Professional Documents
Culture Documents
fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2014.2328870, IEEE Journal of Biomedical and Health Informatics
TITB-00731-2013-JBHI 1
for the United States alone in 2013, and the overall 5-year
Abstract—Computer-aided detection (CADe) of pulmonary survival rate for lung cancer is merely 16%. The survival rate
nodules is critical to assisting radiologists in early identification of increases to 52% if it is localized, and decreases to 4% if it has
lung cancer from computed tomography (CT) scans. This paper metastasized. Therefore, to detect lung cancer at earlier stages
proposes a novel CADe system based on a hierarchical vector
is of great importance [2], and computer-aided detection
quantization (VQ) scheme. Compared with the commonly-used
simple thresholding approach, the high-level VQ yields a more (CADe) in supplement to radiologists’ diagnosis has become a
accurate segmentation of the lungs from the chest volume. In promising tool to serve such purposes [3].
identifying initial nodule candidates (INCs) within the lungs, the While detection of pulmonary nodules has a crucial effect
low-level VQ proves to be effective for INCs detection and on the diagnosis of lung cancer, but the detection is a nontrivial
segmentation, as well as computationally efficient compared to task, not only because the appearance of pulmonary nodules
existing approaches. False-positive (FP) reduction is conducted
varies in a wide range, but also because the nodule densities
via rule-based filtering operations in combination with a
feature-based support vector machine classifier. The proposed have low contrast against adjacent vessel segments and other
system was validated on 205 patient cases from the publically lung tissues. Computed tomography (CT) has been shown as
available on-line LIDC (Lung Image Database Consortium) the most popular imaging modality for nodule detection [2],
database, with each case having at least one juxta-pleural nodule [4], because it has the ability to provide reliable image textures
annotation. Experimental results demonstrated that our CADe for the detection of small nodules. The development of lung
system obtained an overall sensitivity of 82.7% at a specificity of 4
nodule CADe systems using CT imaging modality has made
FPs/scan. Specifically for the performance on juxta-pleural
nodules we observed 89.2% sensitivity at 4.14 FPs/scan. With good progress over the past decade [5], [6]. Generally, such
respect to comparable CADe systems, the proposed system shows CADe systems consist of three stages: (1) image preprocessing,
outperformance and demonstrates its potential for fast and (2) initial nodule candidates (INCs) identification, and (3) false
adaptive detection of pulmonary nodules via CT imaging. positive (FP) reduction of the INCs with preservation of the
true positives (TPs).
Index Terms—Computer-aided detection, CT imaging, false In the preprocessing stage, the system aims to largely
positive reduction, lung nodules, vector quantization.
reduce the search space to the lungs, where a segmentation of
the lungs from the entire chest volume is usually required.
I. INTRODUCTION Because of the high image contrast between lung fields and the
surrounding body tissue, image intensity-based simple
CCORDING to the up-to-date statistics from American thresholding is effective, and is currently the most commonly
A Cancer Society [1], lung cancer is the leading cause of
cancer-related deaths with over 159,000 deaths estimated
used technique for lung segmentation [7], [8]. However, the
determination of an accurate threshold is greatly affected by
image acquisition protocols, scanner types, as well as the
This work was supported in part by the NIH/NCI under grants #CA082402 inhomogeneity of intensities in the lung region, especially
and #CA143111. towards the segmentation of pathological lungs with severe
Hao Han is with the Department of Radiology, Stony Brook University, pathologies [9], [10]. This work proposes an adaptive solution
Stony Brook, NY 11794 USA (e-mail: haohan@mil.sunysb.edu).
Lihong Li is with the Department of Engineering Science and Physics, to mitigate the difficulty of thresholding-based method in lung
College of Staten Island of The City University of New York, Staten Island, NY segmentation.
10314 USA. After defining the search space (i.e., the lung volume),
Fangfang Han is with the Northeastern University, Shenyang, Liaoning
110819 China. INCs detection is the next step to build a CADe system.
Bowen Song is with the Department of Radiology, Stony Brook University, Various INCs detection techniques have been extensively
Stony Brook, NY 11794 USA. studied in recent years, such as multiple thresholding [11], [12],
William Moore is with the Department of Radiology, Stony Brook
University, Stony Brook, NY 11794 USA. nodule enhancement filtering [13], [14], mathematical
* Zhengrong Liang is with the Department of Radiology, Stony Brook morphology [15], [16], and genetic algorithm template
University, Stony Brook, NY 11794 USA (telephone: 631-444-2508, e-mail: matching [17], [18], among many others. The most commonly
jerome.liang@stonybrookmedicine.edu).
2168-2194 (c) 2013 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2014.2328870, IEEE Journal of Biomedical and Health Informatics
TITB-00731-2013-JBHI 2
2168-2194 (c) 2013 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2014.2328870, IEEE Journal of Biomedical and Health Informatics
TITB-00731-2013-JBHI 3
2168-2194 (c) 2013 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2014.2328870, IEEE Journal of Biomedical and Health Informatics
TITB-00731-2013-JBHI 4
6
x 10 Chest Body Volume Gray Level Distribution
3.5
Fat
3
Number of Occurences
2.5
Muscle
2
(a) (b) (c)
1.5 Air Fig. 4. Comparison of simple thresholding and VQ for lung segmentation.
Panel (a) shows a raw CT image with dense pathologies on the left lung lobe.
