122

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 54, NO. 1, JANUARY 2007

Atlas-Based Fuzzy Connectedness Segmentation

and Intensity Nonuniformity Correction
Applied to Brain MRI
Yongxin Zhou and Jing Bai*, Fellow, IEEE

AbstractA framework that combines atlas registration, fuzzy

connectedness (FC) segmentation, and parametric bias field
correction (PABIC) is proposed for the automatic segmentation
of brain magnetic resonance imaging (MRI). First, the atlas is
registered onto the MRI to initialize the following FC segmentation. Original techniques are proposed to estimate necessary
initial parameters of FC segmentation. Further, the result of
the FC segmentation is utilized to initialize a following PABIC
algorithm. Finally, we re-apply the FC technique on the PABIC
corrected MRI to get the final segmentation. Thus, we avoid expert
human intervention and provide a fully automatic method for
brain MRI segmentation. Experiments on both simulated and real
MRI images demonstrate the validity of the method, as well as
the limitation of the method. Being a fully automatic method, it
is expected to find wide applications, such as three-dimensional
visualization, radiation therapy planning, and medical database
construction.
Index TermsAtlas-based segmentation, bias field correction,
brain MRI, fuzzy connectedness.

I. INTRODUCTION
N brain magnetic resonance imaging (MRI), the segmentation of brain tissues is an important first step for numerous
applications including but not limited to surgical planning,
radiation therapy planning, and three-dimensional (3-D) visualization. Manual segmentation by radiologists is reliable, but
with no doubt tedious and time-consuming. There are strong
demands to accomplish this task automatically by computer.
However, several problems associated with the MRI images
have to be addressed first, including noise, intensity nonuniformity (INU) artifact, partial volume effect, as well as anatomical
variability across individuals. Remarkably, these problems do
not put too much difficulty on human segmentation. We believe
that experts extensive anatomical knowledge play a key role
in his success. This inspires us to complement the classical

Manuscript received August 11, 2005; revised May 21, 2006. This work was
supported in part by the China National Nature Science Foundation, in part by
the Tsinghua-Yue-Yuen Medical Science Foundation, and in part by the Specialized Research Fund for the Doctoral Program of Higher Education (SRFDP).
Asterisk indicates corresponding author.
Y. Zhou is with the Department of Biomedical Engineering, Tsinghua University, Beijing 100084, China (e-mail: zhouyongxin@263.net).
*J. Bai is with the Department of Biomedical Engineering, Tsinghua University, Beijing 100084, China (e-mail: deabj@tsinghua.edu.cn).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TBME.2006.884645

solely image-based algorithms with an atlas, which leads to the

strategy of atlas-based segmentation.
In the strategy of atlas-based segmentation, one intuitive way
is to view segmentation as registration. The basic tenet is that
a transformation can be found, to transfer presegmented labels
in the atlas onto the subject [1][4]. However, due to strong
anatomical variations and limited transformation freedoms, segmentations relying exclusively on registration can not always
satisfy our demand [5], [6]. Thus, it may be necessary to complement the registration with some kind of segmentation techniques. Digital brain atlas containing prior expectations about
the spatial location of tissue classes is used to initialize segmentation algorithms [7], [8]. The construction of statistical atlas is
not an easy work, and is often specially designed for the following specific segmentation methods.
The fuzzy connectedness framework was first proposed in
[9] and further extended in [10] and [11], and a tutorial on FC
methods appears in [12]. The FC works on voxel basis, and
therefore can better segment objects with irregular or complex
shapes, which is of special value in brain MRI. The FC method
takes simultaneously into consideration the degree of space adjacency, degree of intensity adjacency, and degree of intensity
gradient adjacency between two voxels. This comprehensive
consideration endows the FC method with distinguished ability
in medical image segmentation.
Being intensity based method, the FC framework is resistant
to some degree of INU artifact in MRI. But the FC framework
tends to be affected by severe INU in MRI. In these cases, some
kind of INU correction method is necessary to improve the accuracy of the FC method. The parametric bias field correction
(PABIC) method proposed in [13] is adopted to estimate and
correct the INU in MRI in this paper.
In the applications of FC method and the PABIC method, the
initialization of necessary parameters is a question need to be
solved first. Human interaction is a commonly adopted method
to assign these parameters, which adds additional burdens to
operators. In this paper, we have been interested in utilizing
a presegmented atlas (VIP-Man in [14]) to initialize both the
methods. Compared with statistical atlas, the construction of
presegmented atlas is relative easy, which requires the segmentation of only one subject. Further, we combine the FC method
and PABIC method together to complete the segmentation of
brain MRI. Four steps are included in the algorithm.

