close

Вход

Забыли?

вход по аккаунту

?

j.compbiomed.2017.10.019

код для вставкиСкачать
Accepted Manuscript
Wall-based measurement features provides an improved IVUS coronary artery risk
assessment when fused with plaque texture-based features during machine learning
paradigm
Sumit K. Banchhor, Narendra D. Londhe, Tadashi Araki, Luca Saba, Petia Radeva,
John R. Laird, Jasjit S. Suri
PII:
S0010-4825(17)30345-1
DOI:
10.1016/j.compbiomed.2017.10.019
Reference:
CBM 2812
To appear in:
Computers in Biology and Medicine
Received Date: 4 September 2017
Revised Date:
19 October 2017
Accepted Date: 19 October 2017
Please cite this article as: S.K. Banchhor, N.D. Londhe, T. Araki, L. Saba, P. Radeva, J.R. Laird, J.S.
Suri, Wall-based measurement features provides an improved IVUS coronary artery risk assessment
when fused with plaque texture-based features during machine learning paradigm, Computers in Biology
and Medicine (2017), doi: 10.1016/j.compbiomed.2017.10.019.
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Wall-based measurement features provides an improved IVUS coronary artery risk
assessment when fused with plaque texture-based features during machine learning
paradigm
Sumit K. Banchhor1, MTech, Narendra D. Londhe1, PhD, Tadashi Araki2, MD, Luca Saba3, MD,
Petia Radeva4, PhD, John R. Laird5, MD, Jasjit S. Suri6,7, PhD, MBA, Fellow AIMBE
1
Department of Electrical Engineering, NIT Raipur, CG, India
Division of Cardiovascular Medicine, Toho University Ohashi Medical Center, Tokyo, Japan
3
Department of Radiology, University of Cagliari, Italy
4
Department of Mathematics and Computer Science, University of Barcelona, Barcelona, Spain
5
UC Davis Vascular Centre, University of California, Davis, CA, USA
6
Monitoring and Diagnostic Division, AtheroPoint™, Roseville, CA, USA
Address correspondence to/Manuscript Reprints:
AC
C
EP
TE
D
M
AN
U
Manuscript category: Original research article
SC
Dr. Jasjit S. Suri, PhD, MBA, Fellow AIMBE (Corresponding Author)
Monitoring and Diagnostic Division
AtheroPoint™, Roseville, CA, USA
Zip Code: 95661
Phone: (916)-749-5628
E-mail: [email protected]
RI
PT
2
1
ACCEPTED MANUSCRIPT
Abstract
Background: Planning of percutaneous interventional procedures involves a pre-screening and risk
stratification of the coronary artery disease. Current screening tools use stand-alone plaque texturebased features and therefore lack the ability to stratify the risk.
Method: This IRB approved study presents a novel strategy for coronary artery disease risk
stratification using an amalgamation of IVUS plaque texture-based and wall-based measurement
features. Due to common genetic plaque makeup, carotid plaque burden was chosen as a gold
standard for risk labels during training-phase of machine learning (ML) paradigm. Cross-validation
RI
PT
protocol was adopted to compute the accuracy of the ML framework. A set of 59 plaque texturebased features was padded with six wall-based measurement features to show the improvement in
stratification accuracy. The ML system was executed using principle component analysis-based
framework for dimensionality reduction and uses support vector machine classifier for training and
testing-phases.
Results: The ML system produced a stratification accuracy of 91.28%, demonstrating an
SC
improvement of 5.69% when wall-based measurement features were combined with plaque texture-
based features. The fused system showed an improvement in mean sensitivity, specificity, positive
predictive value, and area under the curve by: 6.39%, 4.59%, 3.31% and 5.48%, respectively when
M
AN
U
compared to the stand-alone system. While meeting the stability criteria of 5%, the ML system also
showed a high average feature retaining power and mean reliability of 89.32% and 98.24%,
respectively.
Conclusions: The ML system showed an improvement in risk stratification accuracy when the
wall-based measurement features were fused with the plaque texture-based features.
Highlights:
Amalgamation of IVUS plaque texture-based and wall-based measurement features.
•
Principle component analysis-based framework was used for dimensionality reduction.
•
During the training phase, carotid plaque burden was chosen as a gold standard.
•
Support vector machine was used as a classifier for training and testing-phases.
•
Proposed ML system demonstrate improvement in risk stratification accuracy.
TE
D
•
Keywords: Atherosclerosis, Cardiovascular disease, Carotid artery, Coronary arteries, Ultrasound
AC
C
EP
imaging
2
ACCEPTED MANUSCRIPT
1. Introduction
The atherosclerotic cardiovascular disease accounts for the largest number of deaths in the
USA [1]. The disease of atherosclerosis over time causes calcium to build-up in the coronary
arteries [2]. During the advanced stage of the disease, the combined risk of the patient includes
higher plaque growth leading to plaque rupture. This also includes the risk of development of
different components such as: fibro-fatty, macrophages, calcium, and fatty tissue. When these
components increase in size, there is a risk of an increase in stenosis and stress on the fibrous cap
RI
PT
thickness which can cause the risk of rupture leading to myocardial infarction (MI). All of the
above can be categorized as the “risk of arterial wall rupture” or “risk for MI”. Thus, rupture of the
arterial wall cap can cause calcium to dislodge, blocking the oxygen-rich blood flow in the arteries,
leading to myocardial infarction or stroke [3]. Current screening methods like CT or MRI [4,5]
SC
suffers either from excess radiation or magnetic interference. Further, these devices take longer
time to reconstruct the images thereby lacking real-time interface [4]. Intravascular ultrasound
(IVUS) screening, on the other hand, has low-radiation exposure, is economic compared to
M
AN
U
MR/CT, ergonomic, and offers real-time diagnosis [6,7].
Prior to stenting and percutaneous interventional procedures, cardiologists are interested in
performing pre-screening and risk stratification of coronary artery disease (CAD). Studies for risk
stratification of cardiovascular events are mainly categorized into two groups. The first group
attempts to predict the risk by quantifying the plaque characteristics (i.e., texture-based features)
while the second group predicts the risk by quantifying wall-based measurement features [8-13].
TE
D
Christodoulou et al. [8] in 2003, proposed a neural network for carotid plaque classification. Ten
dominant texture feature sets were selected from a total of 61 texture features showing a low
accuracy of 73.10% on a data size of 230 images. Two years later, Kyriacou et al. [9] showed a
carotid classification system that used neural network classifier with 10 different textures and
AC
C
EP
carotid wall-based features and achieved even a slightly lower accuracy 71.2% on a data size of
3
ACCEPTED MANUSCRIPT
274 images. The same group in 2009 applied support vector machine (SVM) classifier on the same
database using only the texture features and showed an improvement in accuracy by 2.5%. In a
hybrid neural network approach on B-mode carotid ultrasound images, Mongiakakou et al. [10] in
2007 used 21 statistical and law’s features on 108 images. The neural network was trained on the
combined use of genetic algorithms and back propagation with momentum and adaptive learning
rate and showed an accuracy of 99.10%.
Our team led by Suri have been working on the characterization of carotid plaque. Acharya et
RI
PT
al. [12] in 2012, proposed an Atheromatic™ system for plaque stratification into symptomatic and
asymptomatic plaques showing an accuracy of 82.40% and 81.70%, using SVM and AdaBoost
classifiers, respectively. The same group [13] in 2012, obtain an accuracy of 83% by fusing the
plaque texture-based and wall-based measurement features using an SVM classifier. A year later,
SC
the same group [14], obtain a further high accuracy of 85.3% by fusing discrete wavelet transform,
higher order spectra and textural features on a large data size consisting of 492 images. A year
later, Pedro et al. [15] fused the clinical and texture features for the classification of carotid plaque.
M
AN
U
An enhanced activity index was proposed and was correlated with the presence or absence of
ipsilateral appropriate ischemic symptoms. Leave-one-patient-out was applied to 146 carotid
plaques obtained from 99 patients and a cross-validation accuracy of 77% was obtained. Araki et
al. [16] showed a CADx system by using the SVM that demonstrated a training and testing-based
ML system using plaque texture features for coronary artery risk assessment. Later, the same group
[17] modified and improved their CADx system by introducing the principal component analysis
TE
D
(PCA)-based polling technique for selection of the grayscale dominant features for improving the
stratification accuracy. These prior studies had ignored how plaque growth affects the walls of the
arteries and lacked the prominent features contributed by the wall-based parameters. This study is
an extension to above studies by using an amalgamation of IVUS plaque texture-based with wall-
AC
C
EP
based measurement features. This is motivated by the current strategy by Suri and his team in
4
ACCEPTED MANUSCRIPT
stroke imaging where carotid IMT wall thickness variability was fused with carotid longitudinal
grayscale features to improve the stroke risk stratification [18,19]. But in our current study, circular
wall parameters along with the plaque calcium are derived as measurement features from IVUS
coronary walls. Thus, the objective is to demonstrate the importance of wall-based measurement
features and its integration with plaque texture-based grayscale features for better ML system
design.