1
Panel (b) illustrates the lung mask obtained by simple thresholding, and panel
(c) is the lung mask obtained by the two-class high level VQ, which is robust
to intensity inhomogeneity.
0.5
Bone
2168-2194 (c) 2013 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2014.2328870, IEEE Journal of Biomedical and Health Informatics
TITB-00731-2013-JBHI 5
5 High-level VQ
Low-Intensity Class
Number of Occurrences
4
Connected Component Analysis
3 Initial Lung Mask
Morphological Closing
2
Lung Volume
1
Low-level VQ
0
-1000 -400 -800-200-600 0 200 400
Gray-level in HU INCs
Fig. 5. Histogram of lung voxel intensities from one chest CT scan of the LIDC
database. Each bin corresponds to the frequency of observing voxels having Fig. 6. Flowchart of the proposed hierarchical VQ scheme for INCs detection.
intensity within a specific range. Statistically, the observed distribution can be
decomposed into four normal-tailed Gaussian mixtures as represented by the
four bell-shaped curves.
R2 = max(dx, dy, dz) / min(dx, dy, dz). (4)
In addition, because pulmonary nodules are usually compact in
C. False Positive Reduction from INCs shape, we define the following compactness rule:
R3 = (dx * dy * dz) / {[max(dx, dy, dz)]^3}. (5)
1) Rule-based Filtering Operations In Eqs. (4) and (5), dx, dy, and dz are the maximal projection
It is challenging to thoroughly separate nodules from lengths (in mm) of an INC object along X, Y, and Z axes,
attached structures due to their similar intensities, especially for respectively. Rules defined in Eqs. (3)-(5) together can capture
the juxta-vascular nodules (the nodules attached to blood the major shape characteristics of pulmonary nodules. In this
vessels). Since the thickness of blood vessels varies study, the threshold values for R1, R2, and R3 were determined
considerably (e.g., from small veins to large arteries), a 2D experimentally. Given the true nodule (or TP) annotations in
morphological opening disk with radius of 1 up to 5 pixels was the studied dataset, we can compute the above three shape
adopted to detach vessels at different degrees. Of note, 2D features for each true nodule. Then based on the range of each
rather than 3D opening operation is favored here because of the feature we can set a conservative threshold for each filtering
anisotropic nature of the LIDC data, where 2D operations have rule. Combining these three expert rules, a candidate will be
been found to outperform 3D operations for relatively treated as a false detection and excluded from further analysis if
thick-slice data [11]. Since vessels in lung parenchyma appear the following expression is true:
to have various radii, opening disks of different radii were R1 < 3.0 || R1 > 30.0 || R2 > 3.0 || R3 < 0.1. (6)
adopted to treat varying levels of vessel attachment. In order to
keep small nodules while removing attached vessels, the lower 2) Feature-based SVM Classification
bound of opening disk radius was set to be the smallest -- one A supervised learning strategy is carried out using the
pixel. The upper bound on disk size was experimentally SVM classifier to further reduce FPs. Our feature-based SVM
determined by examining the sensitivity for annotated nodules classifier relies on a series of features extracted from each of the
in the studied dataset.
remaining INC after rule-based filtering operations. Table I
The rule-based filtering operation was separately
lists the definitions of extracted features by four categories,
conducted for INCs obtained at each opening level, as well as
including geometric or shape features, intensity features,
for the original INCs without opening operation to preserve
gradient features, and Hessian eigenvalue based features. Both
small solitary nodules. The final INCs after filtering operations
2D and 3D features are extracted for the first three groups of
were formed by a logical union operation of the filtered INCs at
features. It is noted that the image resolution difference is taken
all levels.
into account for the calculation of all related features. All 2D
For the INCs obtained at each opening level, obvious FPs are
features are extracted from the largest area slice of INC
first filtered out using the size rule on volume-equivalent
segmentation, and the 2D segmentation at maximum area slice
diameter, which is calculated by
is dilated for 5 layers to obtain the “outside” region for 2D
R1 = [(INC voxels * voxel volume) / (4π/3))^(1/3). (3)
feature computation. For 3D features, the “outside” region is
Because blood vessels are elongated in shape, we also establish
defined by dilating the 3D INC segmentation using a spheroidal
the elongation rule to exclude vessel-like structures with large
elongation values. And the elongation in 3D space is defined by
2168-2194 (c) 2013 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2014.2328870, IEEE Journal of Biomedical and Health Informatics
TITB-00731-2013-JBHI 6
TABLE I 120
OVERVIEW OF FEATURES FOR SVM CLASSIFICATION
100
No. Feature Category
Number of Occurences
80
1 Area 2D geometric
2 Diameter (max 2D bounding box length) 2D geometric 60
3 Eccentricity = ellipse foci / major axis length 2D geometric
4 Circularity = Area/[π(Diameter/2)^2] 2D geometric
40
5 Volume 3D geometric
6 Elongation = max(dx, dy, dz) / min(dx, dy, dz) 3D geometric
7 Compact1 = projection on x-y plane / (dx*dy) 3D geometric 20
8 Compact2 = (dx*dy*dz) / [max(dx, dy, dz)]^3 3D geometric
9-10 Mean and stdev of 2D square compactness 3D geometric 0
0 5
15 10
20 25 30 35
11-16 Min, max, mean, stdev, skewness, and kurtosis 2D intensity Nodule Size (mm)
17-18 Inside vs. outside mean and stdev separation 2D intensity
Fig. 7. The size (volume-equivalent diameter) distribution of the nodules in
19 Inside vs. outside contrast 2D intensity
the 205 datasets of this study.
