**2. Follicle detection**

The methods employed by [Hiremath and Tegnoor] for follicle detection based on different segmentation, feature extraction and classification techniques are described below:

**a.** Edge based method

The automatic method for the detection of follicles in ultrasound images of ovaries using edge based method uses, Gaussian lowpass filter for preprocessing, canny operator for edge detection for segmentation, 3σ-intervals around the mean for the purpose of classification [21]. The experimentation has been done using sample ultrasound images of ovaries and the results are compared with the inferences drawn by medical expert. Initially, the image is processed with Gaussian low pass filter yielding denoised image. Then canny operator is applied to detect the edges from the denoised image. Morphological dilation is performed by using the disk shaped structuring element with radius 1 to fill the weak edges. which yields the segmented image. Possible holes inside the segmented regions are filled. Any spurious regions due to noise are eliminated by morphological erosion. The regions in segmented image having smaller area than the threshold T (empirical value) are removed. The segmented regions are labeled. Set all the nonzero pixels of the border of segmented image to zero, which yields the final segmented image. Thus the plausible follicle regions are the labeled segments.

The geometric features are extracted for the known follicle regions. The main aim of geometric feature extraction is to recognize geometric properties of ovarian follicles in ultrasound images. The ovarian follicles are oval shaped compact structures, which resemble the circular/ellipse and are, thus, characterized by the seven geometric features, namely, the area A, the ratio R of majoraxislength to minoraxislength, the compactness Cp, the circularity Cr, the tortousity Tr,

the extent E and the centriod Y= <sup>1</sup> <sup>B</sup><sup>2</sup> <sup>∑</sup> i,j=1 B CS-i,-j(F-1(λ(F(CSi,j(X))))) .

– Watershed transform

172 Advancements and Breakthroughs in Ultrasound Imaging

– Scanline thresholding

– Active contour method

– 3σ interval based classifier.

– Linear discriminant classifier

For the purpose of experimentation, the databases D1, D2 and D3, of ultrasound images of ovaries are prepared in consultation with the medical expert, namely, Radiologist and Gyneocologist, for the present study of the follicle detection and ovarian classification in ovarian images. Some of the images are captured by the Toshiba [Model SSA-320A/325A] diagnostic ultrasound system with the transvaginal transducer frequency 26 Hz. Some of the images are obtained from the publicly available websites [www.radiologyinfo.com; www.ovaryresearch.com]. The image dataset D1 consists of the 80 ultrasound ovarian images, with the size 256x256. The image dataset D2 consists of the 90 ultrasound ovarian images, with the size 512x512. The image dataset D3 consists of the 70 ultrasound ovarian images, with the size 512x512. It contains the images of 30 normal (healthy) ovaries, 20 polycystic ovaries and

The methods employed by [Hiremath and Tegnoor] for follicle detection based on different

The automatic method for the detection of follicles in ultrasound images of ovaries using edge based method uses, Gaussian lowpass filter for preprocessing, canny operator for edge detection for segmentation, 3σ-intervals around the mean for the purpose of classification [21]. The experimentation has been done using sample ultrasound images of ovaries and the results

segmentation, feature extraction and classification techniques are described below:

– Geometric features

– Texture features

– K-NN classifier

– Fuzzy classifier

– SVM classifier

**1.3. Image data set**

20 cystic ovaries.

**2. Follicle detection**

**a.** Edge based method

**3.** Feature extraction

**4.** Classification

The classification stage comprises two steps, namely, the training phase and the testing phase.

*Training phase* : In the training phase, the geometrical features, are computed for regions known to be follicles in the training images in consultation with the medical expert. Then, the sample means and standard deviations of the geometric parameters R, Cp, Cr, Tr, E and, *C* =(*Cx*, *Cy*) are computed which are used to set the rules for classification of labelled segments into follicles and non-follicles. The *<sup>R</sup>*¯, *<sup>C</sup>*¯ *p*, *C*¯ *r*, *T*¯ *<sup>r</sup>*, *<sup>E</sup>*¯, (*<sup>C</sup>*¯ *x*, *C*¯ *y*) denote the mean values, and, the *σR*, *σCp*, *σCr*, *σTr*, *σE*, (*σCx*, *σCy*) denote the standard deviation values, of the geometric parameters R, Cp, Cr, Tr, E and C=(Cx,Cy).

