Early Detection and Segmentation of Ovarian Tumor Using Convolutional Neural Network with Ultrasound Imaging

Women of all ages might be affected by ovarian and adnexal tumors, which are frequent gynecological issues. Less than three cm in diameter follicular and corpus luteum cysts appear monthly in premenopausal women. Cyst development is still 13–16% more likely in postmenopausal women. Therefore, ovarian and adnexal tumors need to be assessed, tracked, and treated when performing pelvic ultrasound exams on women in clinical settings [1]. The majority of ovarian cysts is physiological, is more prevalent in premenopausal women, and eventually goes away. Ovarian cysts that are not physiological make up a very small portion of all cysts and will continue to grow. An abdominal or vaginal ultrasound can be done by a gynecologist to differentiate Malignant cysts differ from benign cysts. Frequently, monitoring benign ovarian cancers is possible without adjusting the interval. Only if benign tumors persist or grow further, with the potential for torsion and rupture, are operations undertaken. Nevertheless, to lower subsequent mortality, malignant tumors necessitate prompt cytoreductive surgery and then chemotherapy.

The most deadly malignancy of the female reproductive system is ovarian cancer. Around 240,000 women were given ovarian cancer diagnoses worldwide in 2012, and 150,000 of them women passed away from the disease [2]. 53% of ovarian cancer patients, according to United States Cancer Statistics, pass away within 5 years of their diagnosis, and this death rate can rise as a result of diagnostic lag time and late-stage diagnosis [1].

Gynecologists have a tough time diagnosing ovarian tumors since they affect women of different ages. Upon discovery of an ovarian mass, the clinician's top concern is to ascertain if it is a benign or malignant tumor and to evaluate the alternatives for the best care [2], the most often used imaging technique for evaluating ovarian cancers is pelvic ultrasonography, which offers details about the tumor such as its size, consistency, and type. Because results can differ between interpreters, ultrasonography's main drawback is that it is operator-dependent [3]. Recent studies have looked at the value of texture analysis in the diagnosis of breast lesions on digital X-ray mammography using deep learning and the detection of cervical neoplasms on colposcopy images [4]. By using 187 ultrasound images of ovarian tumors, created a Support Vector Machine (SVM) to distinguish between benign and malignant lesions. Additionally, a three-dimensional texture analysis algorithm was created to assess structural differences in the extracellular matrix between normal ovary and serous ovarian tumors [5].

After ovarian cysts are discovered, the typical clinical difficulty that faces obstetricians, gynecologists, and ultrasound examiners is differentiating between benign and harmful ovarian cancers [6]. Ovarian tumors come in many different varieties, and many of them share traits that make clinical diagnosis more difficult. An unclear diagnosis can make choosing a course of treatment more challenging and make a patient feel more anxious. For instance, debunking surgery, which necessitates significant abdomen incisions, is a typical surgical procedure for treating malignant tumors.

Minimally invasive surgery can be used to monitor or treat benign ovarian tumors. Patients' wounds and prognosis are both impacted by an incorrect diagnosis [7].

Since the concept of the CNN was first presented in 2012, it has been widely applied to tasks like image classification, object identification, and image segmentation [8]. Computer-Aided Diagnosis Systems (CADs) are evolving quickly because of the growth of big data in medicine and deep learning [9]. By 2040, the market for image-based artificial intelligence medical diagnostics would have grown by almost 10000%, predicts IDTechEx, a well-known British research company. Deep learning has so far been widely used in the diagnosis of a variety of ailments, such as the detection of benign and malignant thyroid nodules, thyroid nodule identification, and lung cancer [10].

Deep learning techniques have been used in a few research projects to diagnose ovarian tumors, although the majority of research projects base their categorization of ovarian cancer on artificial image segmentation. With medical photos, locating and defining the tumors' boundaries is equally crucial. The task of segmenting an ovarian tumor is extremely difficult for a number of reasons, though. Due to the limited soft tissue contrast, some tiny tumors can readily be mistaken for adjacent tissues [11].

Moreover, due to individual variance, the ovarian tumor's shape and size are unstable and change irregularly. A polycentric distribution of ovarian carcinoma metastases results from the aberrant cells' capacity to infiltrate or spread to different sections of the body as the cancer progresses.

Obstetricians, gynecologists, and ultrasound examiners must deal with the typical clinical challenge of differentiating between benign and malignant ovarian tumors after finding ovarian cysts. Ovarian tumors come in several different varieties, and many of them share traits that make clinical diagnosis more difficult. We want to create a technique that quickly and accurately supports the various diagnoses of ovarian cancers in ultrasound pictures using the CNN model.

Using 1813 ovarian ultrasound pictures obtained from 1254 individuals, the efficiency of the suggested CNN is verified. The deep learning visualization technique and decline in sorting efficiency verify the qualitative and quantitative effects of picture disruption on texture analysis [12].

Our proposed CNN is what we consider a workable deep learning-based diagnostic tool for differentiating ovarian cancers. As compared to cutting edge approaches, the proposed method has demonstrated greater performance. An innovative deep learning strategy is suggested for the segmentation of ovarian tumors from ultrasound pictures, summarizing the key contributions of our study. This is, as far as we are aware, the first deep learning technique for segmenting ovarian cancers using ultrasound pictures. To address the polycentric and size variation issues associated with ovarian tumors, a multi-scale graph learning framework has been developed. This framework makes use of long-range relationships among the ovarian tumors to improve segmentation performance.

