**5.3 AR-DLac scheme**

198 New Advances in the Basic and Clinical Gastroenterology

Previous studies (Charisis *et al.*, 2010a, 2011) have shown that RGB is the most efficient space for WCE image analysis (compared to other colour spaces, i.e., HSV and CIE Lab). More specifically, the majority of ulcer texture information resides in green component of the RGB space. This conclusion coincides with the yellow-greenish appearance of ulcer regions. Thus, the proposed AR-DLac scheme is applied on the green channel extracted from each image.

The next step of our approach includes image purification by applying a denoising procedure. In order to facilitate texture-pattern extraction, the images need to be refined and smoothed by eliminating any distorting information. Endoscopic images from a WCE examination are prone to misleading content. Hardware limitations (quality of the image sensor and lens, non-adjustable light source) and adverse filming conditions (non-uniform lighting, reflections of the light on intestinal juices and lens cover, peptic content) are likely to cause high levels of noise to reside in total or part of the image (e.g., underexposed). To address this issue, we apply BEEMD analysis and each image is decomposed in eight 2D IMFs and residue. The first two IMFs contain the high frequency components of the image, i.e., the noise that may exist. Therefore, they are discarded and not utilized in the subsequent analysis and for image reconstruction. Figure 8 presents a worst case scenario, where artificial high level noise was added to an ulcer image. The distorted image is decomposed with BEEMD into eight IMFs and residue. The result (reconstructed image) proves that BEEMD is capable of dealing successfully with extreme cases of noise. Sheer

Fig. 8. Ulcer image with high-level, artificial noise decomposed in 8 IMFs plus residue and

**5.2 Image denoising** 

purified by BEEMD analysis.

In order to follow the WCE image characteristics and focus upon the ones that mostly relate to ulcer, an AR approach was developed. The aforementioned capabilities of EMD were exploited by developing a new DLac-based approach for the optimized selection of IMFs that correspond to ulcer characteristics of a WCE image. Towards this direction, DLac analysis is applied to every IMF of a decomposed image. The selected IMFs are used either to reconstruct a new image (R-case), or provide separate images (NR-case) that represent specific modes of oscillations coexisting in the initial WCE image. Apart from the optimal IMF selection, we are able to investigate how ulcer texture information are distributed across the frequency scales of WCE images.

#### **5.3.1 Proposed DLac analysis**

The great advantage of DLac is the ability to perform texture analysis in various scales. The coarseness of the scale is primarily determined by the size of window *w*, which designates the size of the neighbourhood for box mass calculation; the greater the window, the coarser the analysis scale. In the case of ulcer tissue recognition, a multi-scale texture analysis is required considering the great variability in size and appearance of ulcer regions. In this context, DLac is calculated for a variety of window sizes, given a constant, relative small box size *r*, in order to achieve pattern analysis at different scales, while identifying slight variations in neighbouring pixels (due to small value of *r*). An example of DLac-w (*r*=3, *w*=4-30) curves that correspond to images b, d, e and f from Fig. 1 is given in Fig. 9a. The curves are distinct, however, obvious discrimination is not achieved. To deliver greater differentiation between the curves an identical reference level has to be secured (Hadjileontiadis, 2009). Thus, DLac-w curves are normalized to the DLac value that corresponds to the smallest *w*. The resulting curves (Fig. 9b) provide quite clear discrimination between the four patterns. From now on, any reference to DLac-w curves implies normalized curves.

#### **5.3.2 Optimized IMF selection**

The selection of optimum IMFs is based on the characteristics of DLac-w curve of each IMF. The motive for such an approach lies in the concept that IMFs with possible useful texture

Enhanced Ulcer Recognition from Capsule Endoscopic Images Using Texture Analysis 201

function L(w). The normalized DLac-w curves in Fig. 9b expose an attitude that bears

was chosen to model the normalized DLac-w curves (Hadjileontiadis, 2009). Parameter *a* represents the concavity of hyperbola, *b* portrays the convergence of L(w), and *c* is the translational term. The best interpretation of DLac-w by the model L(w) is computed as the solution of a least squares problem where parameters a, b, c are the independent variables. Parameters a, b, c embody the global behaviour of the DLac curve, i.e., all the substantial information. However, preliminary studies have shown that local behaviour of DLac curve is also important in ulcer tissue classification. Consequently, FV consists of parameters a, b, c plus five DLac values that correspond to the five smallest values of *w* (i.e., L(6) - L(10); L(5) is discarded as it always equals to one due to normalization). The values L(6) – L(10) were selected empirically after exhausting experiments. In this manner, FV embodies both local and global trend of DLac curve achieving a significant size reduction (eight rather than 26