0018-9294/$25.00 2007 IEEE

ZHOU AND BAI: ATLAS-BASED FC SEGMENTATION AND INTENSITY NONUNIFORMITY CORRECTION APPLIED TO BRAIN MRI

123

Fig. 1. Flowchart of the algorithm.

1) Atlas registration:

2) FC segmentation:

3) PABIC correction:

4) Re-FC segmentation:

The presegmented atlas is registered

onto the MRI through a rigid registration
method.
The FC method is utilized to segment the
MRI for the first time. With the registered
atlas as the initial segmentation, we
propose a novel method to estimate all
necessary FC parameters automatically
and subject-adaptively.
The PABIC method is adopted to estimate
and correct the INU artifacts in MRI. The
FC-segmented MRI is taken as the initial
segmentation to initialize the PABIC
method.
The FC method is applied again, while
taking the PABIC corrected MRI as the
subject, and the FC-segmented MRI as the
initial segmentation

Fig. 1 shows the relationship among the four steps.

The above described algorithm is elaborated upon in
Section II. Section III contains the experimental validation of
the algorithm, and finally some discussions and conclusions are
drawn in Section IV.
II. METHOD
A. Atlas Registration
The goal of atlas registration is to eliminate the overall position and scale differences between the atlas and the MRI. Four
concepts are entailed: normalized mutual information as the
similarity measure, similarity transform, Powell optimization,
and nearest neighbor interpolation.
One special thing is that the VIP-Man atlas is not an intensity
image but an integer-labeled image, that is, all voxels belong to a
specific tissue or organ are labeled with a unique integer. However, the normalized mutual information [15] as the similarity
measure still holds valid here. In other words, the integer-labeled atlas can also be viewed as an intensity image, which has a
relative small intensity range. Obviously, to align the atlas onto
the subject perfectly, a nonrigid deformation is needed. However, our purpose in this step is to provide a good starting point
for the following FC segmentation, not to give a perfect alignment. So we choose similarity transformation, which consists of
9 freedom degrees, three for translation, three for rotation, and
three for scaling. Although we loose some flexibility in similarity transformation, we gain considerable computational efficiency and optimization robust. Meanwhile, we refer to the

Fig. 2. Segmentation result of Case III. (a) Simulated MRI. (b) Registered
atlas. (c) FC segmentation. (d) PABIC corrected MRI. (e) Re-FC segmentation.
(f) Phantom. Comparing (b), (c) and (e), we can see improvements in segmentation. Comparing (e) and (f), we can see that segmentation result of Re-FC is
close to the phantom.

following FC segmentation to cover for the weakness of similarity transformation. Considering the atlass integer labeled nature, the nearest neighbor interpolation is appropriate. Powell
optimization algorithm is utilized to find the optimal similarity
transformation, which maximizes the normalized MI.
Note that after registration, the registered atlas would have the
same dimensions and resolutions with the MRI. The registered
atlas and the MRI are in voxel-wise correspondence.
B. FC Segmentation
The registered atlas can be viewed as an initial segmentation
of MRI. However, there are apparent misalignments between
them [see Fig. 2(a) and (b)]. Instead of referring to other complex techniques to deform the atlas further, we utilize the FC
framework to segment brain tissues directly.
Before the FC method can do its job, some necessary prior
information of tissues has to be specified, including: 1) the intensity probability distribution (IPD); 2) the intensity difference
probability distribution (IDPD); 3) initial seeds. As one of our
innovations, we employ the registered atlas to specify all these
inputs automatically and subject-adaptively.
The registered atlas, denoted as , delineates initial regions
for each brain tissue in MRI, denoted as .
represents the

