An Efficient Follicle Segmentation and Optimized Deep Learning Based PCOS Prediction System

Polycystic Ovary Syndrome (PCOS) is a significant medical ailment affecting women’s reproduction because of hormonal disorders [1, 2]. The collections of fluids are called as follicles that are created in the ovaries and the situation of failing to release the eggs. The different sizes of follicles are developed in the ovary, called as hormonal disorder. This hormonal disorder is known as PCOS. The entire research shows the results of 70% of women are not able to detect the PCOS to cure. The condition of PCOS is described as the failure of the ovary to release the ovum due to the follicle’s formation in the ovaries. The women affected by PCOS have been struggling with hormone balance which creates various health issues such as irregular menstrual cycle, miscarriage, and pregnancy issues. People are suffered from these imbalanced hormones from the age of 16 to 40 years [3]. It will cause the diseases such as cardiovascular disease (CVD), obesity, hypertension, Type 2 Diabetes, gynecological cancer, and risks in pregnancy.

The appropriate method for the detection of PCOS is the Rotterdam criteria. The principle is to identify the ovaries with a capacity larger than or equals to 10 cm cube or the follicle of size 2-9 mm with a count of 10 or more. The ultrasound image (US) of normal and PCOS-affected ovaries is shown in Figure 1.

Figure 1. Normal ovary (left) and Polycystic Ovary Syndrome (PCOS) affected ovary (right)

The early detection of PCOS symptoms helps the affected persons to adapt to the changes in their lifestyle. PCOS women undergo infertility which leads to gynecological cancer [4]. The early identification of PCOS can save their miscarriage and lifespan. The automation systems will overcome these issues. It plays a significant role in healthcare industries [5]. The ML models can deal with enormous dataset processing for diagnosis. ML can be used in medical fields for image processing, documentation, and genetics assessment in the medical field [6].

This paper aimed to develop efficient follicle segmentation and deep learning with meta-heuristic (MH) based classification of UI ovary images. Through the segmentation process, this paper finds the more accurate follicle parameters such as location, size, and shape. This incorporates it with the DL model to find the prediction of the US as PCOS or normal ovary.

3.1 Dataset description

The PCOS dataset [14] helps for this study consists of 541 patients with clinical and physical parameters 223, which form the dataset. This dataset has 41 features [15] that are used for this feature extraction process. Unnecessary data is removed from preprocessing stage. Out of these records, 364 patients exhibit typical data patterns, while 177 patients exhibit signs of PCOS. The data about PCOS has been gathered from ten hospitals across Kerala, India.

3.2 Proposed materials and methods

The proposed PCOS detection and classification system overview architecture is illustrated in Figure 2. The input US image from the dataset is given as input to the model with the features such as BMI, cycle, menstrual LH and FSH data etc. Initially, the image data is preprocessed using techniques like image resizing, histogram equalization, and noise removal to enhance the image efficiency and then the relevant features are selected using Chi-Square (CS) model. Second, the cyst region segmentation is performed using the active contour Otsu method (ACOM), which efficiently segments the position of cyst, size, and location for good classification. Using the method of first convolution layer of CNN, the relevant features are extracted from the segmented image.

The classification task considered in this study is a binary classification problem. Based on the extracted follicle features from ultrasound images, the proposed CNN with Fruitfly Optimization approaches (CNN-FOA) model classifies each input ovary image into one of two categories: The classification is performed using both image-based features obtained from segmented follicle regions and selected clinical features.

Figure 2. Proposed Polycystic Ovary Syndrome (PCOS) detection and classification system overview architecture

3.2.1 Preprocessing

It is used image preprocessing techniques such as image resizing, histogram equalization, and median filter-based noise removal approaches [16].

Image Resizing: The input US image is resized using the Object Carving algorithm, which will minimize artifacts and salient alteration using the quick multi-operator method [17].

Histogram Equalization: The image’s brightness is enhanced using the histogram equalization approach. Using the uniform intensity, the image quality is increased. The image equalization methods are executed using the mathematical expression,

$HE=histeq\left( image \right),H=imhist\left( Image,~64 \right)$

The transformed intensity levels are obtained by mapping the original pixel values using the normalized CDF. This operation improves follicle boundary visibility and contrast, which is critical for accurate segmentation.

Noise Removal using Median Filter: This paper utilized the nonlinear statistical filter called median filter to denoising the image [18]. The output of the median filtering is shown in Eq. (1).

$\left( X,Y \right)=med\{f\left( X-i,~Y-h \right)i,j\in M$         (1)

where, the actual image is represented as f(X,Y), the output filtered image is represented as g(X,Y), and M is the 2D mask with the size n*n [19]. The preprocessing of the input Ultra sound (US) image is given in Figure 3.