20-23 Mean, stdev, skewness, and kurtosis 3D intensity
24-25 Inside vs. outside mean and stdev separation 3D intensity
26 Inside vs. outside contrast 3D intensity
27 XY gradient magnitude separation inside 2D gradient
28-29 Radial gradient mean and stdev inside 2D gradient randomly and equally split into two subsets with the same
30-31 Radial gradient mean and stdev outside 2D gradient proportion of true nodules. Suppose the labeling information in
32-33 Radial gradient mean and stdev separation 2D gradient subset 1 is known, a binary SVM classifier could be trained in
34-35 Mean and stdev of 3D gradient magnitude 3D gradient
36-37 Radial gradient mean and stdev inside 3D gradient subset 1 and applied for the classification of subset 2.
38-39 Radial gradient mean and stdev outside 3D gradient Subsequently, subset 2 is switched to be the training set and
40-41 Radial gradient mean and stdev separation 3D gradient subset 1 is treated as the testing set. This is indeed a two-fold
42-45 Min, max, mean and stdev of tubeness 3D Hessian
46-49 Min, max, mean and stdev of blobness 3D Hessian
cross validation. To alleviate the possible biases due to the
selection of training and testing datasets, we repeated the
random grouping process for sufficient times (e.g., 50). The
classifier performance can be summarized via receiver
operating characteristic (ROC) analysis of the decision scores
structuring element that has a radius of 5 pixels in the X-Y plane
from each SVM run. By searching the optimal operating point
and extends by a single slice along each direction of the Z-axis.
on the averaged ROC curve, we can obtain the classification
Several of the features in Table I warrant some explanation.
sensitivity and specificity, which are defined as follows
Specifically the 2D square compactness is computed on each
TP
relevant slice of the INC segmentation. The separation features Sensitivit y (7)
(such as standard deviation separation) are computed using the TP FN
difference of the inside and outside statistics divided by their TN .
sum. The inside and outside contrast is defined as the difference Specificit y (8)
TN FP
between the inside and outside means divided by the sum of the
inside and outside standard deviations. For the gradient features, Here TP, FN, TN and FP denote the number of true positives,
the XY gradient magnitude separation feature is defined as the false negatives, true negatives and false positives, respectively.
mean separation between the magnitudes of gradients along the
X- and Y-axis directions. Since pulmonary nodules usually have III. RESULTS
symmetric appearance, a radial gradient feature is also The proposed CADe system was validated on a subset of
incorporated. It defines the projection of gradient magnitude on the largest publicly available database – LIDC-IDRI [37]. A
the radial vector, and represents the gradient strength along the complete set of the annotated scans is available online at
radial direction. The 3D Hessian features are computed using http://cancerimagingarchive.net. The LIDC-IDRI images were
eigenvalues of the Hessian matrix, which has been shown in the collected under different scanner manufacturers, variable
literature to be potentially useful to distinguish blob-like and spatial resolution and different X-ray imaging parameters (e.g.
tubular objects [13], [34]. Since the shape of most nodules is slice thicknesses: 0.45-5.0 mm; in-plane pixel size: 0.461-0.977
close to a blob and the shape of blood vessels are tubular, both mm; total slice number: 96-605 slices; tube peak potential
tube-likeness and blob-likeness features [35] are defined to energies: 120-140 kV; and tube current: 40-627 mA).
feed our SVM classifier. Different from most CADe systems that did not report how
In this study, the LIBSVM [36] classifier with the well the system would perform for juxta-pleural nodules [38],
commonly used radial-basis-function (RBF) kernel is in this study we aim to particularly emphasize the performance
employed. The decision rule in our binary SVM classifier is of our CADe system subject to the large presence of
whether the INC tends to be a TP or FP, and the basic idea for juxta-pleural nodules. The LIDC-IDRI database has 226 scans
SVM is to construct an optimal classification boundary that with at least one juxta-pleural nodule annotation. Among those
maximizes the inter-cluster distance. Specifically, the scans, 21 were not included for evaluation due to the following
remaining INCs after the rule-based filtering operations are reasons. One scan shows a tube inside patient’s airway which
2168-2194 (c) 2013 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2014.2328870, IEEE Journal of Biomedical and Health Informatics
TITB-00731-2013-JBHI 7
(a) (b)
(c) (d)
Fig. 8. One original lung CT image and the corresponding PCA selected
component images in the K-L domain. (a) The original image; (b) the first PC;
(c) the second PC; (d) the third PC.