124

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 54, NO. 1, JANUARY 2007

integer label of voxel in , and

represents the intensity
value of voxel in . The initial regions of tissue in
is
classified as
(1)

Let

in (7), we get
(11)

Combining and (6) and (11),, we can get

where
, 2, 3 represents cerebrospinal fluid (CSF), gray
matter (GM), and white matter (WM), respectively, and represents a voxel.
contains voxels
Due to the discrepancy between and ,
can be divided into
from other tissues. So

(2)
Let

and

be the IPD of

and

, respectively
(3)

(4)

(12)
From (12), we can get six equations, in which there are six
can be solved. Then
can be calculated
unknowns. So
from (7).
from (6). And we can finally get
From the initial IPD
to the approximate IPD
,
we intend to conquer the interference of wrong included voxels
in the initial region . It must be pointed out that the validness
of (12) is related to the extent of tissue intensity overlapping.
Section III will give detailed discussion on this.
be the intensity difference histogram of , that
Let
is,
is the number of 6-adjacent voxel pairs whose inis computed as
tensity difference is . The IDPD of

where
and
are the intensity frequency of voxels
and
, respectively, and denotes an intensity value.
in
It can be derived that
(5)
where
(6)
in (5) with a unified
For simplification, we replace
and we will get
approximate IPD

(13)

Since all the three tissues are approximately homogeneous,

would be close to the true IDPD of tissue , in spite of
the inclusion of voxels from other tissues. So we take
as the IDPD of tissue directly.
gives the probability of a voxel with intensity value
belonging to tissue . So we can specify initial seeds for tissue
according to this. Seeds of tissue are specified according
to the following rules: 1) they should lie in tissue s initial
region (that is, ); 2) the seed itself should have a large
;
3) neighboring voxels of the seed should have a large overall
. That is,

(7)
can be viewed as
Equation (7) says that the initial IPD
a weighted combination of all tissues approximate IPDs.
If we can get
, then we can get
. Let
be the
intensity value which maximize tissue s IPD
(8)
Then we make the assumption that

and

(14)

represents the intensity value of voxel in MRI,

where
and represents neighboring voxels of . and are predefined thresholds.
In our method, the following fuzzy affinity function between
voxel and is utilized [10]P

(9)
(15)

That is, despite intensity overlapping between tissues in MRI,

we assume that a tissue would not span its intensity range to the
position of other tissues maximum IPD.
In most cases, the highest peak of
is very close to that
of
. So
can be found through a search on
,

where
,
, and
are the space adjacency function, intensity feature function, and intensity gradient
feature function, respectively [10]

(10)

(16)
(17)

ZHOU AND BAI: ATLAS-BASED FC SEGMENTATION AND INTENSITY NONUNIFORMITY CORRECTION APPLIED TO BRAIN MRI

For isotropic voxels, we define

computational simplification [10]

as following for

if
are 6 - adjacent,
otherwise.

AR = 1 0
SST

(18)

Note that the space adjacent function can be defined much

more sophisticated than (18). A general principle is that the
closer and are to each other, the greater is the value of their
space adjacent function.
The fuzzy connectedness scene (FCS) of each tissue in MRI
can be computed using a dynamic programming algorithm [9],
, has the
respectively. The FCS of tissue , denoted as
same dimensions with
. Voxel value in
is the
fuzzy connected strength of voxel to the seeds of the tissue .
The FCS of each tissue may assign different values to the
same voxel in MRI. To get the segmented MRI, we let tissues to
compete in having voxels as their members. In this competition,
voxel is considered to belong to tissue if it has the largest
FC strength in tissue . So the FC-segmented MRI, which is an
integer-labeled image, can be constructed as shown in (19) at
the bottom of the page [16].