Calcium accumulations always occur in the atheroma region which lies between lumen (inner
RI
PT
wall or internal elastic wall) region and vessel (outer wall or external elastic wall) region [20].
Therefore, an expansion of the walls is purely a reflectance of the growth of calcium in the arteries.
Moreover, due to multi-focal nature of calcium [21], the wall thickness can vary along the circular
walls of the coronary artery. Fig. 1 and Fig. 2 shows typical examples of images showing grayscale
SC
ring, along with calcium, lumen, vessel, and atheroma regional areas corresponding to five different
low-risk and high-risk patients, respectively. Furthermore, our study is based on two hypotheses: (i)
fusion of plaque texture-based and wall-based measurement features can offer an improvement in
M
AN
U
the coronary artery risk stratification; (ii) due to the genetic make-up of the plaque, carotid plaque
burden which is considered as a biomarker for stroke risk [22-28] can be used as a risk label for
patients with coronary artery disease [16,17].
The novelty of this study is to demonstrate an improvement in the accuracy of the CADx
system built for the coronary artery risk assessment by fusing plaque texture-based features with
wall-based measurement features compared to a stand-alone system consisting of only plaque
TE
D
texture-based features. Our objective in this paper is to predict the class label of the plaque type as
AC
C
EP
high-risk or low-risk.
5
M
AN
U
SC
RI
PT
ACCEPTED MANUSCRIPT
AC
C
EP
TE
D
Fig. 1. Typical examples of images showing grayscale ring, calcium regional area, vessel regional
area, lumen regional area, and atheroma regional area from five different low-risk patients.
6
M
AN
U
SC
RI
PT
ACCEPTED MANUSCRIPT
AC
C
EP
TE
D
Fig. 2. Typical examples of images showing grayscale ring, calcium regional area, vessel regional
area, lumen regional area, and atheroma regional area from five different high-risk patients.
7
2. Material and methods
ACCEPTED MANUSCRIPT
From a single case study, twenty-two patients with stable angina pectoris who underwent
percutaneous coronary interventions between July 2009 and December 2010 using iMAP (Boston
Scientific®, Marlborough, MA) IVUS examination were considered for this study. The study
consisted of 22 patients (20 M/2 F) in the age group of 36 to 81 years (average 66±12 years). In this
database, ten patients had a calcified location on the left anterior descending, eight on right, two on
left circumflex and two on the left main coronary artery. Out of 22 patients, ten had proximal, six at
RI
PT
the middle, and six at a distal location. Five patients had a family history of coronary artery disease.
Furthermore, one patient had a prior myocardial infarction while the other had undergone a prior
coronary artery bypass grafting. The mean hemoglobin, LDL, HDL, and total cholesterol were: 5.6
g/dL, 94.4 mg/dL, 48.7 mg/dL, and 168 mg/dL, respectively. Ten patients from the pool of 22
SC
patients were smokers.
The dataset was approved by the Institutional Review Board and written informed consent was
provided by all the patients. In this study, all the patients have undergone both carotid and coronary
M
AN
U
ultrasound examinations. Carotid examinations were performed with a scanner (Aplio XV,
AplioXG, Xario, Toshiba, Inc., Tokyo, Japan) equipped with a 7.5 MHz linear array transducer.
For coronary data acquisition, a 40-MHz IVUS catheter (Atlantis SR Pro; Boston Scientific) was
used. For the carotid database, high-resolution images were acquired as recommended by the
American Society of Echocardiography while, for the coronary database, DICOM image format
was used. While converting into the AVI movies, these DICOM images (16-bits per pixel) were
TE
D
further compressed to JPEG (8 bits per pixel) images. The mean pixel resolution for the carotid and
AC
C
EP
coronary databases were 0.05±0.01 mm/pixel and 0.0167 mm/pixel, respectively.
8
ACCEPTED MANUSCRIPT
2.1 Step 1: IVUS data preparation
The coronary dataset consists of 22 patients with 2109 frames per patient. Our coronary data
size preparation involves two assumptions: (a) the head and tail end frames did not contain any
morphological information about calcium and (b) the artery is not a straight line and longitudinal
and transversal displacements do exist between frames [29,30]. In this database, the change in
calcium was observed in every 10th frame. Using these two assumptions, we accumulated 4930
frames derived from 22 patients. Considering carotid risk threshold [16,17,23], all the patients with
RI
PT
a carotid plaque burden higher than or equal to 0.9 mm are categorized into high-risk and the
patients with a plaque burden lower than 0.9 mm are categorized in the low-risk pool. Accordingly,
14 patients were categorized into high-risk (~63.63%) and rest 8 patients were in the low-risk pool
(~36.36%). Note that our data size in ML framework was not 22 subjects, but 4930 frames
SC
collected from 22 subjects, with pools consisting of 3043 high-risk frames and 1887 low-risk
frames corresponding to 14 high-risk patients and 8 low-risk patients, respectively.
M
AN
U
2.2 Step 2: Wall region of interest estimation
In this study for the ROI estimation, we have employed an ImgTracer™ system (courtesy of
AtheroPoint™, Roseville, CA, USA) recently used by Suri and his team [20,29-31]. Here, two
experts (doctoral students with knowledge about coronary artery disease and IVUS imaging)
generated the vessel wall region by manually tracing the internal elastic lamina (IEL) and external
elastic lamina (EEL) borders. A typical example of manually traced IEL/EEL borders is shown in
TE
D
Fig. 3. In Fig. 3(a), the inner/outer yellow rings indicate IEL/EEL borders. The atherosclerotic wall
AC
C
EP
region treated as the ROI is shown in Fig. 3(b).
9
RI
PT
ACCEPTED MANUSCRIPT
Fig. 3. (color image): (a) Manually traced internal elastic lamina and external elastic lamina
indicated by Inner and outer yellow rings in the vessel wall region obtained by manually tracings
using ImgTracer™. (b) Atherosclerotic grayscale ring image used as the ROI (Courtesy of
AtheroPoint™, Roseville, CA, USA) (Courtesy of source [Banchhor et al. [29]]).
2.3 Step 3: Wall-based feature computation
SC
In this study, we have computed six different wall-based measurement features namely:
coronary calcium area, coronary vessel area, coronary lumen area, coronary atheroma area,
M
AN
U
coronary wall thickness, and coronary wall thickness variability [20]. Coronary calcium area is
computed using a well-established threshold-based segmentation technique [29,31]. However,
coronary vessel area, coronary lumen area, coronary atheroma area was computed by generating
their corresponding binary mask images using the IEL/EEL borders manually traced by experts
using the ImgTracer™ system [20]. Further, using the bidirectional concept of polyline distance
method [32,33], we have computed the coronary wall thickness. Lastly, coronary wall thickness
TE
D
variability is estimated by computing the standard deviation of the coronary wall thickness over all
the frames. For this study, we have extracted 59 different plaque texture-based features [16,17],
AC
C
EP
thus having a total set of 65 features.
10
ACCEPTED MANUSCRIPT
2.4 Step 4: Machine learning (ML)-paradigm for class prediction
Fig. 4 shows the block diagram of the machine learning system used in this study using the
fusion of plaque texture-based and wall-based measurement features (as shown in arrows on left
and right). However, since there are a large number of features (= 65), it is first required to extract
the dominant features from the pool consisting of plaque texture-based features fused with wallbased measurement features. For this reason, we have adopted a PCA-based polling strategy [16].
Furthermore, for training the machine learning classifier to perform tissue classification, we have
RI
PT
used a classifier based on SVM [16,17]. The patient population is divided into two components:
training population and testing population. Training population is used for computing the learning
parameters during the machine learning process. These parameters are computed using training
grayscale coronary wall region and corresponding carotid gold standard risk labels. This gold
SC
standard is derived from the concept of the second hypothesis, which states that the atherosclerotic
plaque has a common genetic makeup, as discussed before in Step 1 [16,17]. In ML paradigm, one
computes the offline grayscale plaque features and trains these features according to the gold
M
AN
U
standard risk labels derived from the carotid artery plaque burden. These offline parameters are
then transformed by the online grayscale wall features computed from the test images to predict its
risk label. Finally, the process of ML is repeated for the cyclic combinations as per the K-fold
combinations. Since there are 10 parts (K10 protocol: 90% training and 10% testing), we, therefore,
rotate this combination 10 times to ensure that each set of 10% testing data gets a chance to become
a training data sets. Each combination yields the stratification accuracy using the ML-system. The
TE
D
mean value of the classification accuracy is then computed which determines the final accuracy of
the ML-system. Furthermore, to separate the features in SVM framework, we have adopted both a
linear and four non-linear kernel functions [17] namely: radial basis function (RBF), polynomial of
AC
C
phases.
EP
order 1, 2, and 3. These kernel functions are used by the SVM during its training and testing-
11
SC
RI
PT
ACCEPTED MANUSCRIPT
M
AN
U
Fig. 4. Improved coronary risk assessment system (cRAS) using machine learning paradigm
utilizing PCA with the fusion of plaque texture-based and wall-based measurement features. Grey
boxes show novel wall-based measurement features.