More specifically, in R-case scenario the selected IMFs and the residue recompose a new image on which the DLac-w curve is calculated and FV is extracted (FVR) as described above. In NR-case scenario, the final FV (FVNR) contains the FV from each selected IMF (FVi, i=3 to 8). However, different number of IMFs is selected for every image with the subsequent problem of changing FV size. To address this issue, FV is designed to include the FVs from IMFs 5, 6 and 8 (i.e., FVNR = {FV5, FV6, FV8}) which are the most frequently chosen IMFs (Fig. 10) for all images in the dataset, implying that they contain the majority of beneficial information. Last but not least, the ability of texture information of each individual IMF to discriminate ulcer from normal images is investigated. In this case, the final FV (FVindiv) consists of the FV from an individual IMF, different each time, i.e. FVindiv = FVi for i=1 to 8. IMFs 1 and 2 are included, despite their noise contamination, for an overall

The last step of ulcer detection scheme is the classification procedure. This procedure involves the classification of extracted FVs into healthy/ulcerous and is accomplished by algorithms called "classifiers". The target of a classifier is to identify the population (e.g., healthy or abnormal) to which a FV belongs on the basis of a training set of FVs whose population is already known. It is usual that 90% of the sample dataset is used for training the classifier and 10% for testing. This procedure is repeated ten times (10-fold cross validation) with random training and test sets in order to acquire more accurate results. The classification performance is measured with the aid of accuracy (acc.), sensitivity (sens.) and specificity (spec.) indexes. The average acc., sens. and spec. are obtained in case of 10-fold cross validation technique. Various classification algorithms have been proposed in the literature. For an extensive evaluation of the classification performance of the proposed AR-DLac scheme, the widely used classification algorithms, i.e., LDA (Linear Discriminant Analysis), QDA (Quadratic Discriminant Analysis), MD (Mahalanobis Distance) (Krzanowski, 2000), and SVM (Support Vector Machine) (Cristiani & Shawe-Taylor, 2000),

wb +c, w=[5-30], (5)

resemblance to the one of hyperbola. On this ground, the function

values).

assessment.

**5.4 Classification** 

were adopted.

L(w) =

a

Fig. 9. DLac (a) and normalized DLac (b) curves for various window sizes (4 to 30 pixels) for images b, d, e and f from Fig. 1.

information should provide DLac-w curves that bear resemblance to the curves estimated on ulcer images. In this context, it has been observed that the slope of DLac-w curves of ulcer images lies within specific limits for the majority of ulcer cases. This observation led us to a slope-based criterion for IMF selection. The IMFs that provide DLac curves with slopes within the limits, specified by DLac curves of ulcer images, were selected. Figure 10 shows the selection probability of each IMF. According to the diagram, IMF5, IMF6 and IMF8 were selected for the majority of images implying that the information included in these IMFs should be taken under consideration.

Fig. 10. Selection probability for each IMF.

#### **5.3.3 DLac-based feature vector**

In image pattern detection applications, feature vector (FV) is the set of features that characterize and represent an image and are utilized for the discrimination between various patterns. The DLac-based FV extraction was implemented in two different approaches, following the R-case and NR-case scenario. From a general perspective, the previously described DLac-w curve is used to form the FV. In our approach DLac-w curve is calculated for *w*=5-30 pixels. These 26 DLac values consist an extensive FV, subject to the "curse of dimensionality". The essence of reducing the feature space dimension without omitting any crucial information introduces the concept of modelling DLac-w curves with another function L(w). The normalized DLac-w curves in Fig. 9b expose an attitude that bears resemblance to the one of hyperbola. On this ground, the function

$$\text{L(w)} = \frac{\text{a}}{\text{w}^{\text{b}}} + \text{c.} \quad \text{w} = [5 \text{-} 30]. \tag{5}$$