H k =

be the

=0
L
=0

SSE 1 L
SST 1 (L 0 n 1 3)
Hik () 0 H ik
H k ( ) :
(L + 1)

(23)
(24)

(25)

6) endwhile.

represents the amplitude of the th component,

In (22),
represents the mean value, and
represents the variance.
Among the components, we choose the one with the largest
amplitude to represent tissue , and assign its mean value and
variance to PABIC as the statistics of tissue .
PABIC requires a class mask to separate the regions that are
well described by the class statistics from the rest of the image.
The generation of such a mask is straightforward here.
if
else

C. PABIC Correction
In this step, the FC-segmented MRI, denoted as
in (19), is taken as an initial segmentation. Let
intensity histogram of tissue calculated on

125

(26)

More details about the implementation of PABIC is referred

to [13].
. Note that
We denote the PABIC corrected MRI as
is an intensity image.

(20)
where denotes the number of elements. PABIC is sensitive to
the choice of class statistics [13]. To eliminate the disturbance
and partial volume efof wrong segmented voxels in
, we decompose
into one or more Gaussian
fect in
components, using the following algorithm. In the following alrepgorithm, is the number of Gaussian components, and
resents the adjusted -square, which controls the optimal value
for .
Gaussian Components Decomposition Algorithm
1) Let n = 0, AR = 0, and L be the maximum intensity value
in M (x);
2) While AR < 0:97 and n < 5, do line 3-5
3)
4)

n + 1;
^ k () to Hk () by minimizing the SSE
Fit H

D. Re-FC Segmentation
So far, we get
, which is the result of FC segmentation, and we get
, in which the INU disturbance is reduced. In this step, we run the FC-segmentation algorithm again,
as the initial segmentation, and
while taking
as the subject. Note that as voxel intensities in MRI may be
changed by PABIC correction, tissue IPDs need to be re-calacting as the atlas.
culated with
Different from the FC-segmentation in the second step, we
do not need IPD correction here any more. The IPDs of tissues
can be inputted into the FC algorithm
derived from
directly.

SSE =
H^ k () =

L
=0
n
j =1

Hk () 0 H^ k ()
aj exp

III. RESULTS

0 ( 0cj bj )

(21)
2

(22)

A. Simulated Data Test

In the BrainWeb database (http://www.bic.mni.mcgill.ca/
brainweb), T1-weighted brain data sets (181 217 181
) were generated by the
voxels of 1.0 1.0 1.0

5) Calculate the AR of the current fit

if
else

To validate the performance of our method, we tested it both

on simulated 3-D MR images from the BrainWeb Simulated
Brain Database [17] and on real brain MRI data sets from the
Internet Brain Segmentation Repository (IBSR) [18].

, and

(19)

126

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 54, NO. 1, JANUARY 2007

TABLE I
SIMILARITY INDEXES OF DATA SETS FROM BRAINWEB

TABLE II
OVERLAP RATIOS OF DATA SETS FROM IBSR

Fig. 4. Similarity indexes during iterations of PABIC and Re-FC. Iteration

number 0 represents the FC segmentation. Iteration number 1 represents the
Re-FC segmentation. Iteration number 2 represents the Re-Re-FC segmentation, and so on.

BrainWeb simulator with different noise levels and different

INU levels [17]. We selected the following cases.
Case I)

Fig. 3. Tissue IPDs estimated (solid lines) in the four steps of segmentation
compared with true tissue IPDs (dashed lines). (a) Tissue IPDs in MRI image are
estimated using the registered atlas. (b) Tissue IPDs are modified using the technique proposed in II.B. FC Segmentation. (c) Tissue IPDs in the PABIC-corrected MRI image are estimated using the FC segmentation result. (d) Tissue
IPDs in the PABIC-corrected MRI image are calculated using the Re-FC segmentation result.

T1 weighted images with 3% noise level and 0%

INU level.
Case II) T1 weighted images with 3% noise level and 20%
INU level.
Case III) T1 weighted images with 3% noise level and 40%
INU level.
The percent noise number represents the percent ratio of
the standard deviation of the white Gaussian noise versus the
signal for a reference tissue. For a 20% INU level, the multiover
plicative INU field has a range of values of
the brain area. For other INU levels, the field is linearly scaled
accordingly.

ZHOU AND BAI: ATLAS-BASED FC SEGMENTATION AND INTENSITY NONUNIFORMITY CORRECTION APPLIED TO BRAIN MRI

127

Fig. 5. Overlap ratios of the brain data sets from IBSR.