3. Results
The main observations here are to see the effect of fusion of wall-based measurement features
with plaque texture-based features on the stratification accuracy in ML framework. These results
TE
D
will characterize: (i) the PCA polling process, best kernel design during the classification process
and (iii) cutoff values between memorization vs. generalization for a dataset size.
3.1 Experimental result 1: Dominant features selection
Best dominant feature combination set can increase the accuracy of the SVM classifier. The
AC
C
EP
objective of this experiment is to find the best matching set of features to yield the highest accuracy
12
ACCEPTED MANUSCRIPT
using PCA-based polling strategy. This is achieved by taking different cutoffs ranging from 0.90 to
0.99 in the increment of 0.01 for a fixed data size.
Table 1, shows the dominant features selected for different PCA-based cutoffs for: (a) standalone plaque texture-based features and (b) plaque texture-based features fused with wall-based
measurement features. In both the cases, we can observe that the number of dominant features
gradually increases with the increase in PCA-based cutoffs. In Table 1, for the fusion of plaque
texture-based and wall-based measurement features, coronary calcium area (feature # 61), coronary
RI
PT
lumen area (feature # 59), and coronary wall thickness variability (feature # 65) are selected as the
dominant features for different PCA-based cutoffs. The selection of dominant features taking
different PCA-based cutoffs for both: (a) stand-alone plaque texture-based features and (b) plaque
3.2 Experimental result 2: Selection of best kernel function
SC
texture-based features fused with wall-based measurement features are shown in Fig. 5.
Since the hyperplane for stratification is governed by the choice of kernel function used [34],
M
AN
U
we, therefore, choose five types kernels [16,17] such as: linear, RBF, Polynomial of order 1, 2, and
3, respectively for optimization in both the paradigms: with and without wall-based measurement
features.
For this protocol, we have fixed the data size. Using five different kernel functions, the
accuracy of the SVM classifier for both: (a) stand-alone plaque texture-based features and (b)
plaque texture-based features fused with wall-based measurement features are shown in Table 2.
TE
D
Fig. 6 (a) and Fig. 6 (b) shows the graphical representation of the accuracy of the SVM classifier
for both: (a) stand-alone plaque texture-based features and (b) plaque texture-based features fused
with wall-based measurement features. We observe that for all the kernel functions, the mean
accuracy of all the PCA-based cutoffs, using the fusion of plaque texture-based and wall-based
AC
C
EP
measurement features are higher as compared to the stand-alone option where plaque texture-based
13
ACCEPTED MANUSCRIPT
features are only considered. Among all the kernel functions, RBF gave the highest accuracy for all
the PCA-based cutoffs, hence is considered as the best among all the four kernel functions. This is
consistent with our other studies [16,17].
Table 1 Dominant features selected at each PCA-based cutoff using (a) stand-alone plaque texturebased features and (b) plaque texture-based features fused with wall-based measurement features.
(a) Using stand-alone plaque texture-based features
Cutoffs
F1 F2 F3 F4 F5 F6 F7 F8 F9
F10
F11
F12
F13
F14
F15
52 56 16 15 6
52 56 16 15 6
52 56 16 15 6
52 56 55 26 24 4
52 56 55 26 24 4
52 56 55 26 24 4 31
52 56 55 26 24 4 31
52 44 56 55 26 42 43 4
52 55 44 4 26 56 43 35 42 37
55 52 26 56 44 4 43 28 37 42
35
31
24
(b) Using plaque texture-based features fused with wall-based measurement features
52 56 65 28 16
0.90
52 59 61 56 26 55
0.91
52 59 61 56 26 55
0.92
52 65 59 26 61 55 44
0.93
52 65 59 26 61 55 44
0.94
52 65 55 61 44 56 59 26
0.95
65 52 61 55 44 26 56 59 24
0.96
65 61 52 55 44 26 59 4 56 21
0.97
61 44 65 59 55 26 52 4 56 24
21
43
0.98
44 61 65 55 26 59 52 43 4
37
24
21
54
35
28
0.99
M
AN
U
SC
RI
PT
0.90
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
F16
42
The numerical numbers listed in the columns have a unique feature name and it is as follows: 4 – Cluster Prominence;
6 - Dissimilarity; 15 - Difference variance; 16 - Difference entropy; 21 - Gray-level non-uniformity; 24 - Low graylevel run emphasis; 26 - Short run low gray-level emphasis; 28 - Long run low gray-level emphasis; 31 - Variance; 35 Skewness; 37 - Contrast (C ); 42 - Contrast (C ); 43 - Busyness; 44 - Complexity; 52 - I ; 54 - Contrst3; 55 1
0
7
Complexity; 56 - Roughness; 59 - Coronary lumen area; 61 - Coronary calcium area; 65 - Coronary wall thickness
variability, F – Feature.
TE
D
3.3 Experimental result 3: Memorization vs. generalization
Since the data size can affect the training coefficients, it is important to know when the
generalization is achieved [16,17]. In our study, we have varied the data size in 10 intervals ranging
from 493 to 4930 in the increment of 493 frames. These 493 frames are randomly selected and then
AC
C
EP
added to the corresponding ongoing pool.
14
RI
PT
ACCEPTED MANUSCRIPT
SC
Fig. 5. Numbers of dominant features vs. PCA-based cut-offs using (a) stand-alone plaque texturebased features and (b) plaque texture-based features fused with wall-based measurement features.
Since our cRAS system is highly dependent upon the grayscale morphological characteristics
M
AN
U
that change from patient-to-patient, and further the calcium deposit varies along with the wallbased measurements, it is, therefore, imperative to establish the metrics by which we can evaluate
the performance of our cRAS system. We, therefore, choose the following evaluation parameters:
(a) Dominant feature retaining power; (b) Receiver operating characteristic curve of the system; (c)
Reliability index; and (d) Stability, respectively. The whole idea is to understand the absolute and
relative performance along with the variations in changing parameters, which helps to understand
TE
D
the bounds of operation. For example, reliability reveals how accurate a system behaves for
different databases. Stability of the system is evaluated if the deviation of the mean accuracy is
within the tolerance limit (say 5%). We want to emphasize that PE is equally adopted for both: (a)
AC
C
EP
stand-alone plaque texture-based features and (b) plaque texture-based features fused with wall-
15
ACCEPTED MANUSCRIPT
based measurement features demonstrating the comparative approach and showing the
effectiveness of the wall-based measurement features on cRAS.
Table 2 Classification accuracy using SVM with varying kernel function (a) stand-alone plaque
texture-based features and (b) plaque texture-based features fused with wall-based measurement
features.
M
AN
U
SC
RI
PT
(a) Using stand-alone plaque texture-based features (in %)
Kernel functions
Cutoffs
Linear
RBF
Poly-1
Poly-2
Poly-3
67.14
76.20
67.14
71.87
74.86
0.90
67.14
76.18
67.14
71.89
74.89
0.91
67.15
76.15
67.15
71.91
74.85
0.92
73.02
87.53
73.02
77.37
82.95
0.93
73.06
87.56
73.06
77.39
82.96
0.94
75.48
89.66
75.48
81.30
85.68
0.95
75.48
89.65
75.48
81.32
85.72
0.96
78.47
91.32
78.47
82.88
86.41
0.97
77.91
92.74
77.91
84.20
89.46
0.98
79.80
93.82
79.80
86.63
92.03
0.99
Average
73.47
86.08
73.47
78.68
82.98
SD
4.87
7.12
4.87
5.45
6.22
(b) Using plaque texture-based features fused with wall-based measurement features (in %)
78.24
86.03
78.24
77.66
81.54
0.90
69.63
85.13
69.63
77.61
80.24
0.91
69.64
85.12
69.64
77.54
80.22
0.92
76.55
91.35
76.55
81.40
85.16
0.93
76.53
91.35
76.53
81.44
85.26
0.94
76.73
91.81
76.73
81.96
85.98
0.95
77.79
93.31
77.79
82.83
87.61
0.96
77.58
95.72
77.58
85.68
91.00
0.97
79.86
96.35
79.86
86.97
93.10
0.98
78.94
96.59
78.94
89.41
94.91
0.99
Average
76.15
91.28
76.15
82.25
86.50
SD
3.59
4.49
3.59
4.11
5.19
Poly - Polynomial
TE
D
Table 3 and Fig. 7 shows the variation of the SVM accuracy with varying data size for both: a)
stand-alone plaque texture-based features and (b) plaque texture-based features fused with wallbased measurement features. We observe that for all the data sizes, the mean accuracy of all the
AC
C
EP
PCA-based cutoffs, using the fusion of plaque texture-based and wall-based measurement features
16
ACCEPTED MANUSCRIPT
was higher compared to a stand-alone plaque texture-based features paradigm as now the dominant
TE
D
M
AN
U
SC
(a)
RI
PT
features are selected from a wider, diverse and strong pool of features.