was chosen to model the normalized DLac-w curves (Hadjileontiadis, 2009). Parameter *a* represents the concavity of hyperbola, *b* portrays the convergence of L(w), and *c* is the translational term. The best interpretation of DLac-w by the model L(w) is computed as the solution of a least squares problem where parameters a, b, c are the independent variables. Parameters a, b, c embody the global behaviour of the DLac curve, i.e., all the substantial information. However, preliminary studies have shown that local behaviour of DLac curve is also important in ulcer tissue classification. Consequently, FV consists of parameters a, b, c plus five DLac values that correspond to the five smallest values of *w* (i.e., L(6) - L(10); L(5) is discarded as it always equals to one due to normalization). The values L(6) – L(10) were selected empirically after exhausting experiments. In this manner, FV embodies both local and global trend of DLac curve achieving a significant size reduction (eight rather than 26 values).

More specifically, in R-case scenario the selected IMFs and the residue recompose a new image on which the DLac-w curve is calculated and FV is extracted (FVR) as described above. In NR-case scenario, the final FV (FVNR) contains the FV from each selected IMF (FVi, i=3 to 8). However, different number of IMFs is selected for every image with the subsequent problem of changing FV size. To address this issue, FV is designed to include the FVs from IMFs 5, 6 and 8 (i.e., FVNR = {FV5, FV6, FV8}) which are the most frequently chosen IMFs (Fig. 10) for all images in the dataset, implying that they contain the majority of beneficial information. Last but not least, the ability of texture information of each individual IMF to discriminate ulcer from normal images is investigated. In this case, the final FV (FVindiv) consists of the FV from an individual IMF, different each time, i.e. FVindiv = FVi for i=1 to 8. IMFs 1 and 2 are included, despite their noise contamination, for an overall assessment.

#### **5.4 Classification**

200 New Advances in the Basic and Clinical Gastroenterology

Fig. 9. DLac (a) and normalized DLac (b) curves for various window sizes (4 to 30 pixels) for

information should provide DLac-w curves that bear resemblance to the curves estimated on ulcer images. In this context, it has been observed that the slope of DLac-w curves of ulcer images lies within specific limits for the majority of ulcer cases. This observation led us to a slope-based criterion for IMF selection. The IMFs that provide DLac curves with slopes within the limits, specified by DLac curves of ulcer images, were selected. Figure 10 shows the selection probability of each IMF. According to the diagram, IMF5, IMF6 and IMF8 were selected for the majority of images implying that the information included in these IMFs

In image pattern detection applications, feature vector (FV) is the set of features that characterize and represent an image and are utilized for the discrimination between various patterns. The DLac-based FV extraction was implemented in two different approaches, following the R-case and NR-case scenario. From a general perspective, the previously described DLac-w curve is used to form the FV. In our approach DLac-w curve is calculated for *w*=5-30 pixels. These 26 DLac values consist an extensive FV, subject to the "curse of dimensionality". The essence of reducing the feature space dimension without omitting any crucial information introduces the concept of modelling DLac-w curves with another

(a) (b)

images b, d, e and f from Fig. 1.

should be taken under consideration.

Fig. 10. Selection probability for each IMF.

**5.3.3 DLac-based feature vector** 

The last step of ulcer detection scheme is the classification procedure. This procedure involves the classification of extracted FVs into healthy/ulcerous and is accomplished by algorithms called "classifiers". The target of a classifier is to identify the population (e.g., healthy or abnormal) to which a FV belongs on the basis of a training set of FVs whose population is already known. It is usual that 90% of the sample dataset is used for training the classifier and 10% for testing. This procedure is repeated ten times (10-fold cross validation) with random training and test sets in order to acquire more accurate results. The classification performance is measured with the aid of accuracy (acc.), sensitivity (sens.) and specificity (spec.) indexes. The average acc., sens. and spec. are obtained in case of 10-fold cross validation technique. Various classification algorithms have been proposed in the literature. For an extensive evaluation of the classification performance of the proposed AR-DLac scheme, the widely used classification algorithms, i.e., LDA (Linear Discriminant Analysis), QDA (Quadratic Discriminant Analysis), MD (Mahalanobis Distance) (Krzanowski, 2000), and SVM (Support Vector Machine) (Cristiani & Shawe-Taylor, 2000), were adopted.