A. Evaluation of the PCA for Feature Extraction B. INCs Detection and Rule-Based Filtering Performance
The purpose of local intensity feature extraction is to retain For INCs detection, all of the 490 nodules from the 205
the major image information for reducing the computational chest CT scans were detected by the proposed hierarchical VQ
burden. Fig. 8 shows an example of one original lung CT image approach. Hence the sensitivity is 100%, and the number of FPs
and the corresponding 3-component images, in the K-L domain at this prescreening stage is about 1526 per scan. The
selected by PCA, with a decreasing order of significance. It is subsequent rule-based filtering operations preserved 475 of the
clear that the first 3 PCs represent the dominant information. nodules detected by VQ and brought the overall detection
Those PCs beyond the third one contain even less information. sensitivity to 96.9%. But the average number of FPs was
PCA not only preserves the true image information, but also significantly reduced from 1526 to 138.7 per scan, which
largely boosts the computational efficiency for our VQ corresponds to a FP reduction rate of more than 10 times. Three
algorithm. examples showing the performance of hierarchical INCs
detection followed by rule-based filtering operation are
illustrated through Fig. 9. The first example contains a huge
juxta-pleural nodule. The second example shows a part-solid
2168-2194 (c) 2013 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2014.2328870, IEEE Journal of Biomedical and Health Informatics
TITB-00731-2013-JBHI 8
TABLE II
AUC VALUES FOR SVM CLASSIFICATION OF INCS 1
0.9
Features AUC mean AUC stdev
0.8
Sensitivity
Gradient Features 0.9656 0.0046
Hessian Features 0.8641 0.0115 0.5
Using geometric features
Gradient + Geometric Features 0.9669 0.0055 0.4
Gradient + Intensity Features 0.9771 0.0041 Using intensity features
Gradient + Hessian Features 0.9651 0.0050 0.3 Using gradient features
Gradient + Intensity + Geometric 0.9743 0.0050 Using Hessian features
0.2 Using gradient+intensity features
Gradient + Intensity + Hessian 0.9753 0.0047
All Features 0.9776 0.0044 0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
1 - Specificity
Fig. 10. Averaged ROC curves showing the SVM classifier performance as a
juxta-pleural nodule, while the last example has a function of the step-wise selected features.
juxta-vascular nodule in the left lobe and a solitary nodule in lower than that for the previously selected “gradient + intensity”
the right lobe. In these examples the proposed scheme is feature set (one-sided p < 0.05). Finally, all the four groups of
effective to detect all the TPs while substantially reducing the features were combined together, but using all the extracted
FPs. features still could not yield better performance than using only
the “gradient + intensity” features (two-sided p = 0.5565).
C. Performance of FP Reduction by SVM Classification Thus, the best feature subset for the SVM classifier is the
After rule-based pruning of the obvious FPs, we further combination of gradient and intensity features. Fig. 10
reduced the FPs via feature-based SVM classification. illustrates the averaged ROC curves under different feature
Two-fold cross validation was used to estimate the combinations. The optimal operating point for the best feature
generalization capabilities of the SVM classifier. INCs were subset corresponds to a sensitivity of 92.7 % at a specificity
randomly and equally split into two folds, and each fold level of 93.3%.
consisted of 237 true nodules and 14217 FPs. Each fold was
used as the training set for once, and the other fold was treated D. Performance on Detection of Juxta-pleural Nodules
as the testing set. This two-fold cross validation was repeated
In particular, we further conducted an experiment to
for 50 times to generate the averaged SVM classification result.
examine the performance of our CADe system to the detection
Table II shows the values of area under the ROC curve
of juxta-pleural nodules, where the SVM classification was
(AUC) in different feature spaces based on the 50 times average.
exclusively applied on the juxta-pleural INCs. Technically, the
For group-wise feature comparison, two-sample t-test was
corrected lung mask for each scan is firstly eroded by 5 layers,
utilized to compare the mean of AUC value for different groups
and any INC with at least one voxel outside this eroded mask
of features. Statistical analysis showed that the gradient
will be treated as a juxta-pleural candidate. The INCs that do
features achieved a significantly higher AUC value than any of
not meet such criterion are categorized as internal candidates.
the other groups of features (one-sided p-value < 0.0001), and
The sensitivity and specificity of our CADe system before
the geometric features were suboptimum, better than either of
entering the SVM classifier could be summarized for the
the intensity or Hessian features (one-sided p < 0.0001). The
detection of juxta-pleural nodules exclusively. Among the 15
Hessian features alone yielded the least AUC value,
nodules that were mistakenly excluded by our filtering rules, 7
significantly lower than the intensity features (one-sided p <
are juxta-pleural nodules. This indicates a sensitivity of 97.8%
0.0001). According to the principle of forward feature selection,
for the detection of 323 juxta-pleural nodules, and the number
the next step is to compare the classifier performance by joining
of FPs within juxta-pleural INCs was in average 53.8 per scan.