(b)
Fig. 6. cRAS stratification accuracy vs. PCA-based cutoffs for five different kernel functions
(linear, RBF, polynomial-1, polynomial-2, and polynomial-3) using (a) stand-alone plaque texture-
AC
C
EP
based features and (b) plaque texture-based features fused with wall-based measurement features.
17
ACCEPTED MANUSCRIPT
Table 3 Average accuracy of each data sizes using (a) stand-alone plaque texture-based features
and (b) plaque texture-based features fused with wall-based measurement features.
M
AN
U
SC
RI
PT
(a) Using stand-alone plaque texture-based features (in %)
Data size
Cutoffs
493
986
1479
1972
2465
2958
3451 3944 4437 4930
100.00 100.00 100.00 100.00 100.00 100.00 71.06 49.90 76.60 76.20
0.90
100.00 100.00 100.00 100.00 100.00 100.00 71.03 50.00 76.56 76.18
0.91
100.00 100.00 100.00 100.00 100.00 100.00 91.37 92.98 90.32 76.15
0.92
100.00 100.00 100.00 100.00 100.00 100.00 91.39 92.98 90.32 87.53
0.93
100.00 100.00 100.00 100.00 100.00 100.00 95.85 95.01 91.71 87.56
0.94
100.00 100.00 100.00 100.00 100.00 100.00 95.85 95.01 91.66 89.66
0.95
100.00 100.00 100.00 100.00 100.00 100.00 98.09 96.88 94.96 89.65
0.96
100.00 100.00 100.00 100.00 100.00 100.00 98.47 97.42 95.52 91.32
0.97
100.00 100.00 100.00 100.00 100.00 100.00 98.47 98.03 96.56 92.74
0.98
100.00 100.00 100.00 100.00 100.00 100.00 99.27 98.17 97.51 93.82
0.99
Average 100.00 100.00 100.00 100.00 100.00 100.00 91.08 86.64 90.17 86.08
SD
0.00
0.00
0.00
0.00
0.00
0.00
10.92 19.43 7.60
7.12
(b) Using plaque texture-based features fused with wall-based measurement features (in %)
100.00 100.00 100.00 100.00 100.00 100.00 90.75 84.82 90.31 86.03
0.90
100.00 100.00 100.00 100.00 100.00 100.00 90.83 84.79 90.35 85.13
0.91
100.00 100.00 100.00 100.00 100.00 100.00 95.68 96.20 93.70 85.12
0.92
100.00 100.00 100.00 100.00 100.00 100.00 96.36 96.21 93.68 91.35
0.93
100.00 100.00 100.00 100.00 100.00 100.00 96.39 97.08 96.00 91.35
0.94
100.00 100.00 100.00 100.00 100.00 100.00 98.60 98.32 95.94 91.81
0.95
100.00 100.00 100.00 100.00 100.00 100.00 99.30 98.38 97.48 93.31
0.96
100.00 100.00 100.00 100.00 100.00 100.00 99.42 99.03 97.97 95.72
0.97
100.00 100.00 100.00 100.00 100.00 100.00 99.44 99.08 98.23 96.35
0.98
100.00 100.00 100.00 100.00 100.00 100.00 99.50 99.07 98.69 96.59
0.99
Average 100.00 100.00 100.00 100.00 100.00 100.00 96.63 95.30 95.24 91.28
SD
0.00
0.00
0.00
0.00
0.00
0.00
3.40
5.64
3.12
4.49
3.4 Dominant feature retaining power of cRAS
Dominant feature retaining power (DFRP) is the ability of the cRAS to retain the best
dominant features responsible for producing a high accuracy for different PCA-based cutoffs. It is
the ratio of similar dominant features between any two cutoffs say ‘m’ and ‘n’ ( ) and the
TE
D
number of dominant features selected taking the cutoff ‘m’ ( ). Using the notation * for the
product, we can mathematically compute DFRP in percentage as [17]:
AC
C
EP
(%) = ∗ 100
(1)
18
RI
PT
ACCEPTED MANUSCRIPT
SC
Fig. 7. Average accuracy vs. changing data size for K = 10 and T = 20 using (a) stand-alone plaque
texture-based features and (b) plaque texture-based features fused with wall-based measurement
features.
In this study, the DFRP taking different PCA-based cutoffs for both: (a) stand-alone plaque
texture-based features and (b) plaque texture-based features fused with wall-based measurement
M
AN
U
features are shown in Table 4. The mean DFRP of all the PCA-based cutoffs, using the fusion of
plaque texture-based and wall-based measurement features is almost similar (= 89.32%) as
compared to stand-alone plaque texture-based features (= 90.16%).
3.5 Receiver operating characteristics
True positive rate (Sensitivity) and false positive rate (100-Specificity) are mostly used to
TE
D
measure the diagnostic capability of the analysis. It is a way to identify how well the classification
methods can detect the true calcium. True positive rate (
) and false positive rate (
) can be
mathematically formulated as:
AC
C
EP
=
Truepositive
(Truepositive Falsenegative)
(2)
19
ACCEPTED
MANUSCRIPT
Truenegative
=
(3)
(Truenagative Falsepositive)
True positive/False positive are defined as the number of times the high-risk patient is
correctly/incorrectly identified with respect to the carotid plaque burden (gold standard) risk labels.
Similarly, True negative/False positive is defined as the number of times the high-risk patient is
incorrectly identified.
RI
PT
Table 4 Dominant feature retaining power calculation for different PCA-based cutoffs using (a)
stand-alone plaque texture-based features and (b) plaque texture-based features fused with wallbased measurement features.
TE
D
M
AN
U
SC
(a) Using stand-alone plaque texture-based features
Similar
dominant
Dominant
Dominant
Dominant feature retaining
Cutoffs (m and n)
features
Features at m
Features at n
power (in %)
(SDFm-n)
5
5
5
100.00
0.90 and 0.91
5
5
5
100.00
0.91 and 0.92
5
6
2
40.00
0.92 and 0.93
6
6
6
100.00
0.93 and 0.94
6
7
6
100.00
0.94 and 0.95
7
7
7
100.00
0.95 and 0.96
7
8
5
71.43
0.96 and 0.97
8
10
8
100.00
0.97 and 0.98
10
13
10
100.00
0.98 and 0.99
(b) Using plaque texture-based features fused with wall-based measurement features
5
6
2
40.00
0.90 and 0.91
6
6
6
100.00
0.91 and 0.92
6
7
5
83.33
0.92 and 0.93
7
7
7
100.00
0.93 and 0.94
7
8
7
100.00
0.94 and 0.95
8
9
8
100.00
0.95 and 0.96
9
10
8
88.89
0.96 and 0.97
10
12
10
100.00
0.97 and 0.98
12
16
11
91.67
0.98 and 0.99
Receiver operating characteristic can be graphically represented using true positive rate and
false positive rate and is generally used to quantify the diagnostic accuracy of the analysis. Using
the carotid plaque burden, we compute the receiver operating characteristic for the optimized PCA-
AC
C
EP
based cutoff (i.e., 0.99) for both: (a) stand-alone plaque texture-based features and (b) plaque
20
ACCEPTED MANUSCRIPT
texture-based features fused with wall-based measurement features as shown in Table 5, while the
corresponding visual curves are shown in Fig. 8. Note that we computed sensitivity, specificity,
positive predictive value, and area under the curve for the optimized kernel only which is RBF. As
can be seen, the ACU for plaque texture-based fused with wall-based measurement features was
0.91 compared to 0.86 for the stand-alone plaque texture-based cRAS system.
RI
PT
Table 5 Receiver operating characteristic for the highest PCA-based cutoff (= 0.99)
and Radial basis function kernel functions using (a) stand-alone plaque texturebased features and (b) plaque texture-based features fused with wall-based
measurement features.
(a) Using stand-alone plaque texture-based features
Mean Sensitivity
Mean Specificity
Mean PPV
Mean AUC
84.94±8.44
87.92±4.99
91.71±3.96
0.86±0.07
(b) Using plaque texture-based features fused with wall-based measurement features
90.74±5.22
92.14±3.42
94.86±2.39
0.91±0.04
TE
D
M
AN
U
SC
PPV - Positive predictive value, AUC – Area under the curve
AC
C
EP
Fig. 8. Receiver operating characteristic for the optimized PCA-based cutoff of 0.99
using (a) stand-alone plaque texture-based features and (b) plaque texture-based features
fused with wall-based measurement features.
21
ACCEPTED MANUSCRIPT
3.6 Reliability index of cRAS
In this study, the behavior of the system is analyzed by computing the reliability index (RI) of
the system for both: (a) stand-alone plaque texture-based features and (b) plaque texture-based
features fused with wall-based measurement features and is mathematically given as:
$% (%) = 1 −
'%
∗ 100
(%
(4)
where N is a set the 10 datasets ranging from 493 to 4930 in the increment of 493 frames, σ and µ
RI
PT
corresponds to the standard deviation and mean of each dataset computed for different PCA-based
cutoffs ranging from 0.90 to 0.99 for the optimized RBF kernel function. The overall reliability
index using the fusion of plaque texture-based and wall-based measurement features was higher (=
98.24%) as compared to stand-alone plaque texture-based features (= 94.86%), as shown in Table
6.