Enhanced Ulcer Recognition from Capsule Endoscopic Images Using Texture Analysis 203

from 51.9% to 61.3%, while the one of IMF2 varies from 53.3% to 62.0%. The sensitivity index is even lower (up to 29.1 percentage points) for the majority of classifiers (i.e., LDA, QDA and SVM). This performance implies that the texture information that lies in IMFs 1 and 2 is not eligible for ulcer detection. This behaviour is consistent with the concept of noise "contamination" of IMF1-2 and validates the noise reduction procedure described in §5.2 and illustrated in Fig 8. IMFs 1-2 contain the high frequency components of the image (i.e., the noise) and, therefore, should be discarded. On the contrary, the classification accuracy of IMF3-8 is 24.8% to 35.5% improved. IMF3 and IMF5 exhibit the lowest (77.4%) and highest (84%) performance (in terms of accuracy), respectively. Despite the superior classification rates, the performance of individual IMFs indicates that texture information that resides in a single image component is inadequate for efficient ulcer detection. Additionally, low classification sensitivity (<76%) suggests extensive misidentification of ulcer regions as healthy. It should be highlighted that IMFs 5, 6 and 8, that are the most commonly selected IMFs (Fig. 10) in IMF selection procedure, deliver the three highest classification accuracy rates among the IMFs. The convergence of these results testifies the

As far as the classification algorithms are concerned, the results in Table 1 imply that the most efficient classifiers include SVM and QDA. SVM achieves the best performance for the majority of IMFs (1-4, 8) due to its more advanced nature. However, the capabilities of QDA should not be underestimated since it exhibits 4.6, 5.7 and 4.8 percentage points higher classification accuracy than SVM for IMF5-7. LDA also proves competent, delivering slightly inferior performance. At last, MD is the most inappropriate classifier for our approach. The extremely high classification sensitivity (up to 99.2% for IMF6) in conjunction with the extremely low classification specificity (down to 15.6% for IMF6) denote over fitting to ulcer

> **IMF**  1 2 3 4 5 6 7 8.

Acc. 55.1±1.7 58.7±1.7 76.9±0.9 78.9±0.7 81.6±0.6 **80.0±1.1** 78.7±1.0 82.3±0.8 Sens. 35.9±2.4 51.8±2.4 71.1±1.4 71.3±1.4 70.4±0.9 **72.2±1.7** 73.4±1.5 82.1±1.4 Spec. 74.4±2.1 65.5±2.4 82.8±1.4 86.5±0.7 92.7±1.1 **87.7±1.6** 83.9±1.4 82.5±0.7

Acc. 59.5±0.8 59.9±1.4 77.1±0.9 79.7±0.7 **84.0±0.9** 78.2±0.8 **79.5±1.3** 80.5±0.9 Sens. 30.4±1.4 42.3±2.2 66.2±1.4 64.9±1.1 **72.0±1.3** 67.1±1.6 **74.0±1.9** 65.7±1.5 Spec. 88.6±1.1 77.5±1.7 88.1±1.5 94.4±0.9 **96.0±1.2** 89.2±0.6 **85.1±1.9** 95.3±1.1

Acc. 51.9±0.7 53.3±1.4 61.2±1.1 66.3±1.0 61.3±1.0 57.4±1.4 62.4±1.1 68.5±1.1 Sens. 98.3±0.8 96.6±1.2 93.8±1.0 93.8±1.0 97.1±0.9 99.2±0.9 98.3±0.9 96.5±1.0 Spec. 5.6±1.3 10.0±2.7 28.5±1.9 34.6±1.9 25.5±1.8 15.6±2.5 26.5±2.1 40.5±1.9

Acc. **61.3±1.6 62.0±1.5 77.4±1.1 79.8±1.0** 79.4±0.8 72.5±1.1 72.7±1.1 **82.5±0.5**  Sens. **42.9±2.0 59.8±1.8 74.2±1.2 75.9±1.3** 74.3±1.0 64.1±1.6 62.4±1.3 **67.7±0.9**  Spec. **79.8±2.2 64.2±2.6 80.5±1.6 83.5±1.6** 84.5±1.3 80.9±1.3 83.1±1.7 **93.3±1.0** 

Table 1. Mean classification accuracy, sensitivity and specificity (%) for each individual IMF

(FVi, i=1 to 8), for all classifiers (LDA, QDA, MD, SVM).

optimal IMF selection procedure.

texture information.

**Classifier** 

LDA

QDA

MD

SVM