Now, the classification rules for follicle recognition are formulated as following: A region with area A, ratio R, compactness Cp, circularity Cr, Tortousity Tr, extent E, and centriod *C* =(*Cx*, *Cy*), is classified as a follicle, if the following conditions are satisfied:

$$\mathbf{A} > \mathbf{T}\_{\mathbf{A}} \tag{1}$$

$$
\overline{R} - a\sigma\_R < R < \overline{R} + a\sigma\_R \tag{2}
$$

$$
\overline{\mathbb{C}p} - a\sigma\_{\mathbb{C}p} < \mathbb{C}p < \overline{\mathbb{C}p} + a\sigma\_{\mathbb{C}p} \tag{3}
$$

$$
\overline{\operatorname{Cr}} - a\,\sigma\_{\operatorname{Cr}} < \operatorname{Cr} < \overline{\operatorname{Cr}} + a\,\sigma\_{\operatorname{Cr}} \tag{4}
$$

$$
\overline{Tr} - \alpha \sigma\_{Tr} < Tr < \overline{Tr} + \alpha \sigma\_{Tr} \tag{5}
$$

**b.** Watershed segmentation method

**c.** Optimal thresholding method

**2.1. HRGMF and thresholding**

experts.

The follicle detection in ultrasound images of ovaries using watershed segmentation method uses the watershed transform for segmentation, geometric features for feature extraction and 3σ-intervals around the mean for the purpose of classification [22]. We follow three steps in the segmentation, first, the internal marker image is obtained, second, external marker image is obtained, and third, modifying gradient of the original image, by mask. In next phase, segmentation is carried out by applying marker controlled watershed transform to modified gradient image, region filling and clearing the border. Finally, preserve the regions which are most probably appear as follicles, while all other regions are removed. The feature extraction is based on seven geometric parameters of the follicles and the classification of regions for follicle detection is based on 3σ intervals around the mean. The experimental results are in

Follicle Detection and Ovarian Classification in Digital Ultrasound Images of Ovaries

http://dx.doi.org/10.5772/56518

175

In the automatic detection of follicle in ultrasound images of ovaries using optimal threshold‐ ing method, sobel operator and morphological opening and closing is used for preprocessing, optimal thresholding for segmentation and 3σ-intervals around the mean for classification are used [23]. Initially, the image is processed with finding the gradient by using sobel operators in horizontal direction. For the gradient image morphological opening and closing is applied by using appropriate structuring element. Then thresholding is applied to the filtered image by optimal thresholding method [24]. In the process, many undesired spurious regions are also obtained (e.g., regions inside the endometrium). These spurious regions must be removed as much as possible and all regions touching the borders are removed, which yields the segmented image. The feature extraction is based on seven geometric parameters of the follicles and the classification of regions for follicle detection is based on 3σ intervals around the mean. The experimental results are in good agreement with the manual follicle detection by medical

Ultrasound ovary images show planar sections through the follicles. These images are characterized by specular reflections and edge information, which is weak and discontinuous. Therefore, traditional edge detection techniques (e.g. Sobel, Prewitt) are susceptible to spurious responses when applied to ultrasound imagery. The main reason, that the follicle segmentation using solely edge information is rather unsuccessful, is a high level of added speckle noise [25]. The follicles being fluid-filled sacs appear as dark oval regions because they display similar fluid echotextures, which are more or less darker than their neighbourhood.

The main idea behind the HRGMF based method is as follows: [26]: In the first phase, the homogeneous region growing mean filter (HRGMF) is applied [27]. By using this filter, first homogeneous region can be identified and then this region is grown until it satisfies the similarity criteria. The value of filtering point is replaced by arithmetic mean of the grown region. In next phase, segmentation is carried out, by binarising the filtered image with three

The follicles could therefore be treated as homogeneous dark regions.

good agreement with the manual follicle detection by medical experts.

