Pulmonary Nodules Segmentation Method Based on Auto-encoder
Guodong Zhang
a,c
, Mao Guo
a
, Zhaoxuan Gong
a
, Jing Bi
a
, Yoohwan Kim
c
, Wei Guo
a,b,*
a
School of Computer Science, Shenyang Aerospace University, Shenyang, Liaoning China 110136;
b
Shenyang Institute of Computing Technology, Chinese Academy of Sciences, Shenyang, Liaoning
China 110168;
c
Department of Computer Science, University of Nevada Las Vegas, Las Vegas, NV
USA 89154
ABSTRACT
In this paper, we proposed a semi-automatic pulmonary nodule segmentation algorithm, which is operated within a
region of interest for each nodule. It mainly includes two parts: the unsupervised training of auto-encoder and the
supervised training of segmentation network. Applying an auto-encoder's unsupervised learning, we obtain a feature
extractor that consists of its encoded part. Through adding some new neural network layers behind the feature extractor
and do supervised learning on it, we get the final segmentation neural network. Compared with the traditional maximum
two-dimensional entropy threshold segmentation algorithm, the dice correlation coefficient of this algorithm is 1% - 9%
higher in 36 regions of interest segmentation experiments.
Keywords: Pulmonary nodule segmentation, auto-encoder, feature extraction, medical image process
1. INTRODUCTION
Lung cancer has become a major tumor disease that threatens the survival of humans. According to the reliable medical
research, it shows that one of the early symptoms of lung cancer is the production of pulmonary nodules
1
. Based on this,
the detection of pulmonary nodules is essential for the detection and diagnosis of lung cancer. Furthermore, the growth
rate of lung nodules is crucial for the judgment of lung cancer. Accordingly, the premise of knowing the exact growth
rate for pulmonary nodules is to obtain the area or volume of the pulmonary nodules. Therefore, there is a growing
demand for segmentation of pulmonary nodules.
Computed Tomography (CT) images are widely used for lung tissue segmentation because of their high resolution and
isotropy CT. With CT scanning slices, we can get a lot of human tissue information, and through the extracted features,
image segmentation and other operations, we can provide physicians with a qualitative or quantitative reference for
application. As a result, it is convenient for doctors to conduct medical diagnosis and helps them avoid misdiagnosis or
frequent misdiagnosis.
As for the segmentation of pulmonary nodules, scholars put forward many related methods, including threshold method
2
,
variable model method
3
, clustering method
4
, and graph cut method
5
and so on. These methods have different advantages
and disadvantages for pulmonary nodule segmentation; however, they all have a common disadvantage that they all need
experts’ intervention. Human scientists are required to select features and design corresponding algorithms. The
superiority of these traditional segmentation methods relies heavily on the selection of corresponding features.
In order to reduce the impact of feature representation on the result, a large number of machine learning methods have
been applied to pulmonary nodule segmentation. Ahuja employed a dynamic programming method to pulmonary nodule
segmentation
6
, but it required a physician specified point. Although Antonelli et al.
7
described a method to segment the
entire CT image using the robust fuzzy C-means algorithm, it required the use of empirical morphological knowledge for
false positive removal. Along with the advent of deep learning
8
in 2006, the applications of deep neural networks have
become more and more popular. The main application of deep neural networks in computer vision is to extract the
features of graphic images. There is a successful application of convolutional neural network in pulmonary nodule
segmentation
9
, but all of these methods need a huge of labeled data.
We proposed an auto-encoder-based neural network which requires only a small amount of label data for model training.
With the exception of the nodule's region of interests (ROIs), it does not require any empirical knowledge for nodule
segmentation.
*zhanggd@sau.edu.cn and guowei@sau.edu.cn; phone (86)24-8972-3471