Tissue Characterization 2421 Methodology

We begin our process of tissue characterization taking the IVUS image and transforming it to cartesian coordinates (Fig. 2.20(a)). Once the cartesian

Figure 2.19: Plaque segmentation (see text).

transformation is done, artifacts are removed from the image (Fig. 2.20(b)). There are three main artifacts in an IVUS image: the transducer wrapping, which creates a first halo at the center of the image (in the cartesian image the echo is shown at the top of the image); the guide-wire effect, which produces an echo reverberation near the transducer wrapping; and the calibration grid, which are markers at a fixed location that allow the physicians to evaluate quantitatively the morphology and the lesions in the vessel. With the artifacts removed, we proceed to identify intima and adventitia using the process described in the former section. At this point, we have the plaque located and we are concerned with tissue identification (Fig. 2.20(c)). The tissue classification process is divided

Figure 2.20: Tissue characterization diagram.

in three stages: First, the soft-hard classification (Figs. 2.20(d) and 2.20(e)), in which the soft plaque, the hard plaque, and calcium are separated. In the second stage, the calcium is separated from the hard plaque (Figs. 2.20(f) and 2.20(g)). At the last stage, the information is fused and the characterization is completed. We will refer later to this diagram to explain some parts of the process. Recall that the plaque is the area comprised between the intima and the adventitia. With both borders located we can focus on the tissue of that area.

For such task, the three stages scheme formerly described is used. Regarding the first stage of the process, a classification is performed on the feature space. At this point, a feature space and a classifier must be selected. To help to choose which feature space and which classifier to use, we try each of the feature spaces with a general purpose classifier, the k-nearest neighbors method used as a ground truth classifier. Regardless the classifier used, the information provided at the output of the system is a pixel classification. Using these data we can further process the classification result incorporating region information from the classification process and obtain clear and smoother borders of the soft and the mixed plaques. Different processes can be applied to achieve this goal, two possible approaches are region-based area filtering and classification by density filtering. In a region-based area filtering the less significant regions in terms of size are removed from the classification. On the other hand, the other method relies on keeping the regions that have high density of classification responses. As the classification exclusively aims to distinguish between soft and hard plaque, a separate process is added to separate hard plaque from calcium.

Once soft and hard plaque are distinguished, we proceed to identify what part of the hard plaque corresponds to calcium. One can argue why not to include a third class in the previous classifier. The reason we prefer not to do so is because experts' identification of calcium plaques is performed by context. Experts use the shadowing produced by the absorbtion of the echoes, behind a high-echoreflective area, to label a certain area as calcium. In the same way, we take the same approach. On the other hand, the fact of including a third class only hinders the decision process and increases the classifier complexity. Therefore, the calcium identification process is made by finding the shadowing areas behind hard plaque. Those areas are easily identified because the soft-hard classification also provides this information (Fig. 2.20) since shadowing areas are classified as nontissue. We can see a plausible way of finding calcified areas. Figure 2.20(f) shows the classification result under the adventitia border of the "hard" tissue. Dark gray level areas are regions with soft plaque and, therefore, do not provide information of the calcium composition of the plaque. We use one of the previous classified images, the soft-hard classification or the blood-plaque classification. In white, it is displayed the regions of tissue under the adventitia border in the area of interest. Figure 2.20(g) shows in light gray the areas of shadowing, and therefore, the areas labelled as calcium.

To end the process, the last stage is devoted to recast the resulting classification to its original polar domain by means of a simple coordinate transformation.

2.4.2.2 Experimental Results

To evaluate the results, a classification of over 200 full-tissue regions from 20 different patients has been performed. The training set is a subset of two thirds of the overall data determined using the bootstrapping strategy. The rest of the data has been used as test set. Previously, different physicians have determined and delineated plaque regions in each full-tissue image.

The first experiment is to set a ground truth for the feature spaces, as a measure to evaluate their description power. We have used k-nearest neighbors as a ground truth classifier. To choose the number of neighbors, we select a feature space and evaluate the performance for different values of k. Tables 2.2, 2.3 and 2.5 show the error rates for pixel classification (RAW Error) and postprocessed classification taking into account neighboring information and density of classifier cluster responses (Post Error). These tables also show the percentages of false positives (FP) and false negatives (FN) for both errors. The FP and FN are included as they give information of the possible geometry of samples in the feature space.