SC
Table 6 Reliability Index (RI) at different data sizes (N) using (a) stand-alone plaque texture-based
features and (b) plaque texture-based features fused with wall-based measurement features.
M
AN
U
(a) Using stand-alone plaque texture-based features
Data size
493
986
1479
1972
2465
2958 3451 3944 4437 4930 Average
RIN (%) 100.00 100.00 100.00 100.00 100.00 100.00 88.01 77.58 91.57 91.73
94.86
(b) Using plaque texture-based features fused with wall-based measurement features
RIN (%) 100.00 100.00 100.00 100.00 100.00 100.00 96.48 94.08 96.73 95.08
98.24
3.7 Stability of cRAS
In this study, we have also analyzed the stability of the system for both: (a) stand-alone plaque
texture-based features and (b) plaque texture-based features fused with wall-based measurement
TE
D
features. The deviation of the accuracy from the mean accuracy corresponding to all the PCA-based
cutoffs for each data sizes was computed. The mean deviation for all the data sizes using the fusion
of plaque texture-based with wall-based measurement features was under the tolerance limit of 5%
compared to stand-alone plaque texture-based cRAS, that came out to be under the tolerance limit
AC
C
EP
of 15%, as shown in Table 7.
22
ACCEPTED MANUSCRIPT
Table 7 Deviation of accuracy from mean accuracy for different data sizes using (a) stand-alone
plaque texture-based features and (b) plaque texture-based features fused with wall-based
measurement features.
M
AN
U
SC
RI
PT
(a) Using stand-alone plaque texture-based features (in %)
Data size
Cutoffs
493
986
1479
1972
2465
2958
3451
3944
4437
4930
0.00
0.00
0.00
0.00
0.00
0.00
20.02
36.74
13.57
9.88
0.90
0.00
0.00
0.00
0.00
0.00
0.00
20.05
36.64
13.61
9.9
0.91
0.00
0.00
0.00
0.00
0.00
0.00
0.29
6.34
0.15
9.93
0.92
0.00
0.00
0.00
0.00
0.00
0.00
0.31
6.34
0.15
1.45
0.93
0.00
0.00
0.00
0.00
0.00
0.00
4.77
8.37
1.54
1.48
0.94
0.00
0.00
0.00
0.00
0.00
0.00
4.77
8.37
1.49
3.58
0.95
0.00
0.00
0.00
0.00
0.00
0.00
7.01
10.24
4.79
3.57
0.96
0.00
0.00
0.00
0.00
0.00
0.00
7.39
10.78
5.35
5.24
0.97
0.00
0.00
0.00
0.00
0.00
0.00
7.39
11.39
6.39
6.66
0.98
0.00
0.00
0.00
0.00
0.00
0.00
8.19
11.53
7.34
7.74
0.99
Average
0.00
0.00
0.00
0.00
0.00
0.00
8.02
14.67
5.44
5.94
SD
0.00
0.00
0.00
0.00
0.00
0.00
6.92
11.75
5.00
3.38
(b) Using plaque texture-based features fused with wall-based measurement features (in %)
0.00
0.00
0.00
0.00
0.00
0.00
5.88
10.48
4.93
5.25
0.90
0.00
0.00
0.00
0.00
0.00
0.00
5.80
10.51
4.89
6.15
0.91
0.00
0.00
0.00
0.00
0.00
0.00
0.95
0.90
1.54
6.16
0.92
0.00
0.00
0.00
0.00
0.00
0.00
0.27
0.91
1.56
0.07
0.93
0.00
0.00
0.00
0.00
0.00
0.00
0.24
1.78
0.76
0.07
0.94
0.00
0.00
0.00
0.00
0.00
0.00
1.97
3.02
0.70
0.53
0.95
0.00
0.00
0.00
0.00
0.00
0.00
2.67
3.08
2.24
2.03
0.96
0.00
0.00
0.00
0.00
0.00
0.00
2.79
3.73
2.73
4.44
0.97
0.00
0.00
0.00
0.00
0.00
0.00
2.81
3.78
2.99
5.07
0.98
0.00
0.00
0.00
0.00
0.00
0.00
2.87
3.77
3.45
5.31
0.99
Average
0.00
0.00
0.00
0.00
0.00
0.00
2.63
4.20
2.58
3.51
SD
0.00
0.00
0.00
0.00
0.00
0.00
1.98
3.50
1.52
2.55
4. Discussion
This study demonstrated an ML risk assessment and stratification system, where it adopted a
fusion of plaque texture-based features with wall-based measurement features. ML system fusing
TE
D
plaque texture-based features with wall-based measurement features outperformed compared to
stand-alone plaque texture-based features. Thus, we validated our hypothesis. Because the
atheroma region causes the IEL and EEL walls to expand bidirectional [20], there was a clear
motivation to use wall-based measurement features. Further, since the atherosclerotic calcium is
AC
C
EP
multi-focal [2,21,35] and the detection process was well established [20], our cRAS system
23
ACCEPTED MANUSCRIPT
leverages on this burden to improve the overall accuracy of the risk stratification. Note that our
modeling assumes negligible effect due to heart motion [36] thereby assuming it to be simple,
pragmatic and ensuring high-speed processing due to the multiresolution approach for calcium
detection. Our system demonstrated a high accuracy of stratification based on the training model
that uses the concept of the genetic make-up of carotid plaque burden. Carotid plaque burden
(considered as a biomarker for stroke risk) can be used as a risk label for patients with coronary
artery disease, which validated our second hypothesis. Finally, we want to emphasize that the
RI
PT
cRAS system did take advantage of selection of the optimized features using polling strategy in
PCA paradigm ensuring best performance. From Fig. 5, it was observed that for different PCA-
based cutoffs, increase in the number of dominant features is higher for the fusion of plaque
texture-based and wall-based measurement features as compared to stand-alone plaque texture-
SC
based paradigm. Also in Table 1, the selection of coronary calcium area, coronary lumen area, and
coronary wall thickness as dominant features proves that wall-based measurement features are as
4.1 A note on population size
M
AN
U
important as plaque texture-based features in coronary artery risk assessment.
One of the requirements in ML framework is large enough population size for achieving
generalization while maintaining satisfactory accuracy for risk stratification. We had approached
our design strategy based on the number of frames rather than the number of patients as the
population pool during the cross-validation protocol design. In our ML system design, even though,
TE
D
we have a low population size of 22 subjects, but we had a total of 4930 frames. These 4930 frames
were stratified into 3043 high-risk frames and 1887 low-risk frames derived from 14 high-risk
patients and 8 low-risk patients, respectively. Our analysis demonstrated that our cRAS system has
AC
C
EP
enough pool of frames to design an accurate risk assessment system.
24
ACCEPTED MANUSCRIPT
4.2 A note on kernel functions
In Table 2, we can clearly observe that the lowest classification accuracy is obtained for linear
and polynomial order 1 kernel functions as the optimum separating hyperplane cannot separate all
the database into two distinct class. The classification accuracy of both linear and polynomial order
1 are same as there is not much difference between there kernel functions. With the increase in the
order of the polynomial kernel functions, we observe an increase in the classification accuracy as
now the size of the function class increases. Lastly, because of its Gaussian contribution, it was
4.3 A note on performance evaluation of our cRAS
RI
PT
observed that the highest classification accuracy is achieved with the RBF kernel function.
A prerequisite for an ML system is to understand the variability due to: (i) PCA-based
dominance feature selection and (ii) type of cross-validation protocol while fusing the wall-based
SC
features with grayscale morphological characteristics of plaque region derived from the IVUS
video frames. It is thus imperative to evaluate the cRAS system by understanding the metrics which
M
AN
U
evaluates the dynamics of the performance evaluation. We, therefore, took special steps in
computing four different performance parameters namely: (a) dominant feature retaining power; (b)
Receiver operating characteristic curves; (c) reliability index; and (d) stability under both: (a)
stand-alone plaque texture-based features and (b) plaque texture-based features fused with wallbased measurement features. These analyses were demonstrated in Table 4, Table 5, Table 6, and
Table 7, respectively. The above curves/tables demonstrate encouraging results on the design for
TE
D
risk assessment.
4.4 Comparison against current literature and benchmarking
IVUS is one of the speedily emerging medical imaging modality showing promising sign
towards high-resolution imaging [37]. This opens new doors for many medical image analysis
AC
C
EP
methods and techniques to extract and quantify plaque [20,23,29] and finally, to risk characterize
25
ACCEPTED MANUSCRIPT
the plaque into low and high-risk bins using ML-based strategies [16,17]. Not much work has been
done so far in the area of IVUS-based CAD risk stratification; however, several studies in the
literature have been seen with carotid plaque characterization and risk stratification. Table 8 shows
the comparison between these techniques using eight different attributes such as: year, artery type,
population size, and feature type, the number of features, feature selection techniques, classifier,
and cross-validation accuracy.