any of the other three groups of features with the gradient
Again, two-fold cross validation was used to evaluate the
features selected in the previous step. Two-sample t-test
performance of the SVM classifier. Each fold consisted of 158
showed that the AUC value in the “gradient + intensity” feature
true nodules and 5512 FPs. This random cross-validation
space was significantly higher (one-sided p < 0.0001) than the
process was also repeated for 50 times to generate a series of
AUC value generated in either the “gradient + geometric” or
AUC and ROC results.
the “gradient + Hessian” feature space. It also outperformed the
The test AUC values of the SVM classification based on
previously selected gradient features (one-sided p < 0.0001).
different features are summarized in Table III. For the
Furthermore, we compared the performance of the SVM
group-wise feature comparison, two-sample t-test indicated
classifier with the addition of either geometric or Hessian
that the gradient features outperformed any of the other three
features to the “gradient + intensity” feature space. Statistical
groups of features (one-sided p < 0.0001). The intensity
analyses showed the AUC value with the addition of either
features were suboptimal, better than either of the geometric or
geometric or Hessian features was, however, significantly
2168-2194 (c) 2013 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2014.2328870, IEEE Journal of Biomedical and Health Informatics
TITB-00731-2013-JBHI 9
TABLE III
AUC VALUES FOR SVM CLASSIFICATION OF JUXTA-PLEURAL INCS 1
0.9
Features AUC mean AUC stdev
0.8
Geometric Features 0.8907 0.0099 0.7
Intensity Features 0.9031 0.0091
Sensitivity
0.6
Gradient Features 0.9659 0.0050
Hessian Features 0.8225 0.0153 0.5
Gradient + Geometric Features 0.9651 0.0064 Using geometric features
Gradient + Intensity Features 0.9703 0.0050 0.4 Using intensity features
Gradient + Hessian Features 0.9621 0.0060 0.3 Using gradient features
Gradient + Intensity + Geometric 0.9685 0.0065 Using Hessian features
Gradient + Intensity + Hessian 0.9677 0.0055 0.2 Using gradient+intensity features
All Features 0.9684 0.0058 0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
1 - Specificity
Fig. 11. Averaged ROC curves of the step-wise selected features for the SVM
Hessian features (one-sided p < 0.0001), while the Hessian classification of juxta-pleural nodule candidates.
features had the least contribution to the SVM classification.
based on the gradient field, Hessian matrix, and the size and
Upon forward selection of another group of features, the
shape of the segmented object. Reference [23] proposed a
“gradient + intensity” features generated a significantly higher
fixed-topology ANN classifier based on 45 geometric, position,
AUC value than either of the “gradient + geometric” or
and intensity features.
“gradient + Hessian” features (one-sided p < 0.0001). There
Although the detection parameters and target datasets
was also a significant increment of AUC value compared to that
(number of cases, number of nodules, and agreement level of
for the gradient features alone (one-sided p < 0.0001). In the
nodules) employed in those methods are different, it is still
next feature selection step, we added either geometric or
meaningful to attempt making relative comparisons. Summary
Hessian features to the “gradient + intensity” feature space.
of comparison with the existing systems is shown in Table IV,
Two-sample t-test showed there was no significant difference
where the ROC analysis results have been interpreted in terms
in AUC value by combining the geometric features (two-sided
of the free-response form. In this study, by considering the
p = 0.1396), and the AUC value after the inclusion of Hessian
detection rate of 96.9% after our VQ and rule-based filtering
features was even significantly lower (one-sided p = 0.0093).
operations, the proposed CADe system achieved an overall
Finally, the AUC value using all the extracted features was
detection sensitivity of 82.7% at 4 FPs/scan. Consequently, our
compared with that using the “gradient + intensity” features
CADe system shows not only comparable detection
only. Statistical test showed that the AUC by utilizing all
performance but also less computational cost by using the fast
available features was, however, significantly lower (one-sided
hierarchical VQ approach. On average, the detection of INCs
p = 0.0417) than the AUC using only a subset of features at a
consumed about 20-25 seconds per scan on a DELL Personal
significance level of 0.05. The averaged ROC plot of Fig. 11
Computer with a processor speed of 2.26GHz. This is about 10
indicates that the optimal operating point in the best “gradient +
times faster than existing methods.
intensity” feature space corresponds to a sensitivity of 91.2%
Moreover, we also compared the performance of our
and a specificity of 92.3%.
