Добавил:
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
Скачиваний:
0
Добавлен:
29.08.2026
Размер:
89 Мб
Скачать
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Testing Near Wall and Far Wall Strips
Grayscale Feature Extraction System
Grayscale Texture Features
Dominant Feature Extraction Using
Statistical Method
Dominant Features
Risk
Stratification
(Classifier)
Low Risk
High Risk
Stenotic Features
as Ground Truth
Training Near Wall and Far Wall Strips
Grayscale Feature Extraction System
Grayscale Texture Features
Dominant Feature Extraction Using
Statistical Method
Dominant Features
Offline Classification (SVM)
Machine Learning Parameters
Offline System Online System
Figure 13.8. Carotid disease risk assessment system for the near wall and the far wall.
morphological features by the ofine training parameters. Thus the basic model uses the cross-validation approach for computing the accuracy, sensitivity and specicity of the sRAS using three kinds of protocols: K5, K10 and JK. This will be discussed in the experimental protocol section.
13.3.3 Texture features
13.3.3.1 Gray-level co-occurrence matrix (GLCM)
Using the statistical tool, one can use GLCM for extracting textural information of the ultrasound image by considering the neighborhood pixel relationship [54, 55]. Considering the grayscale 2D image to be represented by I having the gray levels
…−L
0, 1, , 1),one can compute the
(i, j)th entry of Mxrepresents the probability of the number of occasions a pixel
P
d
g
with intensity i is adjacent to a pixel with intensity j. Dividing each element of M the total number of co-occurrence pairs in M occurrence matrix. One can compute the adjacency by taking any specied direction such as horizontal, vertical, right, left and diagonal. Finally, the texture features were computed by taking the average of the chosen direction of the co-occurrence matrix. We extracted four kind of features, shown in appendix A, table A1.
13.3.3.2 Gray-level run length matrix (GLRLM)
Run length is dened as a set of collinear pixels having the same gray level in a particular direction [56]. Given the reference pixel, one can compute GLRLM and
GLCM matrix Mxof order L, where, the
by
x
will yield the normalized co-
x
13-10
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
this measures the gray intensity pixel in a particular direction. GLRLM is a 2D matrix in which element p(x, y) gives the total number of consecutive runs of length y at gray level x. A total of 11 features were extracted using GLRLM, as shown in appendix A, table A2. Note that M represents the number of gray levels and L represents the maximum run length.
13.3.3.3 Chaotic features
The fractal dimension (FD) feature is a well known feature which measures the chaotic pattern. The FD is measured by computing the irregularity over multiple scales [57]. The FD feature is relevant since the plaques pattern has randomness and is multifocal. Thus, one can apply FD computation for stratication of plaque into high-risk and low-risk bins. For a given self-similar object of N parts scaled by a ratio r from the whole, its FD is given by
=
NrFD log
.
1
log
)
(
r
13.3.3.4 Classification using SVM
SVM is the most fundamental classier strategy for separating data points into different classes [58]. In our scenario, we utilize a two-class problem, i.e. high risk and low risk. The objective is to nd the best hyper-plane which straties the two classes with the largest margin. This margin denes the maximal width of the two slabs parallel to the hyper-plane having no interior data points. Appendix A shows the working of SVM in detail, where different kernels are being used such as: linear-, polynomial- and radial-basis functions [59, 60].
(13.1)
13.4 Experimental protocol
This section presents three sets of experimental protocols adapted for optimization of machine learning parameters, understanding the effect of dominant features on classication accuracy for risk assessment, and understanding the effect of a change in data size on machine learning performance.
13.4.1 Experiment 1: Kernel optimization during machine learning training phase
Experiment 1 is focused on choosing the best kernel for the SVM classier during the training and testing phases. Once the best kernel is selected, we use this kernel for the other two experimental protocols. For the selection of the best kernel, we run the machine learning system using ve sets of kernels: Linear, RBF, Poly-1, Poly-2 and Poly-3. The optimization is executed using a cross-validation protocol by taking a particular partition (K) and a particular LDT and studying the effect of classication accuracy for risk assessment with respect to a change in dominant features. The kernel which yields the highest classication accuracy over all the dominant features is selected.