$$
\overline{\overline{E}} - \alpha \sigma\_{\overline{E}} < \mathcal{E} < \overline{\mathcal{E}} + \alpha \sigma\_{\overline{E}} \tag{6}
$$

$$
\overline{\mathbb{C}\mathfrak{x}} - a\sigma\_{\mathbb{C}\mathfrak{x}} < \mathbb{C}\mathfrak{x} < \overline{\mathbb{C}\mathfrak{x}} + a\sigma\_{\mathbb{C}\mathfrak{x}} \tag{7}
$$

$$
\overline{\mathbb{C}y} - \alpha \sigma\_{\mathbb{C}y} < \mathbb{C}y < \overline{\mathbb{C}y} + \alpha \sigma\_{\mathbb{C}y} \tag{8}
$$

The constants TA and α are empirically determined. In our experiments, we have determined that the empirical values are TA =30, α=3. The equations (1) to (8) constitute the 3σ-classifier, which is to be used in the testing phase.

*Testing phase* : During the testing phase, the area A, the ratio R, the compactness Cp, the circularity Cr, the tortousity Tr, the extent E and the centriod, *C* =(*Cx*, *Cy*) are computed for a region of the segmented image and apply the above classification rules to determine whether it is a follicle or not. The comparison of the experimental results for the two sam‐ ple images with the manual segmentation done by medical expert is presented in the Fig‐ ure 4. There is good agreement between the segmented image and the manual segmentation done by the medical expert, which demonstrates the efficiency of the edge based method. The manual segmentation is done by the team of three medical experts (Radiologists (2) and Gynaecologist (1)).

**Figure 4.** Some typical results: a) Original images, b) Segmentation by edge based method, c) Classified image, d) Manual segmentation by medical expert.

#### **b.** Watershed segmentation method

*Tr Tr Tr Tr Tr* -a < < +a

*E E E E* -a <E< +a

*Cx Cx Cx Cx Cx* -a < < +a

*Cy Cy Cy Cy Cy* -a < < +a

The constants TA and α are empirically determined. In our experiments, we have determined that the empirical values are TA =30, α=3. The equations (1) to (8) constitute the 3σ-classifier,

*Testing phase* : During the testing phase, the area A, the ratio R, the compactness Cp, the circularity Cr, the tortousity Tr, the extent E and the centriod, *C* =(*Cx*, *Cy*) are computed for a region of the segmented image and apply the above classification rules to determine whether it is a follicle or not. The comparison of the experimental results for the two sam‐ ple images with the manual segmentation done by medical expert is presented in the Fig‐ ure 4. There is good agreement between the segmented image and the manual segmentation done by the medical expert, which demonstrates the efficiency of the edge based method. The manual segmentation is done by the team of three medical experts

**Figure 4.** Some typical results: a) Original images, b) Segmentation by edge based method, c) Classified image, d)

 s

 s

> s

> s

image

(5)

(6)

(7)

(8)

d) Manual segmentation

s

s

s

s

a) Original image b) Segmented image c) Classified

which is to be used in the testing phase.

174 Advancements and Breakthroughs in Ultrasound Imaging

(Radiologists (2) and Gynaecologist (1)).

Manual segmentation by medical expert.

The follicle detection in ultrasound images of ovaries using watershed segmentation method uses the watershed transform for segmentation, geometric features for feature extraction and 3σ-intervals around the mean for the purpose of classification [22]. We follow three steps in the segmentation, first, the internal marker image is obtained, second, external marker image is obtained, and third, modifying gradient of the original image, by mask. In next phase, segmentation is carried out by applying marker controlled watershed transform to modified gradient image, region filling and clearing the border. Finally, preserve the regions which are most probably appear as follicles, while all other regions are removed. The feature extraction is based on seven geometric parameters of the follicles and the classification of regions for follicle detection is based on 3σ intervals around the mean. The experimental results are in good agreement with the manual follicle detection by medical experts.

**c.** Optimal thresholding method

In the automatic detection of follicle in ultrasound images of ovaries using optimal threshold‐ ing method, sobel operator and morphological opening and closing is used for preprocessing, optimal thresholding for segmentation and 3σ-intervals around the mean for classification are used [23]. Initially, the image is processed with finding the gradient by using sobel operators in horizontal direction. For the gradient image morphological opening and closing is applied by using appropriate structuring element. Then thresholding is applied to the filtered image by optimal thresholding method [24]. In the process, many undesired spurious regions are also obtained (e.g., regions inside the endometrium). These spurious regions must be removed as much as possible and all regions touching the borders are removed, which yields the segmented image. The feature extraction is based on seven geometric parameters of the follicles and the classification of regions for follicle detection is based on 3σ intervals around the mean. The experimental results are in good agreement with the manual follicle detection by medical experts.

