Добавил:
Sekretar
kiopkiopkiop18@yandex.ru
t.me/Prokururor I Вовсе не секретарь, но почту проверяю
Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз:
Предмет:
Файл:Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_3592_Библиотеки_им_академика_М_И_Перельмана
.pdf
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 12.10. Scatter diagram showing a mild correlation between age and cIMT readings. Panels (a1), (b1)
and (c1) show a correlation between age and left cIMT and panels (a2), (b2) and (c2) show a correlation
between age and right cIMT using AtheroCloud™, manual and sonographer’s readings, respectively.
12.5.7 Receiver operating characteristic (ROC)
The sensitivity and specificity are computed using the following four parameters:
true positive (TP), true negative (TN), false positive (FP) and false negative (FN).
TP is defined as the number of times cIMT correctly identified with respect to the
manually computed cIMT for the cut-off risk threshold. FN is defined as the number
12-18

Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 12.11. Combined cumulative distribution error curves for (i) AtheroCloud™ versus manual,
(ii) sonographer versus manual and (iii) AtheroCloud™ versus sonographer.
Table 12.4. Statistical tests between AtheroCloud™, manual and sonographer cIMT readings.
Mann–Whitney
Combinations
AtheroCloud™
versus manual
Sonographer
versus manual
AtheroCloud™
versus manual
Sonographer
versus manual
Two-tailed z-test Chi-squared-test
Contingency
zp-value
−1.06 < 0.2888 0.984 < 0.0001 = 0.1419
−2.13 < 0.0328 0.950 < 0.0001 = 0.0976
Right cIMT
−0.71 < 0.4795 0.985 = 0.0002 = 0.4321
−2.84 < 0.0045 0.946 = 0.0002 = 0.0079
coefficient p-value p-value
Left cIMT
of times cIMT is incorrectly identified w.r.t the manually computed cIMT for the
cut-off risk threshold. TN and FP are defined as the number of times cIMT is
correctly or incorrectly identified for cut-off risk threshold, respectively. The
sensitivity and specificity can be mathematically defined as
12-19

Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Sensitivity
=
(TP FN)
TP
+
and Specificity
=
TN
(TN FP)
+
.
(12.3)
The ROC can be plotted by using two parameters: (i) the true positive rate
(sensitivity) and (ii) the false positive rate (specificity). The area under the ROC
curve (AUC) is the ability of the test to correctly classify readings into two
diagnostic groups (diseased/normal). A higher AUC justifies the accuracy of the
system. The ROC analysis was performed on the AtheroCloud™ and sonographer
methods against the manual readings, as shown in figure 12.12.
As most of the previous clinical studies have shown that the cIMT cut-off risk
threshold for cerebrovascular accident (stroke) or CVE (heart attack) is widely
distributed, ranging from 0.7 mm [23] to 1.26 mm [28] with an average cut-off of
0.9 mm [25, 26], we, therefore, evaluate the AtheroCloud™ reading and sonographer’s reading against the cIMT cut-off risk threshold of 0.9 mm. The difference in
cIMT is due to variations in clinical demographics and focus of these studies.
Clinical demographics include age, gender, ethnicity, body mass index, LDL, HA1c,
the presence of CAD and family history for CVD. Using the cIMT cut-off risk
threshold of 0.9 mm, the AUC for AtheroCloud™ and sonographer readings were:
0.99 and 0.81, respectively.
12.5.7.1 Operating point computation
Figure 12.13(a) and (b) shows the comparison plots of sensitivity and specificity for
AtheroCloud™ and sonographer cIMT readings, respectively. These plots were
computed by taking the manual cIMT readings with a cut-off of 0.9 mm. These were
Figure 12.12. Receiver operating characteristic curve analysis for AtheroCloud™ versus sonographer cIMT
readings. The AUCs are 0.99 and 0.81 for AtheroCloud™ and sonographer readings.
12-20

Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
used to compute the operating point for both these methods. The operating point
(also called equal error rate) is the point of maximum sensitivity and specificity for
the system. Using the equal error rate concept, the operating point for
AtheroCloud™ and sonographer cIMT readings were 0.83 mm and 0.9 mm,
respectively. The corresponding specificity and sensitivity of AtheroCloud™ were
98.63% and 91.34%, while for the sonographer system they were 76.71% and
76.38%, respectively.
Figure 12.13. Sensitivity and specificity curves for (a) AtheroCloud™ and (b) sonographer cIMT and their
respective operating points.
12-21

Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 12.14(a) and (b) shows the interactive dot diagram for AtheroCloud™ and
sonographer cIMT readings with reference to manual cIMT. An interactive dot
diagram is a dual plot with a horizontal line showing the cut-off point with the best
separation (minimum false negative and false positive results) between the two
readings.
Figure 12.14. Interactive dot diagram for (a) AtheroCloud™ and (b) sonographer cIMT and their respective
threshold, sensitivity and specificity. Classification of AtheroCloud and sonographer cIMT in high-risk and
low-risk bins.
12-22

Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
The data of the negative and positive groups are shown as dots on the two vertical
axes. Figure 12.14(a) shows the AtheroCloud™’s two clusters corresponding to lowand high-risk represented by 0 and 1 in the plot with manual readings as a reference,
with a cut-off of 0.83 mm. Figure 12.14(b) shows the sonographer’s two clusters
corresponding to low- and high-risk represented by 0 and 1 in the plot with manual
reading as a reference, with a cut-off of 0.9 mm.
12.5.8 Risk stratification
Risk stratification is a tool to recognize high-risk patients for better management of
carotid disease and stroke. It can identify and quantify differences and changes in
the risk profiles of patients suffering from coronary artery disease [54]. In this study,
we have used the conventional Framingham risk score (FRS) to stratify the
population into low-, medium- and high-risk bins.
12.5.9 Framingham risk score
The role of the FRS is to identify the patient’s chances of developing the
cardiovascular disease in a specified period of time, typically between 10 to 30
years [55, 56]. Further, the FRS of the patient indicates if they are likely to benefit
from risk prevention strategies such as early screening. The FRS has been slightly
controversial because it does not take cIMT into account, but it has shown which
individual patients are likely to benefit from which kinds of drugs.
A risk score is derived for each patient using the gender-specific prediction
formulae proposed by Wilson [57], based on the following conventional cardiovascular risk factors: age, total cholesterol and HDL cholesterol, systolic blood
pressure, diabetes and smoking status. The values adapted for these parameters
were listed in the demographics section. For computing the FRS, we took all the
parameters concerned except blood pressure, which was not available for all the
patients during the retrospective study. At 10 years, the CHD risk for individuals
with low Framingham risk is 10% or less, intermediate Framingham risk 10% to
20%, and high Framingham risk have 20% or more CHD risk. Previous studies [55]
have shown that the FRS has been validated in the United States for both genders in
both European Americans and African Americans. Even though there are studies
which have claimed improvement via the FRS, but there is little evidence for
improved prediction using the FRS [55 ]. We did not obtain any relationship between
cIMT and FRS scores, but for the sake of relevance, we have computed FRS and
given stratification ranges.
12.5.9.1 Risk stratification based on the FRS
Table 12.5 shows the 10 year mortality rate over the range of the FRS in both men
and women. In men, 70.66%, 18.66% and 10.66% of the total population had FRS
risk levels < 10%, 10% to 19%, and ⩾ 20%, respectively. For women, the entire
population is under the 10% risk level.
12-23

Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Table 12.5. 10 year mortality rates in men and women by the FRS.
FRS Men Women
Low < 10% 14 (18.66%) 25 (100%)
Intermediate 10%–19% 53 (70.66%) —
High ⩾ 20% 8 (10.66%) —
Figure 12.15. Scatter diagram showing a high correlation between age and FRS readings for (a) males and (b)
females, respectively.
12.5.9.2 Age versus FRS
The scatter diagram of age versus FRS for males and females is shown in figure
12.15(a) and (b), respectively. We can observe a higher CC of female age with the
FRS (0.70) compared to males (0.61). Although the CC was very high between age
and FRS, there was no relationship between cIMT and FRS for this study.
12.6 Discussion
12.6.1 Our system
The main objectives of this pilot study were to:
• Propose (a) an automated and (b) user-interactive (semi-automated) cloudbased software system for cIMT measurements in ultrasound carotid scans
using (i) Routine (single) mode and (ii) Pharmaceutical (batch) mode,
intended for patient visit and clinical trials, respectively.
• Validate a cloud-based system against (a) manual tracings (gold standard)
and (b) comparing Routine mode to Pharmaceutical mode for reproducibility
evaluation for the entire database.
• Compare and contrast the cloud-based automated system against sonographer readings by performing exhaustive statistical analysis consisting of PoM,
CC and ROC/AUC curve analysis compared to manual tracings.
12-24

Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
• Benchmark the AtheroCloud™ software system for cIMT measurement
against a desktop-based cIMT measurement system—AtheroEdge™
(AtheroPoint, Roseville, CA, USA) that has been previously benchmarked
[33, 43–45].
• Establish the reliability of the AtheroCloud™ system through statistical tests
such as the two-tailed paired Student’s t-test, two-tailed z-test and Wilcoxon
test.
• Risk assessment and stratification of the population using the FRS.
• Verification of the AtheroCloud™ software application against its functional
requirements, such as automated cIMT report generation in the Routine and
Pharmaceutical modes, and reproducibility between the Routine and
Pharmaceutical modes.
This study shows the accuracy, reliability and functionality for the ultrasoundbased AtheroCloud™ system, which can be adapted for multicenter clinical trials.
Because our carotid scans did not have any bulbs, no reference point was chosen,
and the cIMT was determined all along the carotid artery. We further demonstrated
that mean cIMT readings were similar for the Routine mode and Pharmaceutical
mode, even though the Pharmaceutical mode accepted the entire batch of carotid
scans.
12.6.2 Benchmarking AtheroCloud™ against AtheroEdge™
As seen previously, our system measures cIMT, which is an automated, anytime –
anywhere solution. We demonstrated that the AtheroCloud™ system is more
accurate than sonographer readings. The system was capable of measuring cIMT
ranging up to a 2.0–2.5 mm thickness (heavy plaque in the distal wall) based on
advanced scale-space methods as developed by HSS’s team [3, 11, 13, 54, 58–60]. In
this study, we have compared our proposed AtheroCloud™ system against
AtheroEdge™—a desktop-based commercial software (courtesy of AtheroPoint,
Roseville, CA, USA). The results from a sample Pharmaceutical mode run are
shown in figure 12.16.
We ran the same database pool using the AtheroEdge™ system and compared the
results against the AtheroCloud™ system. The results can be seen in table 12.6.
Column ‘A’ shows the output for the proposed AtheroCloud
™ system and column
‘B’ shows the commercial desktop system AtheroEdge™. As shown in the table, the
Pharma modes for cloud-based versus desktop-based showed nearly similar results:
0.86 ± 0.20 mm versus 0.88 ± 0.20 mm for left cIMT, and 0.77 ± 0.20 mm versus
0.80 ± 0.20 mm for right cIMT, respectively. There was a difference of 2.27% for the
left artery and 3.75% for the right artery. Typically, a 5% error rule is adapted for
regulatory settings and we showed that our system qualified under that benchmarking requirement. Furthermore, we showed that the results were reproducible. We
matched our Pharmaceutical electronic batch reports between the AtheroCloud™
and AtheroEdge™ systems for all the patient ultrasound scans. The two-tailed
paired Student’s t-test, two-tailed z-test and Wilcoxon test demonstrated consistency
12-25

Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 12.16. Pharmaceutical Trial mode automated tracings (yellow) of the carotid intima–media thickness
region showing LI in red and MA in green using AtheroEdge™ software. (Courtesy of AtheroPoint™,
Roseville, CA, USA.)
Table 12.6. Benchmarking of AtheroCloud™ (Routine and Pharmaceutical mode) against AtheroEdge™
(desktop-based system).
AtheroCloud™ (cloud-based)
Neck side
Left cIMT (mm) 0.87 ± 0.20 0.86 ± 0.20 0.88 ± 0.20
Right cIMT (mm) 0.77 ± 0.20 0.77 ± 0.20 0.80 ± 0.20
Routine mode Pharma mode Pharma mode
AtheroEdge™ (desktop-based)
and reliability, making it the most practical and clinically adaptable system. To our
knowledge, there is no existing pilot or commercial study of cloud-based cIMT
measurements for stroke/cardiovascular applications, so we have used desktopbased comparisons.
12.6.3 A brief survey of previous techniques
In past years, many studies have proposed techniques to measure IMT in carotid
ultrasound scans. In 2007, Delsanto et al [61] proposed a Completely User-
12-26

Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
independent Layer EXtraction algorithm based on signal approach (CULEXsa) for
carotid artery segmentation in 2D ultrasound images. The snake-based technique
was used and a segmentation error of less than one pixel was achieved. However, the
algorithm suffers from noise and image artefacts.
In 2010, Molinari et al [13] proposed another automated algorithm called
Completely Automated Layered EXtraction technique based on integrated
approach (CALEXia) for IMT measurement and tried to eliminate the limitations
of their previous work. CALEXia was based on an integrated approach of feature
extraction, line fitting and classification. The IMT measurement error was equal to
0.87 ± 0.56 pixels for CALEXia and 0.12 ± 0.14 pixels for CULEXsa. Despite its
limited performance in segmenting the LI interface, CALEXia outperformed
CULEXsa in segmenting the MA interface. CALEXia also performed correctly in
95% of the tested images (against 92% of CULEXsa). The CALEXia technique was
further validated by the same group on a large database of 200 images [14, 62]. The
study proved CALEXia to be a robust technique for computer-based automated
tracing of the CCA in longitudinal B-mode carotid ultrasound images. However,
more work is needed to increase the performance of the IMT measurement.
In 2011, the groups led by HSS proposed a series of automated IMT measurement techniques, namely: (i) completely automated robust edge snipper (CARES)
[63], (ii) the inter-greedy (IG) method [59], (iii) completed automated local statisticsbased fi rst-order absolute moment (CLASFOAM) [48], (iv) ultrasound double line
extraction system using edge flow (CAUDLES-EF) [64] and (v) carotid artery
intima layer regional segmentation (CAILRS) [38].
In the first technique, i.e. CARES [63], the authors used an integrated approach of
intelligent image feature extraction and line fitting for automatically locating the
carotid artery in the image frame. A Gaussian edge operator was then used for
extracting the wall interfaces. Using 300 carotid ultrasound images, the study
showed an IMT bias of 0.032 ± 0.141 mm. As compared to the IMT bias of their
previously published method CALEXia, CARES improved the IMT accuracy by
67%, while increasing the standard deviation by 3%.
To remove their bias errors, in their second technique, i.e. the IG method [59], an
IG approach was applied to the LI and MA boundaries traced by the three image
segmentation techniques (namely, CALEXia, CULEXsa and Watershed transform
(WS)). The validation was done against manual tracings on a large dataset of 200
images. With the IG method, an IMT error of 0.74 ± 0.75 pixels was observed,
which showed an improvement of 44% against their previous proposed CULEXsa
algorithm.
The third technique, i.e. CLASFOAM [48], was developed to overcome the
limitations of a previously developed snake-based technique. The CALSFOAM
technique consisted of two stages. Stage-I performs automatic recognition of the
carotid artery system by using local statistics and by automatically tracing the profile
of the distal adventitia in an image frame. Stage-II builds an ROI for distal wall
segmentation by using a first-order absolute moment (FOAM) technique.
CALSFOAM was validated against manual tracings on a 300 image multi-institutional
dataset and showed an IMT measurement bias of 0.125 ± 0.103 mm. By fusing
12-27
Соседние файлы в папке Библиотека им академика М.И. Перельмана