13-11
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
13.4.2 Experiment 2: The effect of dominant features on classication accuracy
Experiment 2 is merely an extension of experiment 1, where the cross-validation protocol is implemented by taking the exhaustive combinations of all three partitions, all LDTs ranging from 5 mm to 8 mm (16 sets) and all dominant features (from D = 1toD = 16). During the automated set-up, the training component of the machine learning system utilizes the automated grayscale IMT wall region along with the automated LD measurements, which act as a gold standard (ground truth) for the training phase. Such a cross-validation paradigm is repeated for all three walls: the near wall, far wall and combined wall. These IMT wall regions correspond to experiments 2(a), 2(b) and 2(c), respectively.
13.4.3 Experiment 3: The effect of data size on machine learning performance
Experiment 3 evaluates the effect of an increase in data size on the classication accuracy for risk assessment using the machine learning paradigm. In this experiment, the image data sampling size is made to increase by 10% to study the effect on classification accuracy. It starts from a minimum number of patients (say 40, selected empirically) and is made to increase at a rate of 10%. This protocol is run for the optimal kernel, as obtained in experiment 1. Similar to experiment 2, experiment 3 is also conducted for the automated and manual protocols. This paradigm for increasing data size is repeated for all three walls, the near wall, far wall and combined wall, labeled as experiments 3(a), 3(b) and 3(c), respectively. Since the machine learning is characterized by a combination of population size, types of grayscale features, types of thresholds for ground truth, types of kernels in the classication paradigm during the training and testing phases, and the types of associated risks, we therefore adopted three different kinds of cross-validation protocols (K5, K10 and JK).
13.5 Results
13.5.1 Experiment 1Results: Kernel optimization during the machine learning
training phase
Using the cross-validation protocol for kernel optimization (experiment 1) on N = 407 images, the resultant plot can be seen gure 13.9 and the corresponding table can be seen in appendix B, table B3. The general tendency of the sRAS shows an increase in stratication accuracy with an increase in dominant features (D). The plot further shows that Poly-2 has the highest accuracy for each of the three partition protocols (K = 5, K = 10, K = JK) and thus is selected as the best kernel for our automated stratication based on characterization of tissue morphology. Note that in appendix B, table B3, the values in the table represent the mean values of all the dominant features within each batch of kernels.
13.5.2 Experiment 2Results: The effect of dominant features on classication
accuracy
This experiment shows the cross-validation accuracy by taking: (a) all three protocols (K = 5, K = 10, K = JK); (b) all LDT ranges from 5 mm to 8 mm
13-12
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 13.9. Plot of accuracy versus dominant features for xed data size (N) of 407 images using ve kernels: linear, RBF, Poly-1, Poly-2 and Poly-3.
(16 sets); (c) an increase in dominant features (from D = 1toD = 16); and (d) all three kinds of carotid wall,: far, near and combined. The three kinds of walls correspond to the three different combinations shown in gures 13.1013.12, respectively. Figure 13.10 corresponds to the far wall experiment, while gures
13.11 and 13.12 correspond to the near wall and combined wall experiments,
respectively. Each gure has 2 × 3 matrices consisting of six plots: two rows corresponding to the automated sRAS versus the mRAS and three columns corresponding to the three sets of protocols: K = 5, K = 10 and K = JK. In gure
13.10 (the far wall experiment), the columns for the automated sRAS are labeled as
(a1), (a2) and (a3) while the columns for the mRAS are labeled as (b1), (b2) and (b3). In gure 13.11 (the near wall experiment), the columns for the automated sRAS are labeled as (c1), (c2) and (c3) while the columns for the mRAS are labeled as (d1), (d2) and (d3). In gure 13.12 (the combined wall experiment), the columns for the automated sRAS are labeled as (e1), (e2) and (e3) while the columns for the mRAS are labeled as (f1), (f2) and (f3). The classication accuracy corresponding to nine automated sRASs is shown in appendix B, table B4. The percentage stratication accuracy for different feature combinations (D) and for different partition protocols (K = 5, K = 10 and K = JK) for T = 20 trials are computed using
T
auto
η
ldt
∑∑∑
auto
η
k
()
w
=
lLdDt
××
LDT
w
(, , )
(13.2)
13-13
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 13.10. Experiment 2(a): Accuracy versus dominant features for the automated sRAS (top row) versus the manual mRAS (bottom row) corresponding to the far wall using the three partition protocols: K = 5(first column), K = 10 (second column) and K = JK (third column).