CADe system with four published systems that were devoted to
the identification of juxta-pleural nodules. Reference [41]
E. Comparison with Existing Methods proposed a multiscale α-hull approach to identify juxta-pleural
We compared the detection accuracy and speed of our nodules by patching lung border concavities, and then fed the
CADe system with those of existing systems that also evaluated located nodules to an ANN classifier for FP reduction. In
the LIDC-IDRI database. We selected five CADe systems that another study [17], a conventional template matching approach
showed detection capability or reported detection speed. The was employed to detect nodules in the lung wall area, and only
detection of INCs in [22] was based on a multithreshold four features were used to eliminate FP findings. Reference [21]
surface-triangulation approach, and the features used as input to presented an approach to tackle the detection of internal and
their ANN classifier were volume, roundness, maximum sub-pleural nodules, respectively, and both algorithms were
density, mass, and principal moments of inertia. Reference [39] finally combined as one assembled CADe system for
used a 2D filter to extract the seeds of INCs on each slice image, evaluation. In their method [21], the INCs detection was based
and then applied a SVM classifier with 6 complex features to on a spherical-enhanced filter, and the FP reduction was carried
reduce the FPs. Based on maximum intensity projection out via a voxel-based neural network approach. Later on,
processing and Zernike moments, a feature extraction algorithm another CADe system [42] was particularly developed for
in [20] was employed to enhance the SVM for FP reduction. pleural nodule identification using directional-gradient
Reference [40] extracted the INCs by two segmentation concentration method in combination with a morphological
algorithms of region growing and 3D active contour, while the opening-based procedure. Each identified candidate was
following LDA classification employed three groups of features characterized by 12 morphological and textural features, which
2168-2194 (c) 2013 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2014.2328870, IEEE Journal of Biomedical and Health Informatics
TITB-00731-2013-JBHI 10
TABLE IV
PERFORMANCE COMPARISON OF PROPOSED CADE SYSTEM WITH EXISTING SYSTEMS USING THE LIDC-IDRI DATABASE
Number of Number of Nodule size Agreement Average Detection time
CADe Systems Sensitivity (%)
cases nodules used level FPs/scan (seconds)
Golosio et al. [22] 84 148 >= 3 mm 1 4.0 45.0 -
Opfer and Wiemker [39] - 127 >= 4 mm 1 4.0 76.0 180 - 300
Riccardi et al. [20] 154 387 - 1 4.0 49.0 -
Sahiner et al. [40] 48 73 3 - 36.4 mm 1 4.9 79.0 -
Tan et al. [23] 125 259 >= 3 mm 1 3.0 66.4 -
The proposed system 205 490 >= 3 mm 1 4.0 82.7 20 – 25
TABLE V
COMPARISON OF PROPOSED CADE SYSTEM WITH EXISTING SYSTEMS ON JUXTA-PLEURAL NODULE IDENTIFICATION
Number of Number of Agreement Average
CADe Systems Sensitivity (%)
cases nodules level FPs/scan
De Nunzio et al. [41] 57 78 - - 66.5
Lee et al. [17] 20 24 - 14.15 71.0
Retico et al. [21] 39 27 2 10 74.0
Retico et al. [42] 42 25 2 6 72.0
The proposed system 205 323 1 4.14 89.2
were analyzed by a rule-based filter and a neural classifier. low agreement level also places a challenge to the detection
Table V shows the summary of system performance power of our CADe system.
comparison. The optimal operating of our CADe system It is worth mentioning that a proper correction of the lung
achieved an overall sensitivity of 89.2% at 4.14 FPs per scan. boundary is critical, because failure to include some
Compared with the existing methods in Table V, our CADe juxta-pleural nodules makes it impossible for VQ to recover
system was evaluated via a much larger dataset and, therefore, them in the later stages. Different from some previous works
it outperforms in terms of not only the detection accuracy using 2D closing operation, we adopted the 3D morphological
(including both sensitivity and specificity), but also the closing strategy to close the lung boundary. Fig. 12 portrays two
statistical power. challenging juxta-pleural nodules that were excluded by the 2D
closing operation but recovered by the 3D operation, which
IV. DISCUSSIONS turns to be more robust for lung border smoothing.
Since LIDC-IDRI images were collected from several Though the proposed hierarchical VQ scheme was able to
different institutions, the evaluation of such database was much detect all non-GGO nodules in the studied datasets, more efforts
more challenging than evaluation on images acquired from a are desired to deal with the associated large number of FPs. One
single institution. Another challenge to the capability of our possible solution is to merge clusters close to each other so that
CADe system is to deal with juxta-pleural datasets. As we separated portions of a single object may be associated. Another
know, juxta-pleural nodules are usually semi-spherical or solution is to incorporate more features into the feature vector
spiculate in shape, making them relatively difficult to detect. In for VQ, such as geometric features (e.g., shape index,
addition, because they have similar intensities to the pleural curvedness) and texture features (e.g., local binary pattern,
wall, segmentation algorithms often fail in correcting the lung Haralick and Gabor features). By considering sophisticated
borders to include such nodules. Therefore, it is meaningful to features, FPs with low contrast to true nodules are expected to
stress-test any CADe system in the presence of be identified by some unique feature characteristics other than
irregularly-shaped juxta-pleural nodules. This study collected just the image intensity. In addition, we also investigated the
205 chest CT scans with at least one juxta-pleural attachment to feasibility of boosting low level VQ for the detection of GGO
evaluate the proposed CADe system. For this regard, this study nodules. Some of the GGO nodules in LIDC database may be
is unique. recovered by increasing the maximum class number for VQ and
The nodule size distribution in Fig. 7 shows the selecting two or more high-intensity classes as the INCs.