Feature Selection: Among the 41 features, the Serial no and patient file numbers are discarded. The remaining 39 features are fed as input to the selected appropriate feature. Using the CS model, the feature importance values are examined [19]. The element with the value zero is nonvital, and it is dropped. In this process, the characteristics include age, BMI, cycle, cycle length, marriage status, FSH, LH, AMH, Vitamin D3, weight gain, PRG, hair growth, hair loss, pimples, skin darkening, left and right follicle no and fast food.

Input Image

Resized Image

Histogram Equalization

Noise Removal using Median Filter

Figure 3. Preprocessing outcomes of input US image

3.2.2 Segmentation using active contour Otsu method

Segmentation is partitioning the image into the region of desired features. Due to the occurrence of soft tissues and blood vessel, segmenting the follicles regions from the ovary picture is a difficult process. To discriminate the object from the background, thresholding is the best model based on the minimization function. Otsu’s Method from the study by Alagarsamy et al. [20] has been used to find the optimal threshold which separates the preprocessed into two groups of binarized image. The mean of these groups is computed as in Eq. (2).

${{\mu }_{1}}\left( t \right)=\underset{i=1}{\overset{t}{\mathop \sum }}\,\frac{i.f\left( i \right)}{{{g}_{1}}\left( t \right)}~and~{{\mu }_{2}}\left( t \right)=\underset{i=t+1}{\overset{k}{\mathop \sum }}\,\frac{i.f\left( i \right)}{{{g}_{2}}\left( t \right)}$          (2)

where, k is the number of gray level ranges from 1 to k. The optimal threshold is computed as ${t}'$.

${t}'=arg\underset{t}{\mathop{\min }}\,\sigma _{m}^{2}\left( t \right)$     (3)

where, $\sigma $ is the class variance and it is denoted in Eq. (4).

$\sigma _{m}^{2}\left( t \right)={{g}_{1}}\left( t \right)\sigma _{1}^{2}\left( t \right)+{{g}_{2}}\left( t \right)\sigma _{2}^{2}\left( t \right)$       (4)

where, ${{g}_{1}},~{{g}_{2}},\sigma _{1}^{2}\left( t \right)~and~\sigma _{2}^{2}\left( t \right)$ are the group probabilities and variances which is computed as in Eq. (5).

${{g}_{1}}=\underset{i=1}{\overset{t}{\mathop \sum }}\,f\left( i \right)~and~{{g}_{2}}=\underset{i=t+1}{\overset{k}{\mathop \sum }}\,f\left( i \right)$       (5)

$\sigma _{1}^{2}\left( t \right)=\underset{i=1}{\overset{t}{\mathop \sum }}\,{{\left[ i-{{\mu }_{1}}\left( t \right) \right]}^{2}}\frac{f\left( i \right)}{{{g}_{1}}\left( t \right)}$       (6)

$\sigma _{2}^{2}\left( t \right)=\underset{i=t+1}{\overset{k}{\mathop \sum }}\,{{\left[ i-{{\mu }_{2}}\left( t \right) \right]}^{2}}\frac{f\left( i \right)}{{{g}_{2}}\left( t \right)}$   (7)

The conventional Otsu method is not producing proper results in follicle segmentation and it is modified with the active contour to provide accurate segmentation. The Otsu method is modified with the computation of new threshold value with the iterative manner. Initial threshold t1 is computed by averaging the image value of ovary. Next, the above the mean value and below the threshold value is computed and it is denoted as m1 and m2. The new value of threshold is computed as in Eq. (8).

$t1=\frac{m_{1}^{t}+m_{2}^{t}}{n}$       (8)

where, t is the number of iterations, the $t1$ is forwarded to the next iteration to compute t2 and so on. In the 2^nd iteration, the image value is greater or equal to t1. In this, the mean value below and above the threshold t1 is computed. The iteration process is continuing till the converge [t(k)-t(k-1)] and the last iteration is computed as,

${{t}^{*}}=\frac{m_{1}^{k}+m_{2}^{k}}{n}$       (9)

Now, this iteratively computed threshold ${{t}^{*}}$ is computed in the traditional Otsu method Eq. (2) as in Eqs. (10) and (11).

${{\mu }_{1}}\left( t \right)=\underset{i=1}{\overset{{{t}^{*}}}{\mathop \sum }}\,\frac{if\left( i \right)}{{{g}_{1}}\left( t \right)}$      (10)

${{\mu }_{2}}\left( t \right)=\underset{i={{t}^{*}}+1}{\overset{k}{\mathop \sum }}\,\frac{if\left( i \right)}{{{g}_{2}}\left( t \right)}$       (11)

The total mean of the entire image is computed as in Eq. (12).

${{\mu }_{t}}=\underset{i=1}{\overset{k}{\mathop \sum }}\,if\left( i \right)$        (12)

The position of the initial contour closest to the preferred boundaries is found using the level set method proposed by Osher and Sethian [20]. The segmentation process using the ACOM is illustrated in Figure 4.

