1558 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 64, NO. 7, JULY 2017
Multilevel Contextual 3-D CNNs for False
Positive Reduction in Pulmonary Nodule
Detection
Qi Dou
∗
, Student Member, IEEE, Hao Chen, Student Member, IEEE, Lequan Yu, Student Member, IEEE,
Jing Qin, Member, IEEE, and Pheng-Ann Heng, Senior Member, IEEE
Abstract— Objective: False positive reduction is one of
the most crucial components in an automated pulmonary
nodule detection system, which plays an important role
in lung cancer diagnosis and early treatment. The objec-
tive of this paper is to effectively address the challenges
in this task and therefore to accurately discriminate the
true nodules from a large number of candidates. Methods:
We propose a novel method employing three-dimensional
(3-D) convolutional neural networks (CNNs) for false pos-
itive reduction in automated pulmonary nodule detection
from volumetric computed tomography (CT) scans. Com-
pared with its 2-D counterparts, the 3-D CNNs can encode
richer spatial information and extract more representative
features via their hierarchical architecture trained with 3-
D samples. More importantly, we further propose a simple
yet effective strategy to encode multilevel contextual infor-
mation to meet the challenges coming with the large varia-
tions and hard mimics of pulmonary nodules. Results: The
proposed framework has been extensively validated in the
LUNA16 challenge held in conjunction with ISBI 2016, where
we achieved the highest competition performance metric
(CPM) score in the false positive reduction track. Conclu-
sion: Experimental results demonstrated the importance
and effectiveness of integrating multilevel contextual infor-
mation into 3-D CNN framework for automated pulmonary
nodule detection in volumetric CT data. Significance: While
our method is tailored for pulmonary nodule detection, the
proposed framework is general and can be easily extended
to many other 3-D object detection tasks from volumetric
medical images, where the targeting objects have large vari-
ations and are accompanied by a number of hard mimics.
Index Terms—Computer-aided diagnosis, deep learning,
false positive reduction, pulmonary nodule detection, 3-D
convolutional neural networks.
Manuscript received May 21, 2016; revised August 4, 2016; accepted
September 12, 2016. Date of publication September 26, 2016; date
of current version June 15, 2017. This work was supported in part by
the Research Grants Council of The Hong Kong Special Administra-
tive Region under Project CUHK 412513, in part by the National Nat-
ural Science Foundation of China under Project 61233012, and in part
by the Shenzhen-Hong Kong Innovation Circle Funding under Program
SGLH20131010151755080 and Program GHP/002/13SZ.Asterisk indi-
cates corresponding author.
*
Q. Dou, H. Chen, L. Yu, and P. A. Heng are with the Department of
Computer Science and Engineering, The Chinese University of Hong
Kong, Hong Kong(e-mail: qdou@cse.cuhk.edu.hk).
J. Qin is with the Centre for Smart Health, School of Nursing, The
Hong Kong Polytechnic University.
Digital Object Identifier 10.1109/TBME.2016.2613502
I. INTRODUCTION
A
UTOMATED detection of pulmonary nodules in volumet-
ric thoracic computed tomography (CT) scans plays an im-
portant role in computer-aided lung cancer diagnosis and early
treatment [1]–[3]. The pulmonary nodules are radiologically
visible as small structures that are roughly spherical opacities
within the pulmonary interstitium images [4]. They have been
regarded as crucial indicators of primary lung cancer, which has
been the leading cause of cancer death in recent years [5]. Based
on reliable detection of lung nodules, radiologists and surgeons
can perform size measurements and appearance characteriza-
tions for cancer malignancy diagnosis [6] and, if necessary,
timely surgical intervention in order to increase the survival
chances of patients [7], [8].
An automated pulmonary nodule detection system mainly
consists of two steps: 1) candidate screening and 2) false pos-
itive reduction. In candidate screening, a considerable number
of coarse candidates are rapidly screened throughout the whole
volume using a variety of criteria, e.g., intensity thresholding,
shape curvedness, and mathematical morphology [3], [9], [10].
In false positive reduction, effective classifiers together with
discriminative features are developed to reduce a large number
of false positive candidates. In order to maintain a high sen-
sitivity in candidate screening, the criteria employed in this
step are usually quite straightforward and lenient, and con-
sequently a great number of candidates are s elected out and
forwarded to the second step. In this regard, the false posi-
tive reduction stands as the most crucial component of an au-
tomated pulmonary nodule detection system [ 1] and a lot of
efforts have been dedicated to improving the performance of
this step.
Automated identification of the pulmonary nodules from tho-
racic CT scans is, however, among the most challenging tasks
in computer-aided chest radiograph analysis [11] for at least
the following two reasons. First, the pulmonary nodules have
large variations in sizes, shapes, and locations, as shown in the
green rectangle in Fig. 1. Moreover, the contextual environ-
ments around them are often diversified for different categories
of lung nodules, such as solitary nodules, ground-glass opacity
nodules, cavity nodules, and pleural nodules [12]. Second, some
false positive candidates carry quite similar morphological ap-
pearance to the true pulmonary nodules, as shown in the red
0018-9294 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications
standards/publications/rights/index.html for more information.