Figure 13.11. Experiment 2(b): Accuracy versus dominant features for the automated sRAS (top row) versus the manual mRAS (bottom row) correponding to the near wall using the three partition protocols: K = 5(first column), K = 10 (second column) and K = JK (third column).
13-14
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 13.12. Experiment 2(c): Accuracy versus dominant features for the automated sRAS (top row) versus the manual mRAS (bottom row) correponding to the combined wall using the three partition protocols: K = 5 (rst column), K = 10 (second column) and K = JK (third column).
T
man
η
ldt
w
(, , )
,
(13.3)
where
w
auto
ldt(, , )
∑∑∑
=
lLdDt
××
LDT
man
η
k
()
w
is the accuracy using the automated method for wall type w,
lumen diameter l, dominant feature d and trial number t. Correspondingly,
man
w
ldt(, , )
is the accuracy using the manual method for wall type w, lumen diameter l, dominant feature d and trial number t. Using the above equations, the mean accuracies are calculated and are shown in appendix B, table B5 for automated LD and manual LD. All nine plots of gures 13.1013.12 demonstrate the increase in accuracy with the increase in the number of dominant features used for the training and testing of the machine learning system. There are 16 lumen diameter thresholds in each of the nine plots, corresponding to stenosis severity, which ranges from 5 mm to 8 mm in intervals of 0.2 mm. We observed that even when the LDT decreases from 8 mm (normal artery) to 5 mm (severe artery), the behavior of the machine learning system is consistent in the cross-validation protocol and generalizes as the number of dominant features increases. For each LDT selected, the order of dominant features changes, which is shown in appendix
B, table B6, taken as an example for the far wall at LD = 6.4 mm. Even though the
behavior of the near wall, far wall and combined wall is similar when dominant features are increased, the far wall accuracy curves are more clustered together compared to the near and combined walls. This behavior is attributed to accurate media wall detection in the far wall region compared to the near wall during the B-mode acquisition. As the number of dominant feature combinations increases beyond 6, the classication accuracy stabilizes to nearly 99%, converging to 100%
13-15
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
when D is 16. This behavior is consistent for all three walls, as can be seen in gures
13.10–13.12. The corresponding mean accuracies for automated (sRAS) versus
manual (mRAS) for the three different protocols (K = 5, K = 10 and K = JK) using N = 407 images is shown in appendix B, table B3, which is consistent with the plots in gures 13.1013.12, respectively.
13.5.3 Experiment 3Results: The effect of data size on machine learning
performance
This experiment shows the effect of an increase in data size on the classication accuracy of the machine learning system corresponding to the far wall, near wall and combined wall in gures 13.1313.15, respectively. Each gure has two plots corresponding to the automated sRAS (left) and manual mRAS (right). Each plot of the far wall has three protocols corresponding to three partitions, K = 5, K = 10
Figure 13.13. Experiment 3(a)the effect of increase in data size: automated (left) versus manual (right) in the far wall for the K = 5, K = 10 and K = JK protocols.
Figure 13.14. Experiment 3(b)the effect of increase in data size: automated (left) versus manual (right) in the near wall for the K = 5, K = 10 and K = JK protocols.
13-16
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 13.15. Experiment 3(c)the effect of increase in data size: automated (left) versus manual (right) in the combined wall for the K = 5, K = 10 and K = JK protocols.
and K = JK shown in red, green and blue, respectively. Thus, there are a total of six plots in this experiment shown as 3(a), 3(b) and 3(c), corresponding to the far, near and combined walls. For all six plots, as the data size increases during the machine learning protocol, the classication accuracy increases until it eventually reaches nearly 100%. When the data size is 40 images (empirically selected), the accuracy is close to 80%. As the number of patients in the training set gradually increases, the machine learning classication accuracy increases and stabilizes to the point of diminishing returns and beyond this point there is no further change in the accuracy. The point at which the curve starts to show no more change is the point where learning changes to generalization during the machine learning phase. Our pop­ulation shows that this occurs at around n = 200 for the automated system and n = 240 for the manual system. Such a behavior is seen in both the automated sRAS and manual mRAS. Appendix B, table B7 shows the increase in accuracy with the increase in the number of patients during the training phase. Note that this is the mean accuracy over all the LD thresholds.