It was very recently that Araki et al. [38] in 2017 performed a risk stratification on the data
RI
PT
base of 407 carotid B-mode ultrasound images. Here, the SVM classifier trained by 16 texture
features provides the cross-validation accuracy of 98% for both near and far wall, respectively. The
same group [39] upgraded their system by utilizing the PCA-based pooling strategy for the
dominant feature selection and obtain a high accuracy of 98.55% and 98.83% for carotid near wall
SC
and carotid far wall, respectively.
While the above benchmarking used carotid plaque classification for risk estimation, our team
has been attempting to model coronary plaque risk stratification by fusing coronary and carotid
M
AN
U
atherosclerotic genetic makeup concept [22-28]. This study brings a novel approach of introducing
coronary wall-based measurement features along with grayscale coronary plaque texture-based
features for risk stratification. We, therefore, showed the cRAS for both: (a) stand-alone plaque
texture-based features and (b) plaque texture-based features fused with wall-based measurement
features. The fused system showed an improvement of 5.69%. To the best of our knowledge, this is
the first ML-based CADx system which utilizes a fusion of plaque texture-based and wall-based
AC
C
EP
TE
D
measurement features for coronary artery risk stratification.
26
RI
PT
ACCEPTED MANUSCRIPT
Table 8: Survey of risk stratification techniques in the literature.
Authors
Arterial
type
Population
size
Feature (s)
type
Total
features
Carotid
230
Tex
61
Carotid
274
Tex, Wall
10
Feature selection
technique (s)
Mean, SD,
Distance
NA
SF, Law's
21
Tex
Tex
54
Carotid
274
2012
Acharya et al. [12]
Carotid
346
2012
Acharya et al. [13]
2013
Acharya et al. [37]
2014
Pedro et al. [15]
2016
2016
2017
2017
Araki et al. [16]
Araki et al. [17]
Araki et al. [14]
Saba et al. [38]
2017
Our case
2007
Carotid
346
Carotid
492
Carotid
146
Coronary
Coronary
Carotid
Carotid
2865
2865
407
407
Coronary
4930
TE
D
2005
M
AN
U
Carotid
2009
Christodoulou et al.
[8]
Kyriacou et al. [9]
Mongiakakou et al.
[10]
Kyriacou et al. [11]
2003
SC
Year
Tex, Wall
DWT, HOS,
Tex
Rayleigh
Mixture
Tex
Tex
Tex
Tex
Only Tex, Tex
fused with Wall
Classifier (s)
Cross-validation accuracy
MNN
73.1%
NN
71.2%
ANOVA
HNN
99.1%
10
NA
4
T test
3
T test
SVM
SVM,
AdaBoost
SVM
73.7%
SVM – 82.4%, AdaBoost –
81.7%
83%
7
T test
SVM
91.7%
16
NA
EAI
77%
56
56
16
16
NA
PCA
NA
PCA
SVM
SVM
SVM
SVM
94.95%
98.43%
FW – 98.00%, NW – 98.00%
FW – 98.55%, NW – 98.83%
65
PCA
SVM
Only Tex – 86.08%,
Tex fused with Wall – 91.28%
AC
C
EP
Tex – Plaque texture-based, Wall –Wall-based, NN – Neural Network, SF – Statistical Feature, SD – Standard Deviation, MNN – Modular Neural Network, HNN –
Hybrid Neural Network, SVM – Support Vector Machine, EAI – Enhanced Activity Index, PCA – Principal Component Analysis, DWT – Discrete Wavelet Transform,
HOS – Higher Order Spectra, NW – Near Wall, FW – Far Wall, NA – Not Applicable.
27
ACCEPTED MANUSCRIPT
4.5 Carotid plaque burden as a gold standard for training-phase in ML design
One of the important components of the cRAS system design is the choice for gold standard
during the training-phase. The idea behind the choice of the gold standard is to ensure that we have
an indicator which has a strong link to the coronary artery disease while maintaining the low-cost
design of the cRAS system. Surely, one can think about the histology-based [40] or the calcium
score [41] using the CT as a gold standard. The CT results are not real-time and it is hard to
RI
PT
reconstruct them as a 3D image [42]. They are not easily affordable, truly expensive and tedious
protocols. Furthermore, the focus of our paper is solely based on ultrasound. The second option is
to adopt the concept of genetic make-up between carotid and coronary atherosclerosis disease,
which is now well established [22-28]. Previous studies had proven that carotid intima-media
thickness (cIMT) had influenced heart and death rate which proves the relationship of plaque
SC
burden in coronary and carotid arteries [43-48]. Ogata et al. [49] also showed the correlation of
cIMT with the plaque accumulation in the left main coronary artery. Establishing this concept, we
M
AN
U
thus leveraged our gold standard choice to be carotid artery and its plaque burden. One way to
establish this gold standard is to measure the media wall thickness as an indicator which can be
computed using cIMT measurements [38,39]. Suri has his team has shown numerous studies for
cIMT measurement and its link to various cardiovascular risk events such as ABI [25], syntax score
[26], etc. Since this study collected dual information such cIMT and coronary IVUS images, we,
therefore, used cIMT measurements as the risk label for our cRAS system.
TE
D
4.6 A note on time computation for online risk prediction
A machine learning system is characterized by its training (or offline) and testing (online)
phases. The time complexity of the risk assessment system is based on the hardware components
such as processor speed and computer RAM. The PC configuration of our system had the
AC
C
EP
following: HP Compaq Elite 8300 with Intel Core i7-3770 Processor, 3.40 GHz and 2 GB RAM,
28
ACCEPTED MANUSCRIPT
MATLAB 2013a software with Windows-7 Operating System. Typically, the offline system is not
accounted in the overall time complexity of the ML design, however, our single frame training and
testing time of the proposed cRAS system was 0.0435 seconds and 0.0083 seconds, respectively,
which can be considered reasonably fast looking at the previous risk assessment systems designed
for carotids and coronary applications [17].
4.7 Strength, limitations, and extensions
RI
PT
We demonstrated a machine learning system for risk assessment based on grayscale plaque
morphology padded with the wall-based features. We can characterize the benefits into two
categories: primary benefits and secondary benefits: The key benefit of our design is the
generalized system for risk assessment which can be extended to a more complex design by adding
meaningful features which are clinically more relevant. The successful main idea of coronary’s risk
SC
dependency based solely on genetic make-up can also be further be extended into other designs like
histology-based or CT-based biomarkers. This is another powerful strength of our cRAS system.
M
AN
U
The ability to fuse the wall-based measurement features with grayscale morphologic plaque
texture-based features provides another platform for fusion of feature paradigm. The secondary
benefits are the ability to obtain higher classification accuracy using an SVM-based classifier,
integration of PCA-based polling strategy for dominant feature extraction, ability to optimize the
best kernel function, ability to optimize the data for to obtain the best accuracy along with high
mean sensitivity, specificity, positive predictive value, and area under the curve. The system also
TE
D
has a high average feature retaining power and is reliable and stable.
Despite above strengths, the study suffers from some limitations. The study utilizes manual
tracings of IVUS images for the ROI generation. Dataset used in the current study is controlled as
all the patients come from a diabetic cohort. Intra/inter-observer variability could have been tried
AC
C
EP
but is out of the scope of current study as tracing all the frames is very expensive. DICOM images
29
ACCEPTED MANUSCRIPT
with gating and registration schemes can be incorporated with a large dataset for extensive
evaluations. Optical coherence tomography (OCT) is a high-resolution optical imaging technology
and provides more accurate arterial cross-section compared to IVUS [51]. We understand that there
is a need of further OCT/histology-based validation and automated segmentation of IVUS walls
[7], however, the current results are encouraging.
5. Conclusion
RI
PT
Coronary artery disease risk stratification tool based on IVUS wall grayscale morphological
characterization when fused with wall-based measurement features gave superior performance
using machine learning-based techniques. The system computed six novel wall-based measurement
features such as: coronary calcium area, coronary vessel area, coronary lumen area, coronary
atheroma area, coronary wall thickness, and coronary wall thickness variability, which were fused
SC
with grayscale features, gave an improvement of ~6% in the accuracy for predicting the class label
of the plaque type as high-risk or low-risk. All performance parameters showed similar behavior.
M
AN
U
Our cRAS system equally showed improvement in stability and reliability. Since the machine
learning system was automated, it can be adopted one step closer to clinical use for cardiovascular
imaging laboratories.
Conflict of interest: Dr. Jasjit S. Suri has a relationship with AtheroPoint™, Roseville, CA, USA
which is dedicated to Atherosclerosis Disease Management, including Cerebrovascular and
Cardiovascular Imaging.
AC
C
EP
TE
D
Contributions:
Sumit K. Banchhor: Programming, data analysis/interpretation, and drafting article.