Table 2.2 illustrates the results regarding the selection of the number of neighbors k. It can be seen that for k = 7 a lower pixel error rate is obtained. Therefore, the performance of the feature spaces will be evaluated using 7-nearest neighbors. The result of the classification of the test data using all feature spaces and 7-nearest neighbors classifier is shown in Table 2.3. Observing the RAW data error rate, the best overall feature spaces are the co-occurrence matrices, local binary patterns, derivatives of Gaussian, and accumulation local moments. These results are confirmed looking at the postprocessing error rate and ratifies

Table 2.2: Selection of the parameter k, using local binary pattern feature space as a reference

Table 2.2: Selection of the parameter k, using local binary pattern feature space as a reference

k value

RAW error

FP

FN

Post error

FP

FN

3

33.94

25.13

8.80

10.16

3.46

6.69

7

25.67

9.67

16.23

3.45

2.67

0.81

15

32.93

26.19

6.74

5.81

3.46

2.34

Table 2.3: Feature space performance discriminating hard plaque from soft plaque using k-nearest neighbor

Table 2.3: Feature space performance discriminating hard plaque from soft plaque using k-nearest neighbor

Feature space

RAW error

FP

FN

Post error

FP

FN

Co-occurrence measures

22.36

10.91

11.45

10.88

4.19

6.68

Derivative of Gaussian

27.81

23.51

4.95

16.29

16.67

0.04

Gabor filters

35.26

18.86

17.22

16.26

16.49

0.07

Wavelets

45.05

20.52

24.90

31.78

24.40

7.68

Accumulation local moments

31.72

16.42

15.30

12.17

11.36

0.81

Local binary patterns

25.67

9.67

16.23

3.45

2.67

0.81

the qualitative evaluation shown in Table 2.4, where we observe that the same feature spaces are the ones that perform best. Analyzing each of the feature spaces in terms of FP and FN rates, we can deduce that Co-occurrence feature space has good discrimination power, having a "symmetric" nature where both FP and FN rates are comparable. In the same sense, we can deduce that the overlapping of both classes is similar. Derivatives of Gaussian's filter space have tendency to over-classify hard plaque. The classes in the feature space are not very well defined as hard plaque must have a higher scatter than the soft plaque. Gabor filter's bank gives a good description of both classes as they have similar false rates. However, both classes are very overlapped giving a hard time to the classification process. Wavelets overlapping of classes in the feature space is extremely high; therefore, it describes bad each of the classes. Accumulation local moments have similar description power than Gabor filter's bank; however, the different responses from both allow a much better postprocessing in accumulation local moments. This fact allows us to suppose that the classification error points in the image domain are much more scattered and

Table 2.4: Descriptive table of the discriminative power of each feature space using k-nearest neighbors

Table 2.4: Descriptive table of the discriminative power of each feature space using k-nearest neighbors

Feature space

Qualitative speed

Qualitative performance

Co-occurrence measures

Slow

Good

Gabor space

Slow

Acceptable

Wavelets

Fast

Poor

Derivative of Gaussian

Slow

Acceptable

Accumulation local moments

Fast

Good

Local binary patterns

Fast

Good

(g) (h) Figure 2.21: Tissue pixel classification data using 7-nearest neighbors method on different feature spaces. (a) Original image in cartesian coordinates. (b) Expert manual classification of tissue. (c) Co-occurrence feature space. (d) Gabor feature space. (e) Wavelets feature space. (f) Derivative of Gaussian feature space. (g) Accumulation local moments feature space. (h) Local binary patterns feature space.

with very few local density. Local binary patterns have good descriptive power as well as giving a more sparse pattern in FP and FN in the image domain. Figure 2.21 provides a graphical example of the performance of 7-nearest neighbors method applied to several feature spaces. Observing the images, we realize that scale-space processes, derivative of Gaussian, Gabor filters, and wavelets have poor to acceptable discrimination power, and therefore, are not suitable for the task of tissue discrimination. On the other hand, statistic-based feature spaces and structure feature spaces have acceptable to good performances. Table 2.4 details the conclusions arisen from the experiment. The qualitative speed nomenclature (fast/slow) indicates the viability of the feature space technique to be included in a real time or near-real time process. A "fast" scheme denotes a method over 10 times faster than the "slow" one. Because the results are obtained using prototypes and not a full application, no absolute time measure is provided. Note, also, that the images displayed are pixel-based classification results and have no further smoothing postprocessing. To further develop our

Table 2.5: Feature space performance using FLD and Mahalanobis distance

Feature space

RAW error

FP

FN

Post error

FP

FN

Co-occurrence measures

40.88

34.66

6.20

12.91

12.10

0.81

Accumulation local moments

35.50

20.34

16.16

13.83

13.02

0.81

Local binary patterns

26.37

5.76

20.85

6.93

1.52

5.47

discussion we will only take the three best postprocessed data performing feature spaces: co-occurrence matrix measures, accumulation local moments, and local binary patterns. Up to this point we have neither taken into account complexity of the methods nor time issues. However, these are critical parameters in real applications, thus, we consider them in our following discussions.