13.6 Performance evaluation
13.6.1 Precision-of-merit (PoM) analysis
The performance or PoM of the machine learning system is evaluated by replacing the automated grayscale IMT wall strips region with a manual segmented grayscale IMT wall strip region. Similarly, we replaced the automated LD with manual LD during the training phase of the machine learning system. The protocols adapted during the validation of our machine learning system are exactly similar in nature to the automated and real set-up. This means there are three kinds of partition protocols: K = 5, K = 10 and K = JK. We replicate these protocols for all three sets of walls: the far wall, near wall and combined walls. Mathematically, the PoM is given as follows:
13-17
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
auto
PoM ( ) 100
=−
w
() ()
kk
ηη
w
⎜ ⎝
η
man
w
man
w
()
k
⎞ ⎟
100, (13.4)
×k
⎟ ⎠
where
auto
η ldt(, , )
w
oM ( )
is the PoM for the wall type w with partition protocol k,
k
w
is the accuracy using the automated method for wall type w, lumen diameter l, selected dominant feature d and trial number t, and is computed using equation (13.2). Correspondingly,
man
w
ldt(, , )
is the accuracy using the manual method for wall type w, lumen diameter l, dominant feature d and trial number t and is computed using equation (13.3).
Figure 13.16 (g1), (g2) and (g3) (the left column) shows that the probability of the PoM values for the far wall is greater than that of the near wall for given accuracy values within 5% tolerance (ε) for three kinds of protocols: K = 5, K = 10 and K = JK. Similarly, gure 13.16 (h1), (h2) and (h3) (the right column) shows that the probability of the PoM values for the far wall is greater than of the combined wall within 5% tolerance (ε) for given accuracy values for the three kinds of protocol: K = 5, K = 10 and K = JK. Note that our stratification of probability curves utilizes three bins: mild (or low), moderate (or medium) risk and severe (or high) risk, and their computed LD ranges were: mild (low) risk—7.4 mm to 8.0 mm; moderate (or medium) risk6.4 mm to 7.2 mm; and severe (or high) risk5.0mmto6.2mm.Further,itisclearfromgure
13.16 that a PoM of 95% is achieved when D is greater than or equal to ten dominant
features. We therefore selected D = 10 as a threshold for seeing the overall PoMs.
Appendix B, table B8 shows the PoM values for the far, near and combined walls for three partition protocols: K = 5, K = 10 and K = JK, and total trials (T ) = 20 for three different dominant feature (D) conditions. The mean PoM for all the walls and all three protocols are computed on the basis of D < 10 or D > 10 or all D included. As can be seen, when
D > 10 the PoM increases compared to when D < 10, which is
intuitive and consistent with the behavior.
13.6.2 ROC analysis
After choosing the optimized kernel, i.e. a polynomial of order 2, we computed the ROC curve for all three partition protocols, i.e. K = 5, K = 10 and K = JK, and for the automated sRAS and manual mRAS using the far wall, near wall and combined wall. The results in appendix B, table B9 show the mean sensitivity, mean specicity and AUC for the far, near and combined walls for the K = 5, K = 10 and K = JK partition for 16 values of LD ranging from 5.0 mm to 8.0 mm in the interval of
0.2 mm. The mean AUC value for the whole wall is close to unity which indicates the high accuracy of our system.
13.7 Discussion
13.7.1 Our system
In this study, we present an automated sRAS using morphology-based tissue characterization of plaque build-up in the carotid arterial wall. The sRAS is a
13-18
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 13.16. Left (g1, g2 and g3): probability (Pr) curves for the PoM when far wall > near wall, given the classication accuracy for the three different protocols: K = 5, K = 10 and K = JK, with ε = 5%. Right (h1, h2 and h3): probability curves for the PoM when far wall > combined wall given the classication accuracy for the three different protocols: K = 5, K = 10 and K = JK, with ε = 5%.
machine learning system where the learning phase involves generating the learning coefcients, which are derived using grayscale dominant features and a risk label based on stenosis severity of the artery. Three sets of partition protocols were adopted and online stratication accuracy was measured using a cross-validation paradigm. This scheme was applied to the three wall types in the carotid artery: the near, far and combined walls. Even though the far wall proved to have superior performance, the near wall also performed well (see appendix B , table B6). This proves our hypothesis that the near wall carries an equal risk of atherosclerotic disease and should be taken into consideration for stroke risk assessment. We, however, think more analysis needs to be performed on this aspect in the future.
13-19