### **2.1. HRGMF and thresholding**

Ultrasound ovary images show planar sections through the follicles. These images are characterized by specular reflections and edge information, which is weak and discontinuous. Therefore, traditional edge detection techniques (e.g. Sobel, Prewitt) are susceptible to spurious responses when applied to ultrasound imagery. The main reason, that the follicle segmentation using solely edge information is rather unsuccessful, is a high level of added speckle noise [25]. The follicles being fluid-filled sacs appear as dark oval regions because they display similar fluid echotextures, which are more or less darker than their neighbourhood. The follicles could therefore be treated as homogeneous dark regions.

The main idea behind the HRGMF based method is as follows: [26]: In the first phase, the homogeneous region growing mean filter (HRGMF) is applied [27]. By using this filter, first homogeneous region can be identified and then this region is grown until it satisfies the similarity criteria. The value of filtering point is replaced by arithmetic mean of the grown region. In next phase, segmentation is carried out, by binarising the filtered image with three different thresholds, T1, set as standard deviation of the input image, T2, set as mean of the input image, T3, set as abs(T2-T1). The resultant images are combined after thresholding. The components are labelled, the holes inside the regions are filled, the borders if any are cleared. Finally, the regions which most probably appear as follicles are preserved, while all other regions are removed. The feature extraction is based on seven geometric parameters of the follicles and the classification of regions for follicle detection is based on 3σ intervals around the mean.

ultrasound device is not moist enough). Especially, a disturbing type of noise is the speckle noise. Therefore the more efficient speckle reduction method based on the contourlet transform is used for denoising medical ultrasound images [28]. The contourlets can be loosely inter‐ preted as a grouping of nearby wavelet coefficients since their locaters are locally correlated due to smoothness of the boundary curve [29, 30]. The contourlet transform is a multiscale and multidirectional framework of discrete image [31]. It is the simple directional extension for wavelet that fixes its subband mixing problem and improves its directionality. In this trans‐ form, the multiscale and multidirectional analyses are separated in a serial way. The laplacian pyramid (LP), is first used to capture the point discontinuities followed by a directional filter bank (DFB) to link point discontinuities into linear structure. Thus, we perform a wavelet like transform for edge detection, and then a local directional transform for contour segment detection. In other words, the contourlet transform comprises a double filter bank approach

Follicle Detection and Ovarian Classification in Digital Ultrasound Images of Ovaries

http://dx.doi.org/10.5772/56518

177

The contourlet transform exploits smoothness of contour effectively by considering variety of directions following contour. The contourlet transform can be designed to be a tight frame along with thresholding in order to achieve denoising of the image more effectively. The algorithm for contourlet transform method is as follows [32,33] : Firstly, apply the log trans‐ form to the input ultrasound image. Then, apply the contourlet transform on the log trans‐ formed image upto n levels of Laplacian pyramidal decomposition and m directional decompositions at each level, where n and m depend on the image size. Next, perform thresholding of contourlet transformed image. Lastly, the despeckled image is obtained by performing inverse contourlet transform on the thresholded image. Then, histogram equali‐

Firstly, all the pixels which are darker than their neighborhood row wise (horizontal scan) and then column wise (vertical scan) is collected. Then, the two resultant images are added to yield a segmented image. The regions with area less than a threshold value are removed. The holes inside the region are filled. The nonzero pixels touching the image border are set to zero [34,35].