First, we would like to investigate the segmentation procedure of Case III, to clarify the steps in Section II further. Fig. 2(a)
and (b) shows an example MRI slice and the corresponding slice
from the registered atlas. In spite of the overall alignment, minor
discrepancies can be seen between them. Fig. 3(a) shows the initial IPDs and the true IPDs of tissues in MRI. Note that the initial
IPDs were calculated with the registered atlas as the template,
and the true IPDs were calculated with the phantom provided
by BrainWeb as the template. It can be seen that the initial IPDs
were interfered severely by the discrepancies between the registered atlas and the MRI. Apparently, the initial IPDs contain
voxels which belong to other tissues. After the IPD correction
procedure, the corrected IPDs are drawn in Fig. 3(b). Noticeable
improvements were achieved in the correction. Utilizing the FC
method, we got the FC segmented MRI, displayed in Fig. 2(c).
Compared with the registered atlas, the result of FC segmentation was improved remarkably. With the FC segmented MRI as
the initial segmentation, we utilized the PABIC method to correct INU artifacts. The corrected MRI is showed in Fig. 2(d).
Fig. 3(c) shows the tissue IPDs of the PABIC corrected MRI.
Intensity overlapping among tissues was reduced. This would
increase the accuracy of the Re-FC segmentation. Fig. 2(e) and
(f) shows the comparison of Re-FC segmentation result and the
phantom.
We quoted the similarity index for the quantitative evaluation
of the segmentation result. The similarity index is defined as

(27)
where
denotes the number of voxels classified by both
the proposed method and the phantom as tissue , and
and
represent the number of voxels classified as tissue
by the proposed method and by the phantom, respectively. This
index will reach the maximum value one if the proposed method
perfectly coincides with the phantom, and will decrease as the
quality of the segmentation decline. The quantitative measurements of Case I-III are shown in Table I.
For each case, we calculated the similarity indexes of each
tissue at both the FC segmentation step and the Re-FC segmentation step. We can see that the Re-FC segmentation increases

Fig. 6. Sample images from IBSR. Left: original MRI image; center: segmentation by our method; right: expert hand-guided segmentation.

the similarity indexes in Case III a little more than it dose in

Case I and Case II. We can infer that the Re-FC segmentation
combined with the PABIC correction acts most prominently in
MRI images with high INU levels.
One succeeding question is whether more iterations of PABIC
and Re-FC would improve the results further. We still took Case
III as the example, and applied four iterations of PABIC and
Re-FC on it. Fig. 4 gives the plot of similarity indexes through
the iterations. It can be seen that the similarity indexes fluctuate
about 0.01 from iteration 1 to iteration 4. So more iterations of
PABIC and Re-FC are not worthwhile in practical applications.
B. Real Data Test
As another test, we applied our method on the 20 normal MRI
data sets (256 256 61 voxels of 1 1 3
) provided
by IBSR [18]. In addition, IBSR provides expert hand-guided
segmentations and performance results from five automatic segmentation methods, making it convenient to compare our results
with those reported by the five automatic methods.
We applied our method to the 20 MRI data sets in IBSR.
Prior to segmentation, edge-preserving smoothing using an
anisotropic diffusion algorithm was applied to the datasets [19].
Interslice intensity variations were corrected by a simple intensity histogram correction algorithm, which was able to nearly
match the mean and variance across slices. The results are
summarized in Fig. 5 and Table II. Fig. 6 shows sample images
from the IBSR data sets. The overlap ratio in Table II is defined
as the ratio of the sum of the voxels with the same label in both
the algorithm segmented image and the expert hand-guided
segmented image to the total number of voxels with the same

128

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 54, NO. 1, JANUARY 2007

for brain MRI segmentation. Each of the four steps provides

improvements on the segmentation or on the MRI images.
In FC segmentation, we proposed an originative technique to
depress the negative influence of those wrong classified labels
transferred from the registered atlas. Also, in the initialization
of the PABIC method, we utilize the Gaussian components decomposition to depress the negative influence of partial volume
effects. The central goal is to make automatic estimations of
tissue IPDs, and to avoid human intervention.
In the initialization of FC segmentation, we assume that the
intensity overlapping among tissues should be limited within
some range. However, this assumption may be not true in MRI
images with severe intensity overlapping. The method proposed
may fail in these MRI images. To overcome this shortcoming
and to make the method more robust is the main topic of our
future work.
ACKNOWLEDGMENT
The program of the proposed algorithm is based upon the ITK
software system. (http://www.itk.org)
REFERENCES