representativeness of the studied datasets, where about 74% of It has been widely acknowledged that the inclusion of
the nodules are less than 10 mm. This is clinically valuable in rule-based filtering before the computationally expensive
the sense that the detection of small nodules is crucial to classification step will enhance the performance of CADe
detecting lung cancer at early stages. Besides the substantial systems. Since intensity-based VQ could barely separate
presence of small nodules, the consideration of nodules with nodules from other attaching structures with similar intensities,
preprocessing of the detected INCs is necessary to segment the
juxta-vascular nodules, which are otherwise prone to be
2168-2194 (c) 2013 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2014.2328870, IEEE Journal of Biomedical and Health Informatics
TITB-00731-2013-JBHI 11
2168-2194 (c) 2013 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2014.2328870, IEEE Journal of Biomedical and Health Informatics
TITB-00731-2013-JBHI 12
[10] E. M. van Rikxoort, B. de Hoop, M. A. Viergever, M. Prokop, and B. van [32] K. Fukunaga, Introduction to Statistical Pattern Recognition, Academic
Ginneken, “Automatic lung segmentation from thoracic computed Press, second edition, 1990.
tomography scans using a hybrid approach with error detection,” Med. [33] R. C. Gonzalez and R. E. Woods, Digital Image Processing, Pearson
Phys., vol. 36, no. 7, pp. 2934-2947, 2009. Prenctice Hall, 2007.
[11] T. Messay, R. C. Hardie, and S. K. Rogers, “A new computationally [34] Y. Sato, C. F. Westin, A. Bhalerao, S. Nakajima, N. Shiraga, S. Tamura,
efficient CAD system for pulmonary nodule detection in CT imagery,” and R. Kijinis, “Tissue classification based on 3D local intensity
Med. Image Anal., vol. 14, no. 3, pp. 390-406, 2010. structures for volume rendering,” IEEE Trans. Vis. Comput. Graph., vol.
[12] W. Choi and T. Choi, “Genetic programming-based feature transform and 6, no. 2, pp. 160-180, 2000.
classification for the automatic detection of pulmonary nodules on [35] B. Sahiner, Z. Ge, H.-P. Chan, L. M. Hadjiiski, N. Bogot, P. N. Cascade,
computed tomography images,” Inf. Sci., vol. 212, pp. 57-78, 2012. and E. A. Kazerooni, “False-positive reduction using Hessian features in
[13] Q. Li, S. Sone, and K. Doi, “Selective enhancement filters for nodules, computer-aided detection of pulmonary nodules on thoracic CT images,”
vessels, and airway walls in two- and three-dimensional CT scans,” Med. in Proc. of SPIE, 2005, vol. 5747, pp. 790-795.
Phys., vol. 30, no. 8, pp. 2040-2051, 2003. [36] C-C Chang and C-J Lin, “LIBSVM: A library for support vector
[14] A. Teramoto and H. Fujita, “Fast lung nodule detection in chest CT machines,” ACM Trans. Intell. Syst. Technol., vol. 2, no. 3, 2011.
images using cylindrical nodule-enhancement filter,” Int. J. CARS, vol. 8, [37] S. G. Armato et al., “The lung image database consortium (LIDC) and
no. 2, pp. 193-205, 2013. image database resource initiative (IDRI): A complete reference database
[15] C. I. Fetita, F. Preteux, C. Beigelman-Aubry, and P. Grenier, “3-D of lung nodules on CT scans,” Med. Phys., vol. 38, pp. 915-931, 2011.
automated lung nodule segmentation in HRCT,” in Lecture Notes in [38] Q. Li, “Recent progress in computer-aided diagnosis of lung nodules on
Computer Science. Berlin, Germany: Springer-Verlag, 2003, vol. 2878, thin-section CT,” Comput. Med. Imag. Graph., vol. 31, no. 4-5, pp.
Medical Image Computing and Computer-Assisted Intervention, pp. 248-257, 2007.
626-634. [39] R. Opfer and R. Wiemker, “Performance analysis for computer-aided
[16] K. Awai, K. Murao, A. Ozawa, M. Komi, H. Hayakawa, S. Hori, and Y. lung nodule detection on LIDC data,” in Proc. of SPIE, 2007, vol. 6515,
Nishimura, “Pulmonary nodules at chest CT: Effect of computer-aided pp. 1-9.
diagnosis on radiologists’ detection performance,” Radiol., vol. 230, no. [40] B. Sahiner, L. M. Hadjiiski, H. P. Chan, J. Shi, T. Way, P. N. Cascade, E.