Narendra D. Londhe: Advising and supporting in arranging the IT resources.
Tadashi Araki: Support in data collection and IRB approval.
Luca Saba: Radiological imaging and ground truth development.
Petia Radeva: Support in clinical demographics collection.
John R. Laird: Cardiologist and clinical atherosclerosis discussion.
Jasjit S. Suri: Concept/design and principal investigator of the project.
30
ACCEPTED MANUSCRIPT
Acknowledgements: The authors acknowledge Mr. Harman Suri from Mira Loma, Sacramento,
CA, USA for proof reading and providing the corrections to the manuscript.
Funding: This research did not receive any specific grant from funding agencies in the public,
commercial, or not-for-profit sectors.
References
1.
D. Mozaffarian, E.J. Benjamin, A.S. Go, et al., Executive summary: Heart Disease and Stroke
Statistics-2016 update: A report from the American Heart Association, Circulation 133 (2016)
RI
PT
447–454.
2.
R. Ross, Cell Biology of atherosclerosis, Annu. Rev. Physiol. 57 (1995) 791–804.
3.
P. Libby, P.M. Ridker, G.K. Hansson, Progress and challenges in translating the biology of
atherosclerosis, Nature 473 (2011) 317–325.
4.
K. Ramani, R.M. Judd, T.A. Holly, et al., Contrast magnetic resonance imaging in the
ventricular dysfunction, Circulation 98 (1998) 2687–2694.
P. Schoenhagen, R.D. White, S.E. Nissen, et al., Coronary imaging: angiography shows the
M
AN
U
5.
SC
assessment of myocardial viability in patients with stable coronary artery disease and left
stenosis, but IVUS, CT, and MRI show the plaque, Cleve. Clin. J. Med. 70 (2003) 713–720.
6.
P. Schoenhagen, S. Nissen, Understanding coronary artery disease: tomographic imaging with
intravascular ultrasound, Heart 88 (2002) 91–96.
7.
A. Katouzian, E.D. Angelini, S.G. Carlier, et al., A state-of-the-art review on segmentation
algorithms in intravascular ultrasound (IVUS) images, IEEE Trans. Inf. Technol. Biomed. 16
8.
TE
D
(2012) 823–834.
C.I. Christodoulou, C.S. Pattichis, M. Pantziaris, et al., Texture based classification on
AC
C
EP
atherosclerotic carotid plaques, IEEE Trans. Med. Imaging 22 (2003) 902–912.
31
9.
ACCEPTED MANUSCRIPT
E. Kyriacou, M.S. Pattichis, C.I. Christodoulou, et al., Ultrasound imaging in the analysis of
carotid plaque morphology for the assessment of stroke, Stud. Health Technol. Inform. 113
(2005) 241–275.
10. S.G. Mougiakakou, S. Golemati, I. Gousias, et al., Computer-aided diagnosis of carotid
atherosclerosis based on ultrasound image statistics, laws’ texture and neural networks,
Ultrasound Med. Biol. 33 (2007) 26–36.
11. E. Kyriacou, M. Pattichis, C.S. Pattichis, et al., Classification of atherosclerotic carotid plaques
RI
PT
using morphological analysis on ultrasound images, J. Appl. Intell. 30 (2009) 3–23.
12. R.U. Acharya, O. Faust, A.P.C. Alvin, et al., Symptomatic vs. asymptomatic plaque
classification in carotid ultrasound, J. Med. Syst. 36 (2012) 1861–1871.
13. U.R. Acharya, S.V. Sree, M.M. Krishnan, et al., Atherosclerotic risk stratification strategy for
SC
carotid arteries using texture-based features, Ultrasound Med. Biol. 38 (2012) 899–915.
14. U.R. Acharya, O. Faust, A.P. Alvin, et al., Understanding symptomatology of atherosclerotic
plaque by image-based tissue characterization, Comput. Methods Programs Biomed. 110
M
AN
U
(2013) 66–75.
15. L.M. Pedro, J.M. Sanches, J. Seabra, et al., Asymptomatic carotid disease-a new tool for
assessing neurological risk, Echocardiography 31 (2014) 353–361.
16. T. Araki, N. Ikeda, D. Shukla, et al., A new method for IVUS-based coronary artery disease
risk stratification: a link between coronary & carotid ultrasound plaque burdens, Comput.
Methods Programs Biomed. 124 (2016) 161–179.
TE
D
17. T. Araki, N. Ikeda, D. Shukla, et al., PCA-based polling strategy in machine learning
framework for coronary artery disease risk assessment in intravascular ultrasound: A link
between carotid and coronary grayscale plaque morphology, Comput. Methods Programs
Biomed. 128 (2016) 137–158.
AC
C
EP
18. U.R. Acharya, S.V. Sree, F. Molinari, et al., An automated technique for carotid far wall
32
ACCEPTED MANUSCRIPT
classification using grayscale features and wall thickness variability, J. Clin. Ultrasound. 43
(2015) 302–11.
19. L. Saba, K.M. Meiburger, F. Molinari, et al., Carotid IMT variability (IMTV) and its validation
in symptomatic versus asymptomatic Italian population: can this be a useful index for studying
symptomaticity?, Echocardiography 29 (2012) 1111–9.
20. S.K. Banchhor, N.D. Londhe, T. Araki, et al., Relationship between Automated Coronary
Calcium Volumes and a set of Manual Coronary Lumen Volume, Vessel Volume and
RI
PT
Atheroma Volume in Japanese Diabetic Cohort: An Intravascular Ultrasound-based Study, J.
Clin. Diagn. Res. 11 (2017) TC09–TC14.
21. W. Casscells, M. Naghavi, J.T. Willerson, Vulnerable atherosclerotic plaque a multifocal
disease, Circulation 107 (2003) 2072–2075.
SC
22. F. Molinari, K.M. Meiburger, G. Zeng, et al., Automated carotid IMT measurement and its
validation in low contrast ultrasound database of 885 patient Indian population epidemiological
study: results of AtheroEdge™ Software, Int. Angiol. 31 (2012) 42–53.
M
AN
U
23. L. Saba, S.K. Banchhor, H.S. Suri, et al., Accurate cloud-based smart IMT measurement, its
validation and stroke risk stratification in carotid ultrasound: A web-based point-of-care tool
for multicenter clinical trial, Comput. Biol. Med. 75 (2016) 217–34.
24. L. Saba, R. Sanfilippo, M. di Martino, et al., Volumetric Analysis of Carotid Plaque
Components and Cerebral Microbleeds: A Correlative Study, J. Stroke Cerebrovasc. Dis. 26
(2017) 552–558.
TE
D
25. N. Ikeda, R. Kogame, M. Iijima, et al., Impact of carotid artery ultrasound and ankle-brachial
index on prediction of severity of SYNTAX score, Circ. J. 77 (2013) 712–716.
26. N. Ikeda, A. Gupta, N. Dey, et al., Improved correlation between carotid and coronary
atherosclerosis SYNTAX score using automated ultrasound carotid bulb plaque IMT
AC
C
EP
measurement, Ultrasound Med. Biol. 41 (2015) 1247–1262.
33
ACCEPTED MANUSCRIPT
27. T. Araki, N. Ikeda, N. Dey, et al., Calcium lesion volume measurement on intravascular
ultrasound imaging and its association with carotid intima-media thickness, J. Ultrasound Med.
34 (2015) 469–482.
28. T. Araki, N. Ikeda, N. Dey, et al., A comparative approach of four different image registration
techniques for quantitative assessment of coronary artery calcium lesions using intravascular
ultrasound, Comput. Methods Programs Biomed. 118 (2015) 158–172.
29. T. Araki, S.K. Banchhor, N.D. Londhe, et al., Reliable and accurate calcium volume
RI
PT
measurement in coronary artery using intravascular ultrasound videos, J. Med. Syst. 40 (2016)
51.
30. S.K. Banchhor, T. Araki, N.D. Londhe, et al., Five multiresolution-based calcium volume
measurement techniques from coronary IVUS videos: A comparative approach, Comput.
SC
Methods Programs Biomed. 134 (2016) 237–58.
31. S.K. Banchhor, N.D. Londhe, T. Araki, et al., Well-Balanced System for Coronary Calcium
Detection and Volume Measurement in a Low Resolution Intravascular Ultrasound Videos,
M
AN
U
Comput. Biol. Med. 84 (2017) 168–181.
32. J.S. Suri, R.M. Haralick, F.H. Sheehan, Greedy algorithm for error correction in automatically
produced boundaries from low contrast ventriculograms, Pattern Anal. Appl. 3 (2000) 39–60.
33. L. Saba, F. Molinari, K.M. Meiburger, et al., Inter- and intra-observer variability analysis of
completely automated cIMT measurement software (AtheroEdge™) and its benchmarking
1261–1272.
TE
D
against commercial ultrasound scanner and expert readers, Comput. Biol. Med. 43 (2013)
34. V. Vapnik, Statistical learning theory. Vol 1, New York, Wiley, 1998.
35. I. Tabas, K.J. Williams, J. Borén, Subendothelial lipoprotein in retention as the initiating
process in atherosclerosis: update and therapeutic implications, Circulation. 116 (2007) 1832–
AC
C
EP
1844.
