490 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 6, JUNE 2001
Automatic Lung Segmentation for Accurate
Quantitation of Volumetric X-Ray CT Images
Shiying Hu, Eric A. Hoffman, Member, IEEE, and Joseph M. Reinhardt*, Member, IEEE
Abstract—Segmentation of pulmonary X-ray computed tomog-
raphy (CT) images is a precursor to most pulmonary image anal-
ysis applications. This paper presents a fully automatic method for
identifying the lungs in three-dimensional (3-D) pulmonary X-ray
CT images. The method has three main steps. First, the lung re-
gion is extracted from the CT images by gray-level thresholding.
Then, the left and right lungs are separated by identifying the an-
terior and posterior junctions by dynamic programming. Finally, a
sequence of morphological operations is used to smooth the irreg-
ular boundary along the mediastinum in order to obtain results
consistent with those obtained by manual analysis, in which only
the most central pulmonary arteries are excluded from the lung
region.
The method has been tested by processing 3-D CT data sets from
eight normal subjects, each imaged three times at biweekly inter-
vals with lungs at 90% vital capacity. We present results by com-
paring our automatic method tomanually traced borders from two
image analysts. Averaged over all volumes, the root mean square
difference between the computer and human analysis is 0.8 pixels
(0.54 mm). The mean intrasubject change in tissue content overthe
three scans was 2.75%
2.29% (mean standard deviation).
Index Terms—Image segmentation, pulmonary imaging, volu-
metric imaging, X-ray CT.
I. INTRODUCTION
H
IGH-RESOLUTION X-ray computed tomography (CT)
is the standard for pulmonary imaging. Depending on the
scanner hardware, CT can provide high spatial and high tem-
poral resolution, excellent contrast resolution for the pulmonary
structures and surrounding anatomy, and the ability to gather a
complete three-dimensional (3-D) volume of the human thorax
in a single breath hold [1]. Pulmonary CT images have been
used for applications such as lung parenchyma density analysis
[2], [3], airway analysis [4], [5], and lung and diaphragm me-
chanics analysis [6], [7]. A precursor to all of these quantitative
analysis applications is lung segmentation. With the introduc-
tion of multislice spiral CT scanners, the number of volumetric
studies of the lung is increasing and it is critical to develop fast,
Manuscript received June 30, 2000; revised March 20, 2001. This work
was supported in part by a Biomedical Engineering Research Grant from the
Whitaker Foundation and by the National Institutes of Health (NIH) under
Grant HL64368-01 and Grant HL60158-02. The Associate Editor responsible
for coordinating the review of this paper and recommending its publication
was A. Manduca. Asterisk indicates corresponding author.
S. Hu is with the Department of Biomedical Engineering, University of Iowa,
Iowa City, IA 52242 USA.
E. A. Hoffman is with the Department of Radiology, University of Iowa, Iowa
City, IA 52242 USA.
*J. M. Reinhardt is with the Department of Biomedical Engineering, Univer-
sity of Iowa, Iowa City, IA 52242 USA (e-mail: joe-reinhardt@uiowa.edu).
Publisher Item Identifier S 0278-0062(01)04688-2.
accurate algorithms that require minimal to no human interac-
tion to identify the precise boundaries of the lung.
A number of groups havedevelopedtechniques for computer-
assisted segmentation of pulmonary CT images [2], [8]–[14].
In [10], manually traced boundaries were used to estimate re-
gional gas and tissue volumes in the lungs of normal subjects.
But manual methods are laborious and subject to both interob-
server and intraobserver variations. On two-dimensional (2-D)
transverse slices of a pulmonary CT dataset, the natural contrast
between the low-density lungs and the surrounding high-den-
sity chest wall can be used to guide image segmentation. In
[2], [8], and [9], 2-D edge tracking was used to find the bound-
aries of the left and right lungs. Others have used 3-D region
growing with manually specified seed points [2], [12]–[14]. In
many semi-automatic approaches, some manual interaction is
required to select threshold values or edit the resulting segmen-
tation [2], [8], [9], [12]–[14]. In [8], anterior and posterior junc-
tion lines are provided manually to separate the right and left
lungs in the case where the edge contrast is reduced by the
volume averaging. More recently, Brown et al. [11] provided
a knowledge-based, automatic method to segment chest CT im-
ages. In their method, anatomic knowledge stored in a semantic
network is used to guide low-level image processing routines.
Rather than requiring manual intervention to define the anterior
junction lines as in [8], Brown et al. used dynamic programming
to search for the junction lines automatically.
In this paper, we describe a fully automatic method for iden-
tifying the lungs in CT images. The method has three main
steps. First, the lung region is extracted from the CT images by
gray-level thresholding. The left and right lungs are then sepa-
rated by detecting the anterior and posterior junctions. Finally,
we optionally smooth the lung boundary along the mediastinum.
There are several distinctions between our method and previous
work. First, instead of using a fixed threshold value, we use
an optimal thresholding method [15] to automatically choose
a threshold value that reflects the gray-scale characteristics of
a specific dataset. Second, we use an efficient method to find
the anterior and posterior junction lines between the right and
left lungs. Finally, to obtain more consistent results across time
and to leave lung structures with the lung, we optionally smooth
the irregular boundary along the mediastinum. We present re-
sults comparing our automatic method to manually traced bor-
ders from two image analysts. We have compared the automatic
method to the manual analysis for a total of 12 3-D volumes. Av-
eraged over all volumes, the root mean square (rms) difference
between the computer and human analysis is about 0.8 pixels
(corresponding to about 0.54 mm). By studying the same sub-
ject repeatedly over a short time interval (1–2 weeks), we are
0278–0062/01$10.00 ©2001 IEEE