Preprocessed Image

OTSU method

Iterative OTSU method

Binary Mask

ACOM

Final Output

Figure 4. Segmentation follicles using active contour Otsu method (ACOM)

3.2.3 Fruitfly Optimization algorithm

The Fruitfly Optimization Algorithm (FOA) to use the foraging behavior of fruit flies to determine a globally optimal solution (Pan 2012). The fruit fly can fully utilize its instincts to find features with disorder since it has superior vision and smell compared to other species. In the proposed CNN-FOA framework, the Fruitfly Optimization Algorithm is employed to optimize CNN hyperparameters including weights, biases, batch size, and number of epochs. The population size is set to 30 fruit flies, and the maximum number of iterations is fixed at 100. At each iteration, candidate CNN parameters are evaluated using classification accuracy on the validation set as the fitness value [21]. First, fruit flies use their smell parts to detect the presence of follicles and then fly in that direction. Second, fruit flies can identify follicles and other fruit flies thanks to their keen vision when they get close to the diet source.

3.2.4 Feature extraction and classification using CNN-FOA

From these segmented results, the details about the follicles features are extracted using the first convolution layer of CNN and the follicles are classified using the CNN-FOA model. The features are like as such as perimeter, area, aspect ratio (AR), major axis, minor axis and the eccentricity. In the proposed CNN-FOA model, the first convolution layer has been used for feature extraction process. The CNN architecture is stated in Figure 5. It consists of two convolution, two pooling and four fully connected layers. The first convolution layer has 64 filters and each filer size is 1*3 with astride of size 1, followed by a ReLU activation function [22]. The second convolutional layer uses 128 filters of size 1 × 3. Max-pooling layers with a pool size of 2 × 2 are applied after each convolutional layer to reduce spatial dimensionality. The data fed as input is 2D array representation of the input image. Using the data construction, the CNN model studies the complex features and the convolution procedure have been used to extract the features.

$ReLu\left( X \right)=\text{max}\left( 0,X \right)$         (13)

The over fitting issues are prevented and the complexity is reduced using Dropout layers. The regularization rate is fixed as 0.5. The input data convolution operation for the previous X^l-1 layer is declared in Eq. (14).

${{X}^{l}}={{\omega }^{l}}.{{X}^{l-1}}+{{b}^{l}}$       (14)

where, weight is represented as w and b of l^th layer bias respectively and output is ${{X}^{l}}$. The last pooling layer extracted features listed in Table 1 are fed as input to fully connected layer (FC). To obtain the output with SoftMax activation function, FC4 is employed. CNN, as a regularization technique, utilizes batch normalization (BN) to normalize the input data for FC4. The extracted feature vector is given as input to the classification process from the convolution layer of size (1*64), which. The output layer data computation is stated as,

$Y=T+{{\delta }^{l}}$       (15)

where, T is the target value and $\delta $ is the error of the neuron.

Figure 5. CNN with Fruitfly Optimization approaches architecture based feature extraction and classification

Table 1. Extracted features of follicles

3.2.5 Hyper parameter optimization using FOA

The proposed CNN-FOA based classification process comprised of two process such as parameter optimization and classification process. During the parameter optimization process, the CNN parameters are adjusted using FOA. The fitness function is considered as classification accuracy which isstated in Eq. (16).

$fitness=\frac{\mathop{\sum }_{i=1}^{k}{{A}_{i}}}{k}$          (16)

where, A is the obtained accuracy from CNN classifier. The process involved in steps is stated as follows:

Initialize value of size of population, maximum iterations number, variables upper and lower bounds, problem dimension like bias and weight. Consider the best FF position as the FF position called global optimum. Each FF position is updated as in Eq. (17) with the levy flight method and the fitness is evaluated.

$X_{i}^{levy}={{X}_{i}}+{{X}_{i}}\oplus levy~\left( s \right)$        (17)

where, $X_{i}^{levy}$ is i^th search agent new position of ${{X}_{i}}$. The levy distribution is described as,

$levy~\left( s \right)\sim {{\left| s \right|}^{-1-\beta }},~~~0\le \beta \le 2$         (18)

where, $\beta $ is the levy index and s is the step length.

Check with the best individual population, if it better than the global optimum means update the values in global solution. To initiate the values of weight w and bias b in the CNN model structure. For each of training sample X in D and for each input layer unit i. Check with input unit with the output value ${{X}_{i}}={{Y}_{i}}$. For each and every hidden layer (or output layer) unit i

${{X}_{i}}=\underset{j=1-m}{\mathop \sum }\,{{w}_{ij}}{{X}_{i}}+{{b}_{j}}$        (19)

${{Y}_{i}}=T+{{\delta }^{l}}$    (20)

For each weight ${{w}_{ij}}$ in the network, the weight increment is computed as $\Delta w_{i j}=(l) E_i Y_j$. The weight is then updated as $w_{i j}=w_{i j}+\Delta w_{i j}$. For each bias b in the network. To increase the bias $\Delta b_i=(l) E_i$ and update the bias as $b_j=b_j+\Delta b_j$. Return the follicle detection as normal or PCOS.