*Horizontal Scanline Thesholding (HST)* : The follicle region will have more or less same grey level value for all the pixels within it. The horizontal scanline thresholding method is described as

of the ith row subimage of *f* are given by the equations 9 and 10 :

1 <sup>1</sup> (,) *N*

<sup>1</sup> ( (,) )

*fij m <sup>N</sup>* <sup>=</sup> æ ö s = ç ÷ è ø

2

<sup>=</sup> å (9)

å (10)

*j m fij <sup>N</sup>* <sup>=</sup>

1

*N i i j*

and standard

follows; Consider the input image *f* of the size *MxN* . The sample mean *mi*

*i*

for obtaining sparse expansions for typical images with contours.

zation is applied to enhance the contrast of the despeckled image. **i.** Horizontal and Vertical Scanline (HVST) based method

deviation *σ<sup>i</sup>*

The experimentation has been done using two datasets D1 and D2. The D1 set consists of the 80 sample ultrasound ovarian images of size 256x256, out of which 40 images are used for training and 40 for testing. The D2 set consists of 90 sample ultrasound images of ovaries of size 512x512, out of which 45 images are used for training and 45 for testing. The ten-fold experiments are performed for the classification and the average follicle detection rate is computed.The Table 1 shows the classification results of the proposed method after ten-fold experiments for both the data sets D1 and D2. The average detection rates for the proposed method with D1 and D2 sets are 79.77% and 68.86%, false acceptance rates (FAR) are 25.99% and 28.41%, and false rejection rates (FRR) are 20.52% and 31.12%, respectively.


**Table 1.** Classification results of proposed method after ten-fold experiments.

The proposed HRGMF based method for follicle detection is more effective as compared to the edge based method, watershed based segmentation method and optimal thresholding methods. The experimental results are in good agreement with the manual follicle detection by medical experts, and thus demonstrate efficacy of the method. Thus, the proposed HRGMF based method for follicle segmentation is effective in computer assisted fertility diagnosis by the experts.

#### **2.2. Contourlet transform and scanline thresholding**

The follicle detection rate in the homogeneous region growing mean filter (HRGMF) based method is considerably improved. However, due to speckle noise, finding the object bounda‐ ries is difficult and thus leads to poor segmentation. To improve the segmentation accuracy, the contourlet transform is used for despeckling the ultrasound image followed by histogram equalization, and, thereafter, the horizontal and vertical scanline thresholding (HVST) method is employed for follicle detection using geometric parameters.

*Contourlet transform* : The first step of the algorithm is denoising the image, since ultrasound images are invariably noisy due to the mode of the image acquisition itself. (e.g. head of the ultrasound device is not moist enough). Especially, a disturbing type of noise is the speckle noise. Therefore the more efficient speckle reduction method based on the contourlet transform is used for denoising medical ultrasound images [28]. The contourlets can be loosely inter‐ preted as a grouping of nearby wavelet coefficients since their locaters are locally correlated due to smoothness of the boundary curve [29, 30]. The contourlet transform is a multiscale and multidirectional framework of discrete image [31]. It is the simple directional extension for wavelet that fixes its subband mixing problem and improves its directionality. In this trans‐ form, the multiscale and multidirectional analyses are separated in a serial way. The laplacian pyramid (LP), is first used to capture the point discontinuities followed by a directional filter bank (DFB) to link point discontinuities into linear structure. Thus, we perform a wavelet like transform for edge detection, and then a local directional transform for contour segment detection. In other words, the contourlet transform comprises a double filter bank approach for obtaining sparse expansions for typical images with contours.

The contourlet transform exploits smoothness of contour effectively by considering variety of directions following contour. The contourlet transform can be designed to be a tight frame along with thresholding in order to achieve denoising of the image more effectively. The algorithm for contourlet transform method is as follows [32,33] : Firstly, apply the log trans‐ form to the input ultrasound image. Then, apply the contourlet transform on the log trans‐ formed image upto n levels of Laplacian pyramidal decomposition and m directional decompositions at each level, where n and m depend on the image size. Next, perform thresholding of contourlet transformed image. Lastly, the despeckled image is obtained by performing inverse contourlet transform on the thresholded image. Then, histogram equali‐ zation is applied to enhance the contrast of the despeckled image.