Fig. 7. True tissue IPDs of the two data sets from IBSR. (a) IPDs of data set
202_3; (b) IPDs of data set 17_3.

label in either image. The overlap ratios of results in [8] were

included in Table II as an additional comparison. It can be seen
that the GM overlap ratio of our method is significantly higher
than the GM overlap ratio of [8], while the WM overlap ratio of
our method is little lower than that in [8] but higher than those
of other methods.
In Fig. 5 we can see that most overlap ratios are above 0.7,
indicating the validity of the method. We also notice that the
overlap ratios of some data sets are rather low. As two examples,
we investigated the intensity histograms of data sets 202_3 and
17_3. The segmentation of data set 202_3 is satisfying, while the
segmentation of 17_3 failed. Fig. 7 shows the true tissue IPDs
of the two data sets. We can see that the validity of the method is
affected by the range of tissue IPD overlapping in MRI images.
In FC Segmentation, we assumed that a tissue would not span
its intensity range to the position of other tissues maximum
IPD [see (9)]. It can be seen that in Fig. 7(a) this assumption
holds true, while in Fig. 7(b) it does not. So, the IPD estimation
technique proposed in FC segmentation could not give reliable
estimations of true tissue IPDs. Then FC segmentation would
fail, and the following PABIC and Re-FC would fail too.
IV. CONCLUSION AND DISCUSSION
The main contribution of this work is the combination of atlas
registration, FC segmentation, PABIC correction and Re-FC
segmentation together to provide a fully automatic method

[1] B. M. Dawant, S. L. Hartmann, J. P. Thirion, F. Maes, D. Vandermeulen, and P. Demaerel, Automatic 3-D segmentation of internal
structures of the head in MR images using a combination of similarity and free-form transformations. I. Methodology and validation
on normal subjects, IEEE Trans. Med. Imag., vol. 18, no. 10, pp.
909916, Oct. 1999.
[2] T. Rohlfing, R. Brandt, R. Menzel, and C. R. Maurer, Jr, Segmentation of three-dimensional images using non-rigid registration: Methods
and validation with application to confocal microscopy images of bee
brains, presented at the Medical Imaging 2003: Image Processing
Conf., San Diego, CA, 2003.
[3] M. B. Cuadra, O. Cuisenaire, R. Meuli, and J. P. Thiran, Automatic
segmentation of internal structures of the brain in MR images using
a tandem of affine and non-rigid registration of an anatomical brain
atlas, in Proc. 2001 Int. Conf. Image Processing, Thessaloniki,
Greece, 2001, vol. 3.
[4] J. Ehrhardt, H. a. Handels, T. a. Malina, B. b. Strathmann, W. b. Pltz,
and S. J. a. Pppl, Atlas-based segmentation of bone structures to support the virtual planning of hip operations, Int. J. Med. Informatics,
vol. 64, pp. 439447, 2001.
[5] D. Collins and A. Evans, Animal: Automatic nonlinear image
matching and anatomical labeling, in Brain Warping, A. Toga,
Ed. New York: Academic, 1999, pp. 133142.
[6] D. L. Collins, G. Le Goualher, R. Venugopal, A. Caramanos, A. C.
Evans, and C. Barillot, Cortical constraints for non-linear cortical registration, Lecture Notes in Computer Science, vol. 1131, p. 307, 1996.
[7] K. Van Leemput, F. Maes, D. Vandermeulen, and P. Suetens, Automated model-based tissue classification of MR images of the brain,
IEEE Trans. Med. Imag., vol. 18, no. 10, pp. 897908, Oct. 1999.
[8] J. L. Marroquin, B. C. Vemuri, S. Botello, F. Calderon, and A. Fernandez-Bouzas, An accurate and efficient Bayesian method for automatic segmentation of brain MRI, IEEE Trans. Med. Imag., vol. 21,
no. 8, pp. 934945, Aug. 2002.
[9] J. K. Udupa and S. Samarasekera, Fuzzy connectedness and object
definition: Theory, algorithms, and applications in image segmentation, Graphical Models Image Process., vol. 58, pp. 246261, 1996.
[10] P. K. Saha, J. K. Udupa, and D. Odhner, Scale-based fuzzy connected
image segmentation: Theory, algorithms, and validation, Compu. Vi.
Image Understanding, vol. 77, pp. 145174, 2000.
[11] P. K. Saha and J. K. Udupa, Relative fuzzy connectedness among multiple objects: Theory, algorithms, and applications in image segmentation, Comput. Vision Image Understanding, vol. 82, pp. 4256, 2001.
[12] J. K. Udupa and P. K. Saha, Fuzzy connectedness and image segmentation, Proc. IEEE, vol. 91, no. 10, pp. 16491669, Oct. 2003.
[13] M. Styner, C. Brechbuhler, G. Szckely, and G. Gerig, Parametric estimate of intensity inhomogeneities applied to MRI, IEEE Trans. Med.
Imag., vol. 19, no. 3, pp. 153165, Mar. 2000.