2, pp. 347-d352, 2004. A. Kazerooni, C. Zhou, and J. Wei, “The effect of nodule segmentation on
[17] Y. Lee, T. Hara, H. Fujita, S. Itoh, and T. Ishigaki, “Automated detection the accuracy of computerized lung nodule detection on CT scans:
of pulmonary nodules in helical CT images based on an improved Comparison on a data set annotated by multiple radiologists,” in Proc. of
template-matching technique,” IEEE Trans. Med. Imag., vol. 20, no. 7, SPIE, 2007, vol. 6514, pp. 65140L-1-7.
pp. 595-604, 2001. [41] G. De Nunzio, A. Massafra, R. Cataldo, I. De Mitri, M. Peccarisi, M. E.
[18] J. Dehmeshki, X. Ye, X. Lin, M. Valdivieso, and H. Amin, “Automated Fantacci, G. Gargano, and E. Lopez Torres, “Approaches to juxta-pleural
detection of lung nodules in CT images using shape-based genetic nodule detection in CT images within the MAGIC-5 collaboration,” Nucl.
algorithm,” Comput. Med. Imag. Graph., vol. 31, pp. 408-417, 2007. Instrum. Methods Phys. Res. A, vol. 648, pp. S103-S106, 2011.
[19] X. Ye, X. Lin, J. Dehmeshki, G. Slabaugh, and G. Beddoe, “Shape-based [42] A. Retico, M. E. Fantacci, I. Gori, P. Kasae, B. Golosio, A. Piccioli, P.
computer-aided detection of lung nodules in thoracic CT images,” IEEE Cerello, G. De Nunzio, and S. Tangaro, “Pleural nodule identification in
Trans. Biomed. Engineer., vol. 56, no. 7, pp. 1810-1820, 2009. low-dose and thin-slice lung computed tomography,” Comput. Biol.
[20] A. Riccardi, T. S. Petkov, G. Ferri, M. Masotti, and R. Campanini, Med., vol. 39, no. 12, pp. 1137-1144, 2009.
“Computer-aided detection of lung nodules via 3D fast radial transform, [43] F. Han, H. Wang, B. Song, G. Zhang, H. Lu, W. Moore, H. Zhao, and Z.
scale space representation, and Zernike MIP classification,” Med. Phys., Liang, “A new 3D texture feature based computer-aided diagnosis
vol. 38, no. 4, pp. 1962-1971, 2011. approach to differentiate pulmonary nodules,” in Proc. of SPIE, 2013,
[21] A. Retico, P. Delogu, M. E. Fantacci, I. Gori, and A. Preite Martinez, vol. 8670, pp. 86702Z-1-7.
“Lung nodule detection in low-dose and thin-slice computed
tomography,” Comput. Biol. Med., vol. 38, no. 4, pp. 525-534, 2008.
[22] B. Golosio, G. L. Masala, A. Piccioli, P. Oliva, M. Carpinelli, R. Cataldo,
P. Cerello, F. DeCarlo, F. Falaschi, M. E. Fantacci, G. Gargano, P. Kasae,
and M. Torsello, “A novel multithreshold method for nodule detection in
lung CT,” Med. Phys., vol. 36, no. 8, pp. 3607-3618, 2009.
[23] M. Tan, R. Deklerck, B. Jansen, M. Bister, and J. Cornelis, “A novel
computer-aided lung nodule detection system for CT images,” Med.
Phys., vol. 38, no. 10, pp. 5630-5645, 2011.
[24] D. Chen, Z. Liang, M. R. Wax, L. Li, B. Li, and A. E. Kaufman, “A novel
approach to extract colon lumen from CT images for virtual
colonoscopy,” IEEE Trans. Med. Imag., vol. 19, no. 12, pp. 1220-1226,
2000.
[25] H. Han, L. Li, F. Han, H. Zhang, W. Moore, and Z. Liang, “Vector
quantization-based automatic detection of pulmonary nodules in thoracic
CT images,” in Conf. Record IEEE NSS/MIC, 2013, M22-41.
[26] A. Gersho and R. M. Gray, Vector Quantization and Signal Compression,
Boston: Kluwer Academic, 1992.
[27] C. C. Chang and W. C. Wu, “Fast planar-oriented ripple search algorithm
for hyperspace VQ codebook,” IEEE Trans. Image Process., vol. 16, no.
6, pp. 1538-1547, 2007.
[28] C. Garcia and G. Tziritas, “Face detection using quantized skin color
regions merging and wavelet packet analysis,” IEEE Trans. Multimedia,
vol. 1, no. 3, pp. 264-277, 1999.
[29] K. L. Oehler and R. M. Gray, “Combing image compression and
classification using vector quantization,” IEEE Trans. Pattern Anal.
Mach. Intell., vol. 17, no. 5, pp. 461-473, 1995.
[30] L. Li, D. Chen, H. Lu, and Z. Liang, “Segmentation of brain MR images:
a self-adaptive online vector quantization approach,” in Proc. of SPIE,
2001, vol. 4322, pp. 1431-1438.
[31] L. Li, D. Chen, H. Lu, and Z. Liang, “Comparison of quadratic and linear
discriminate analyses in the self-adaptive feature vector quantization
scheme for MR image segmentation,” in Proc. Int. Soc. Mag. Reson.
Med., 2001, vol. 9, pp. 809.
2168-2194 (c) 2013 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.