#### **i.** Horizontal and Vertical Scanline (HVST) based method

different thresholds, T1, set as standard deviation of the input image, T2, set as mean of the input image, T3, set as abs(T2-T1). The resultant images are combined after thresholding. The components are labelled, the holes inside the regions are filled, the borders if any are cleared. Finally, the regions which most probably appear as follicles are preserved, while all other regions are removed. The feature extraction is based on seven geometric parameters of the follicles and the classification of regions for follicle detection is based on 3σ intervals around

The experimentation has been done using two datasets D1 and D2. The D1 set consists of the 80 sample ultrasound ovarian images of size 256x256, out of which 40 images are used for training and 40 for testing. The D2 set consists of 90 sample ultrasound images of ovaries of size 512x512, out of which 45 images are used for training and 45 for testing. The ten-fold experiments are performed for the classification and the average follicle detection rate is computed.The Table 1 shows the classification results of the proposed method after ten-fold experiments for both the data sets D1 and D2. The average detection rates for the proposed method with D1 and D2 sets are 79.77% and 68.86%, false acceptance rates (FAR) are 25.99%

> Data set D1 set D2 set Classification rate 79.77% 68.86% Type I error (FAR) 25.99% 28.41% Type II error(FRR) 20.52% 31.12%

The proposed HRGMF based method for follicle detection is more effective as compared to the edge based method, watershed based segmentation method and optimal thresholding methods. The experimental results are in good agreement with the manual follicle detection by medical experts, and thus demonstrate efficacy of the method. Thus, the proposed HRGMF based method for follicle segmentation is effective in computer assisted fertility diagnosis by

The follicle detection rate in the homogeneous region growing mean filter (HRGMF) based method is considerably improved. However, due to speckle noise, finding the object bounda‐ ries is difficult and thus leads to poor segmentation. To improve the segmentation accuracy, the contourlet transform is used for despeckling the ultrasound image followed by histogram equalization, and, thereafter, the horizontal and vertical scanline thresholding (HVST) method

*Contourlet transform* : The first step of the algorithm is denoising the image, since ultrasound images are invariably noisy due to the mode of the image acquisition itself. (e.g. head of the

**Classification results for ten-fold experiments**

and 28.41%, and false rejection rates (FRR) are 20.52% and 31.12%, respectively.

**Table 1.** Classification results of proposed method after ten-fold experiments.

**2.2. Contourlet transform and scanline thresholding**

is employed for follicle detection using geometric parameters.

the mean.

176 Advancements and Breakthroughs in Ultrasound Imaging

the experts.

Firstly, all the pixels which are darker than their neighborhood row wise (horizontal scan) and then column wise (vertical scan) is collected. Then, the two resultant images are added to yield a segmented image. The regions with area less than a threshold value are removed. The holes inside the region are filled. The nonzero pixels touching the image border are set to zero [34,35].

*Horizontal Scanline Thesholding (HST)* : The follicle region will have more or less same grey level value for all the pixels within it. The horizontal scanline thresholding method is described as follows; Consider the input image *f* of the size *MxN* . The sample mean *mi* and standard deviation *σ<sup>i</sup>* of the ith row subimage of *f* are given by the equations 9 and 10 :

$$m\_i = \frac{1}{N} \sum\_{j=1}^{N} f(i, j) \tag{9}$$

$$\sigma\_i = \sqrt{\left(\frac{1}{N} \sum\_{j=1}^{N} (f(i,j) - m\_i)^2\right)}\tag{10}$$

Now, set T4 = *mi* and T5 = *σ<sup>i</sup>* . Then multiply T4 and T5 by positive scale factors K1 and K2 (empirically fixed). Binarize ith row subimage using the thresholds K1T4 and K2T5 separately. This procedure is carried out for the rows i=1,...,M and obtain the horizontal mean (standard deviation) thresholded image fhm (fhsd).

*Vertical Scanline Thresholding (VST)*: The VST method is described as follows; Consider the input image *f* of the size *MxN* . The sample mean *mj* and standard deviation *σ<sup>j</sup>* of the jth column subimage are given by the equations 11 and 12:

$$m\_j = \frac{1}{M} \sum\_{i=1}^{M} f(i, j) \tag{11}$$

**ii.** *Edge based method:* The edge based segmentation method, which is described in the

The HVST based method for the follicle detection is more effective as compared to edge based

The active contour method is used for segmentation of ultrasound image to increase the follicle detection accuracy. Either 3σ intervals based classifier or fuzzy classifier may be employed for