Once the feature space is selected, the next decision is to find the most suitable classifier taking into account our problem constraints, if any. We are concerned with speed issues, therefore, simple but powerful classifiers are required. Because the high dimensionality of two of the feature spaces selected (co-occurrence matrix measures have about 24 features per distance and accumulation local moments have 81 features) a dimensionality reduction step is desired. PCA is the first obvious choice, but because great amount of overlapping data the results are worse than using Fisher's linear discriminant analysis which is focalized in finding the most discriminative axes for our given set of data. The result of this experiment is shown in Table 2.5. We use maximum likelihood combined with a Fisher linear discriminant analysis reduction. As local binary patterns do not need dimensionality reduction due to the small amount of features computed (three features), the comparison with this method is done by just classifying with the ML method. As expected, the raw results are much worse with this kind of classifier. Co-occurrence matrix measures take the worse part doubling their error rate. However, local binary patterns, though they have also worse error rate with ML, manage to be the most discriminative of the three methods. This fact is also shown in the postprocessing, where local binary patterns still have the lower error ratio. Co-occurrence matrix measures regain their discrimination power after the postprocessing.

Therefore, using one of the fastest classifiers, ML, one achieves, at least, a classification ratio over 87% (with accumulation local moments). If the selected feature space is local binary patterns, the scheme is the fastest possible scheme as local binary patterns are computationally efficient and low-time consuming as

Figure 2.22: Boosting procedure for tissue characterization at different stages of its progress. (a) Expected hand classification by an expert. (b) First stage of the boosting procedure. (c) Classification with a five classifiers ensemble. (d) Classification with 10 "weaks" ensemble.

Figure 2.22: Boosting procedure for tissue characterization at different stages of its progress. (a) Expected hand classification by an expert. (b) First stage of the boosting procedure. (c) Classification with a five classifiers ensemble. (d) Classification with 10 "weaks" ensemble.

well as the ML classifier does not transform data in another feature space. This scheme is really well suited for real-time or near-real-time applications because of both time efficiency and reliability in the classification. This is, however, by no means the only near-real-time configuration available since accumulation local moments are computationally as fast as local binary patterns. However, the FLD dimensionality reduction hinders the process due to the complexity of the data in its original feature space. To overcome this problem, other classifiers can be used. The necessity to find reliable and fast classifiers lead us to boosting techniques. Boosting techniques allow a fast and simple classification procedure to improve its performance as well as maintaining part of its speed. To illustrate this fact Fig. 2.22 shows the evolution of the classification when more classifiers are added to the strong classifier. Figure 2.22(a) shows the expected hand classification by a physician. Figure 2.22(b) shows the base classification of a single "weak". Figure 2.22(c) illustrates the result of the classification using an ensemble of five classifiers. Figure 2.22(d) shows the resulting classification after the addition of 10 weak classifiers to the ensemble. The error rates at different stages of the process are also shown in Table 2.6. These results are computed using a ML method as a weak classifier on the accumulation local moments space. The numbers show how the error rate is improved, and, though the raw classification error rate is nearly immutable, we can observe that there is a great change in the classification data points distribution in the image domain since the FP and FN rates drastically change. The postprocessing error rate gives better description of what is happening. The disposition of the error points in the classification image is more sparse and unrelated to their neighborhood, allowing better

Table 2.6: Error rates using boosting methods with maximum likelihood with the accumulation local moments space

Ensemble no.

RAW error

FP

FN

Post error

FP

FN

Base error

34.86

28.20

6.98

41.94

40.33

1.10

Ensemble of 5 c.

29.38

16.32

13.32

33.17

31.87

1.10

Ensemble of 10 c.

31.44

7.36

23.37

7.92

3.22

4.76

postprocessing and classification rates. In this case, the classification rate is over 92% with a classifier as fast as applying 10 times a threshold. Therefore, using accumulation local moments and boosting techniques we have another fast and highly accurate scheme for real-time or near-real-time tissue characterization.

Up to this point, we have discussed the reliability of the soft plaque versus hard plaque discrimination process, which is our main concern, since the identification of calcium is reduced to the part of hard plaque with a large shadowing area. Using the method described in the former section, 99% of the calcium plaque is correctly identified. Figure 2.23 shows some results of the tissue characterization process. Figures 2.23(a) and 2.23(b) show the characterization of a soft plaque. In Figs. 2.23(c) and 2.23(d), there are two different kind of plaques detected, calcium (gray region) and soft plaque (white region). Figures 2.23(e) and 2.23(f) show the characterization of the three kind of plaques: fibrotic (light gray region), soft plaque (white region), and calcium (dark gray region).

0 0

Post a comment