ZHOU AND BAI: ATLAS-BASED FC SEGMENTATION AND INTENSITY NONUNIFORMITY CORRECTION APPLIED TO BRAIN MRI

[14] X. G. Xu, T. C. Chao, and A. Bozkurt, VIP-man: An image-based

whole-body adult male model constructed from color photographs of
the visible human project for multi-particle Monte Carlo calculations,
Health Phys., vol. 78, pp. 476486, 2000.
[15] C. Studholme, D. L. G. Hill, and D. J. Hawkes, An overlap invariant
entropy measure of 3D medical image alignment, Pattern Recognit.,
vol. 32, pp. 7186, 1999.
[16] J. K. Udupa, P. K. Saha, and R. A. Lotufo, Disclaimer: Relative fuzzy
connectedness and object definition: theory, algorithms, and applications in image segmentation, IEEE Trans. Pattern Anal. Mach. Intell.,
vol. 24, pp. 14851500, 2002.
[17] C. A. Cocosco, V. Kollokian, R. K.-S. Kwan, and A. C. Evans,
BrainWeb: Online interface to a 3D MRI simulated brain database,
NeuroImage, vol. 5, pp. 24, 1997.
[18] Internet Brain Segmentation Repository. M. G. H. Tech. Rep., Ctr.
Morphometric Anal. IBSR, 2000 [Online]. Available: http://neurowww.mgh.harvard.edu/cma/ibsr/
[19] G. Gerig, O. Kubler, R. Kikinis, and F. A. Jolesz, Nonlinear
anisotropic filtering of MRI data, IEEE Trans. Med. Imag., vol. 11,
no. 2, pp. 221232, Jun. 1992.
Yongxin Zhou received the B.S. degree in automatic
control from Beihang University, Beijing, China, in
1999, and the M.S. degree in biomedical engineering
from Capital University of Medical Sciences, Beijing, China, in 2002. He is currently working toward
the Ph.D. degree in the Department of Biomedical
Engineering, Tsinghua University, Beijing, China.
His research interests include image processing,
pattern recognition.

129

Jing Bai (M87SM92F02) received the M.S. and

Ph.D. degrees from Drexel University, Philadelphia,
PA, in 1983 and 1985, respectively.
From 1985 to 1987, she was a Research Associate and Assistant Professor with the Biomedical
Engineering and Science Institute of Drexel University. In 1988, 1991, and 2000, she become an
Associate Professor, Professor, and Cheung Kong
Chair Professor of Electrical Engineering with
Department of Tsinghua University, Beijing, China.
Her research activities have included mathematical
modeling and simulation of cardiovascular system, optimization of cardiac
assist devices, medical ultrasound, telemedicine, home health care network and
home monitoring devices, and infrared imaging. She has published four books
and over 100 journal papers.
Since 1997, Dr. Bai is an Associate editor for the IEEE TRANSACTIONS ON
INFORMATION TECHNOLOGY IN BIOMEDICINE.