The follicle detection in ultrasound images of ovaries using active contour method uses the contourlet transform based method for preprocessing, active contour method for segmentation and 3σ-intervals around the mean for classification [37]. Initially, the input image is despekled by using the contourlet transform method. Next, histogram equalization is applied to enhance the contrast of the despeckled image. Further, the negative transformation is applied on the histogram equalized image [38], as the proposed segmentation method works on high intensity valued objects. In the segmentation stage, the active contour without edges method is used [39]. The resulting image after applying active contour method contains segmented regions with it. The geometric features are extracted and then 3σ-intervals around the mean is used

The Figure 5 (a) depicts sample original ultrasound image of the ovary. The resultant images obtained at different steps of the proposed method are shown in the Figure 5 (b)-(e). Many undesired spurious regions are also obtained (e.g., regions inside the endometrium). These spurious regions must be removed as much as possible. Therefore, the regions having an area less than T (empirical value) are removed (Figure 5(f)). The Figure 5(g) depicts the segmented follicles (outlined in white) superimposed on the original image. The Figure 5(h) depicts the recognized follicles after applying the classification rules and the Figure 5(i) shows the follicles

The Figure 6 depicts comparison of the active contour method with the HVST based segmen‐ tation method. In the Figure 6, (a) and (e) are two original images, while (b) and (f) are their corresponding segmented images by HVST method. Similarly, (c) and (g) are segmented images of active contour method. It is observed that the follicles, which were not detected by HVST based segmentation method (Figure 6 (b) and (f)), are correctly identified by the active contour method (Figure 6 (c) and (g)). The Figure 6 (d) and (h) show the manual segmentation rendered by the medical expert. Hence, the classification accuracy is improved in the active

Gaussian low pass filter and edge based segmentation.

**i.** Active contour method with 3σ intervals based classification

**3. Active contour method for follicle detection**

follicle detection.

for classification.

annotated manually by the medical expert.

segmentation after applying contourlet transform and histogram equalization.

section 2 (a), is applied to the histogram equalized image obtained after despeckling the input image using contourlet transform [36]. The follicle detection rate is im‐ proved in the edge based segmentation after applying contourlet transform and histogram equalization, as compared to the method in the section 2(a) which employs

Follicle Detection and Ovarian Classification in Digital Ultrasound Images of Ovaries

http://dx.doi.org/10.5772/56518

179

$$
\sigma\_j = \sqrt{\frac{1}{M}} \sum\_{i=1}^{M} (f(i, j) - m\_j)^2 \tag{12}
$$

Now, set T6 = *mj* and T7 = *σ<sup>j</sup>* . Then multiply T6 and T7 by positive scale factors K3 and K4 ( empirically fixed), respectively. Binarize jth column subimage using thresholds K3T6 and K4T7 separately. This procedure is carried out for the columns j=1,...,N and obtain the vertical mean (standard deviation) thresholded image fvm ( fvsd ).

*Image Fusion*: The resultant images fhm and fvm are combined, to yield the image fhvm. Any region touching the borders are removed. The resultant images fhsd and fvsd are combined, to yield the image fhvsd. Any region touching the borders are removed. Finally, the resultant images fhvm and fhvsd are combined, to yield the segmented image fseg, which contains segmented regions in it. The regions in fseg having smaller area than the threshold T8 are removed. The regions are labeled (identified) and possible holes inside them are filled.

The geometric features are extracted and then 3σ-intervals around the mean are used for classification. The Table 2 shows the classification results of the HVST based method after tenfold experiments for both the data sets D1 and D2. The average detection rates for the proposed Method I with D1 and D2 sets are 90.10% and 92.76%, false acceptance rates (FAR) are 11.71% and 9.89%, and false rejection rates (FRR) are 21.55% and 7.23%, respectively.


**Table 2.** Classification results of proposed method after ten-fold experiments.

**ii.** *Edge based method:* The edge based segmentation method, which is described in the section 2 (a), is applied to the histogram equalized image obtained after despeckling the input image using contourlet transform [36]. The follicle detection rate is im‐ proved in the edge based segmentation after applying contourlet transform and histogram equalization, as compared to the method in the section 2(a) which employs Gaussian low pass filter and edge based segmentation.

The HVST based method for the follicle detection is more effective as compared to edge based segmentation after applying contourlet transform and histogram equalization.