34
ACCEPTED MANUSCRIPT
36. S.K. Setarehdan, S. Singh, Advanced Algorithmic Approaches to Medical Image
Segmentation: State-of-the-art Applications in Cardiology, Neurology, Mammography and
Pathology. Springer-Verlag London, 2002.
37. C.M. Campos, R.J. Fedewa, H.M. Garcia-Garcia, et al., Ex vivo validation of 45 MHz
intravascular ultrasound backscatter tissue characterization, Eur. Heart J. Cardiovasc. Imaging
16 (2015) 1112–1119.
38. T. Araki, P.K. Jain, H.S. Suri, et al., Stroke Risk Stratification and its Validation using
Comput. Biol. Med. 80 (2017) 77–96.
RI
PT
Ultrasonic Echolucent Carotid Wall Plaque Morphology: A Machine Learning Paradigm,
39. L. Saba, P.K. Jain, H.S. Suri, et al., Plaque Tissue Morphology-based Stroke Risk Stratification
using Carotid Ultrasound: A Polling-based PCA Learning Paradigm, J. Med. Syst. 41 (2017)
SC
98.
40. D.S. Scott, U.K. Arora, A. Farb, et al., Pathologic validation of a new method to quantify
coronary calcific deposits in vivo using intravascular ultrasound, Am. J. Cardiol. 85 (2000) 37–
M
AN
U
40.
41. N. Schamroth, A. Gutstein, R. Kornowski, Coronary Artery Calcium Score: Where Do We
Stand? Current Uses and Implications in Asymptomatic Patients, Isr. Med. Assoc. J. 19 (2017)
214–215.
42. A. Wahle, M. Sonka, Coronary plaque analysis by multimodality fusion. Stud Health Technol
Inform. In: Suri JS et al. (Eds.), Plaque Imaging: Pixel to Molecular Level, 321–359, 2004.
TE
D
43. F. Molinari, G. Zeng, J.S. Suri, Intima-media thickness: setting a standard for a completely
automated method of ultrasound measurement, IEEE Trans. Ultrason. Ferroelectr. Freq.
Control 57 (2010) 1112–1124.
44. M.J. Sanches, A. Laine, J.S. Suri, Ultrasound Imaging. Springer Science & Business Media,
AC
C
EP
New York, 2011.
35
ACCEPTED MANUSCRIPT
45. L. Saba, J.M. Sanches, L.M. Pedro, et al., Multi-modality atherosclerosis imaging and
diagnosis. Springer, New York, 2014.
46. L. Saba, N. Ikeda, M. Deidda, et al., Association of automated carotid IMT measurement and
HbA1c in Japanese patients with coronary artery disease, Diabetes Res. Clin. Pract. 100 (2013)
348–353.
47. N. Ikeda, T. Araki, K. Sugi, et al., Ankle–Brachial Index and Its Link to Automated Carotid
Artery Disease Patients, Curr. Atheroscler. Rep. 16 (2014) 1–8.
RI
PT
Ultrasound Measurement of Intima–Media Thickness Variability in 500 Japanese Coronary
48. T. Nezu, N. Hosomi, S. Aoki, et al., Carotid Intima-Media Thickness for Atherosclerosis,
Journal of Atherosclerosis and Thrombosis 23 (2015) 18–31.
49. T. Ogata, M. Yasaka, M. Yamagishi, et al., Atherosclerosis found on carotid ultrasonography
SC
is associated with atherosclerosis on coronary intravascular ultrasonography, J. Ultrasound
Med. 24 (2005) 469–474.
50. P.A. Lacroix, V. Aboyans, E. Espaliat, et al., Carotid intima-media thickness as predictor of
M
AN
U
secondary events after coronary angioplasty, Int. Angiol. 22 (2003) 279–283.
51. M. Akhtar and W. Liu, 2016. Use of intravascular ultrasound vs. optical coherence
tomography for mechanism and patterns of in-stent restenosis among bare metal stents and
AC
C
EP
TE
D
drug eluting stents, J. Thorac. Dis. 8 (2016) 104–108.
36
ACCEPTED MANUSCRIPT
Highlights
•
Amalgamation of IVUS plaque texture-based and wall-based measurement features.
•
Principle component analysis-based framework was used for dimensionality
reduction.
During the training phase, carotid plaque burden was chosen as a gold standard.
•
Support vector machine was used as a classifier for training and testing-phases.
•
Proposed ML system demonstrate improvement in risk stratification accuracy.
AC
C
EP
TE
D
M
AN
U
SC
RI
PT
•
ACCEPTED MANUSCRIPT
Authors Biographic
SC
RI
PT
Sumit K. Banchhor, MTech received his BE degree in Electronics and
Telecommunication engineering from the Pt. Ravishankar Shukla University,
Bhilai, Chhattisgarh in 2007 and MTech degree in Digital Electronics from
Chhattisgarh Swami Vivekanand Technical University, Bhilai, Chhattisgarh in
2011. He has been working toward the Ph.D. degree since 2014 from
Department of Electrical Engineering of National Institute of Technology,
Raipur, India.
M
AN
U
Dr. Narendra D Londhe, PhD received his BE degree from Amravati
University in 2000. Later he received his MTech and PhD degrees in the year
2004 and 2011, respectively from Indian Institute of Technology Roorkee. He is
presently working as Assistant Professor in Department of Electrical Engineering
of National Institute of Technology, Raipur, India.
EP
TE
D
Dr. Tadashi Araki, MD received the MD degree from Toho University, Japan
in 2003. His research topics include coronary intervention, intravascular
ultrasound (IVUS) and peripheral intervention. Now, he works in Toho
University Ohashi Medical Center, Tokyo, Japan as coronary and peripheral
interventionalist.
AC
C
Dr. Luca Saba, MD received the MD degree from the University of Cagliari,
Italy in 2002. Today he works in the A.O.U. of Cagliari. Dr. Saba research fields
are focused on Neuroradiology, Multi-Detector-Row Computed Tomography,
Magnetic Resonance, Ultrasound, and Diagnostic in Vascular Sciences. His
works, as lead author, achieved more than 75 high impact factor, peer-reviewed,
Journals. Dr. Saba has written 7 book chapters and he presented more than 400
papers in National and International Congress. Dr. Saba is a member of the
Italian Society of Radiology (SIRM), European Society of Radiology (ESR),
Radiological Society of North America (RSNA), American Roentgen Ray
Society (ARRS) and European Society of Neuroradiology (ESNR).
0
ACCEPTED MANUSCRIPT
RI
PT
Dr. Petia Radeva, PhD is a senior researcher and associate professor at the
University of Barcelona. She received her Ph.D. degree from the Universitat
Autònoma de Barcelona in 1998. She is the head of Barcelona Perceptual
Computing Laboratory (BCNPCL) at the University of Barcelona and the head
of MiLab of Computer Vision Center. Her present research interests are on the
development of learning-based approaches (in particular, statistical methods) for
computer vision and image processing. Some of the projects she is currently
heading are: Machine learning tools for large scale object recognition, Sponsored
Research Agreement on Automatic Stent Detection in IVUS, Study for the
development of polyp detection algorithms, Audience measurements by
Computer Vision, Evaluation of Intestinal Motility by Endoluminal Image
Analysis.
M
AN
U
SC
Dr. John Laird, MD is an internationally renowned interventional cardiologist
who has lectured and performed endovascular procedures around the world. He
has also been a leader in the development of Drug Coated Balloon for peripheral
artery disease. His practice focuses on interventions for carotid artery disease,
abdominal and thoracic aortic aneurysmal disease, renal artery disease and
peripheral artery disease. Among his many areas of leadership and innovation are
the investigation of new stents for carotid and peripheral vascular applications
and the development of laser angioplasty. He uses the Excimer laser as a tool for
complex peripheral interventions and in limb salvage situations. Dr. Laird is
known for innovation, teaching vascular interventions to other physicians, and
his role in organizing and conducting clinical trials of new therapies for vascular
disease.
AC
C
EP
TE
D
Dr. Jasjit S. Suri, PhD, MBA, Fellow AIMBE is an innovator, visionary,
scientist, and an internationally known world leader. Dr. Suri received the
Director General’s Gold medal in 1980 and the Fellow of American Institute of
Medical and Biological Engineering, awarded by National Academy of Sciences,
Washington DC in 2004. He is currently Chairman of Global Biomedical
Technologies, Inc., Roseville, CA, USA. He has published over 500 peerreviewed articles and book chapters and over 100 innovations/trademarks. He is
currently Chairman of AtheroPoint and Global Biomedical Technologies, Inc.,
Roseville, CA, USA.
1
Документ
Категория
Без категории
Просмотров
2
Размер файла
3 375 Кб
Теги
2017, 019, compbiomed
1/--страниц
Пожаловаться на содержимое документа