1.0\linewidth}
\centering
\includegraphics[width=\textwidth]{montage.png}
\end{minipage}
\caption{{\small
Segmentations of a single subject are shown when three atlases are used, with
varying template library size $n$. Blue colouring represents agreement between
the gold-standard and MAGeT brain. Green colouring indicates false positive
voxels labelled by MAGeT brain (i.e. labelled voxels not appearing in the
gold-standard labels), and red colouring indicates false negative voxels (i.e.
voxels labelled in the gold-standard but not by MAGeT brain).
}}
\label{montage}
\end{figure}
MAGeT brain achieves a level of segmentation accuracy that is to within 2.0\%
of the accuracy of the basic multi-atlas approach (Fig. \ref{results};
mean $\kappa = 0.775$ vs mean $\kappa = 0.791$) in the best case. Importantly,
to do this MAGeT brain only requires 8 manual segmented images whereas the
basic multi-atlas requires 17 atlases. This represents a significant
savings in manual effort, and supports our contention that it is possible with
MAGeT brain to trade a small decrease in accuracy for a significant decrease in
the number of manual segmentations needed.
Surprisingly, our analysis does not show any significant improvements in
segmentation accuracy for either method when applying cross-correlation
weighted voting to reduce the number of template labels being fused.
\begin{figure}[t]
\centering
<<performance, cached=TRUE, echo=FALSE, warnings=FALSE >>=
library(ggplot2)
results = read.csv(file='data/results.csv')
xcorr_vote= subset(results, method=='xcorr' | method == 'multiatlas_xcorr_vote' )
qplot(x=top_n, y=k_hippocampus, data=xcorr_vote, colour=as.factor(num_atlases), geom=c('point'), alpha=I(0)) +
geom_smooth(method='loess')+ #line=1.5, formula=y~poly(x,2), method='lm') +
geom_hline(aes(yintercept = mean(subset(results, method=='naive')$k_hippocampus), linetype="dashed")) +
theme(legend.position="none") +
ylab('Mean similarity score (Kappa)') +
xlab('Number of template labels fused') +
annotate("text", 14.3, 0.712, label="Single-atlas", size=4, hjust=0)+
annotate("text", 15.3, 0.754, label="3 atlases", size=4, hjust=0)+
annotate("text", 14.4, 0.766, label="4 atlases", size=4, hjust=0)+
annotate("text", 13.5, 0.770, label="5 atlases", size=4, hjust=0)+
annotate("text", 12.5, 0.773, label="6 atlases", size=4, hjust=0)+
annotate("text", 11.5, 0.776, label="7 atlases", size=4, hjust=0)+
annotate("text", 10.3, 0.779, label="8 atlases", size=4, hjust=0)+
annotate("text", 13.5, 0.793, label="Basic multi-atlas", size=4, hjust=0)+
annotate("text", 14.6, 0.787, label="17 atlases", size=4, hjust=0)+
scale_colour_manual(values=c(
"#D73027","#FC8D59","#FEE08B","#D9EF8B", "#91CF60","#1A9850", "#000000")) +
scale_linetype_manual(values=c(
"dotted", "solid", "solid", "solid",
"solid", "solid", "solid", "dashed")) +
coord_cartesian(ylim=c(0.703,0.803), xlim=c(3,18)) + scale_y_continuous(breaks=seq(0,1,0.01))
@
\caption{{\small
Mean performance of MAGeT brain with varying atlas library size and number of
template labels fused. Also shown is the mean basic multi-atlas
performance when using an atlas library of 17 images and varying the number of
labels fused, as well as the mean performance of single-atlas segmentations.
Data is fit with LOESS local regression smoothing. One standard deviation is
shown in grey.
}}
\label{results}
\end{figure}
\section{Discussion}
In this paper, we have demonstrated that accurate segmentations can be produced
by automatically deriving a template library from a small set of input atlases.
MAGeT brain segmentations were compared to both single-atlas segmentations and
multi-atlas segmentations with cross-correlation voting. For the IBSR dataset,
on average MAGeT brain achieves within 2.0\% of the segmentation accuracy of
the multi-atlas approach but requires only 8 atlases as compared to using the
entire IBSR manually segmented library of 17 atlases.
L\"{o}tj\"{o}nen et al. \citep{Lotjonen2010} report a Kappa of 0.814 on
hippocampal segmentation in the IBSR dataset using a multi-atlas approach where
atlas selection is based on the similarity between atlas and unlabeled subject
after nonlinear registration, and using STAPLE \citep{Warfield2004} for label
fusion. Discrepancies between our results and the above results may be due to
choice of registration algorithm, regularization parameters, or similarity
metric for label fusion. Performance may also be affected by the variability
in the IBSR data set (as previously noted by \citep{Klein2009}). Future work
from our group will attempt to address some of these issues.
\section{Acknowledgements}
Computations were performed on the gpc supercomputer at the SciNet HPC
Consortium \citep{Loken2010}. SciNet is funded by: the Canada Foundation for
Innovation under the auspices of Compute Canada; the Government of Ontario;
Ontario Research Fund - Research Excellence; and the University of Toronto.
This work was supported by the Canadian Institutes of Health Research (CIHR),
National Alliance for Research on Schizophrenia and Depression (NARSAD),
Ontario Mental Health Foundation (OMHF), and the CAMH Foundation (Koerner New
Scientist Program and Paul Garfinkel New Investigator Catalyst Fund).
\bibliographystyle{IEEEbib}
\bibliography{references}
\end{document}