1.png

Figure 1. Proposed methodology for the classification of ovarian tumor

This section explains the method for detecting ovarian cancers that is based on deep learning architectures for segmentation and classification a multi-network feature extraction model and deep learning classifier CNN are suggested for use in the implementation of utilizing medical histopathology pictures.

The Convolutional Neural Network (CNN or ConvNet) is a subtype of Neural Networks that is mainly used for applications in image and speech recognition. Its built-in convolutional layer reduces the high dimensionality of images without losing its information. That is why CNNs are especially suited for this use case. The research's methodology is shown in Figure 1.

3.1 Datasets

The main source of evidence for the efficiency of the suggested strategy was our collection of images of ovarian tumors. As there isn't a publicly accessible liver image dataset, one was used instead to compare with current medical picture segmentation techniques [16]. Zhejiang University's School of Medicine's Affiliated Women's Hospital offered an ovarian tumor dataset used in this work, which includes 196 Ultrasound Images (UI) in total, as shown in Figure 2 some examples from images.

2.png

Figure 2. Examples of ultrasound images [17]

3.2 Pre-processing

3.png

Figure 3. Training ovarian ultrasound image

The Image Archiving and Communication Systems in the hospital are used for image segmentation and analysis was used to retrieve all Ultrasound Images. The procedure is provided, spanning data preprocessing and model training in Figure 3. The raw images showed numbers from medical records and other pertinent data. We eliminated medical record information from the photos in accordance with Upgraded Local Binary Pattern (ULBP) standards to protect data confidentiality and prevent it from influencing next model training and classification. A random selection of 60% of the benign and malignant tumor photos were 40% of the datasets were used as validation datasets, and the remaining 60% as training datasets. Images from training datasets had data augmentation done, the machine learning approach used to broaden data sets used for training that are diverse, prevent approach over-fitting, and enhance accuracy [18].

Data augmentation methods used in the current investigation included random flipping, horizontal mirroring, where images were rotated by 60 degrees or less and random rotation, where images were. As a consequence of adding To obtain a number equivalent to the initial training data, white noise was applied in the training data to each image, The final training data set has 2454 total images.

3.3 Classification model using ConvNN

CNN is an example of a deep learning technique that simulates how the brain processes information. In recent years, the CNN has drawn the attention of researchers from numerous nations since it is thought to be a useful tool for enhancing automated diagnosis systems and achieving better results for illness prediction systems. A subclass of artificial neural networks called "Convolutional Neural Networks" (CNNs) create output states, or "pixels" in a visual analogy, by convolution a filter input signal that the particular bank of neurons has already learned. The CNN has a unique architecture made up of many layers that will be organized according to their capabilities. Each layer of CNN processed information, with the output from one layer feeding into the output of the next. A network's the input layer is the top layer, while the output layer is its bottom layer.The network's hidden layers are all of the layers that are present between the input and output layers. Usually each layer contains a simple algorithm and only one type of activation function.Convolution  activation  pooling  and fully connected layers are possible classifications for these layers [19].

In this paper, only a suggested one-dimensional convolutional neural network model (1D-CNN). In general, CNN model had the benefit of capturing high-order interactions between the image and the target class to make accurate predictions. In other words, the prediction model is used to predict a target class for a given instance [20]. CNN can simulate the human brain and one of the main significant benefits of CNN is to predict the diseases. Generally, the distinctive CNN is made up of two main components which are feature extractor and classifier that are clarifying as follows:

3.3.1 Feature extractor

It is the first phase of CNN works; it makes to extract features and convert them into feature maps. CNN consists of multiple filters each with a specific task, multiple feature maps resulted accordingly which are corresponding to multiple filters. The features extraction process passes through some steps to generate the final low dimensional feature vector which is finally fed into a classifier. Feature extractor consists of multiple layers (multiple convolution layers with optional pooling layers). In the first step, it passes through convolution layer to convolve the filter with the input and produces feature maps which are reduced later with pooling layer. Then, it uses the previously generated feature maps to be as input feature maps and applies the same process on them, it still goes on layer by layer and extract powerful feature and get smaller size feature maps. After that, the final powerful feature, reduced dimension feature maps are flattened to generate low dimensional feature vector to be fed into classifier [21].

3.3.2 Classifier

After feature maps are extracted and dimensions are reduced by selecting the best features between them, low dimensional feature vector fed into a classifier. The classifier returns the likelihood of class that the input may belong to it [22]. The classifier includes one or more completely connected layers to do that purpose [23].

3.4 Building prediction model

Before applying the prediction model, the reduced dataset is divided into two phases called the training set and testing set. Typically, when the dataset divides the data into a training set and a testing set, with the majority of the data utilized for training, testing is done using a smaller percentage of the data. The dataset is divided here into training (sixty percent) and testing set (forty percent). Generally, building the prediction model represents the most important step in the proposed system. In this paper, CNN model is utilized since it is straightforward to examine while also having a high level of dataset managing ability. It also has the ability to reach the highest prediction accuracy. All CNN models are of the same structure which is the best one according to trial and error and the 1D-CNN structure is the best one between the numerous structures experiment. The CNN model with carefully selected input data is applied for its observed effectiveness to offer satisfying results, and boosting prediction performance. Therefore, the proposed system was based on the CNN model for predicting samples [24]. The CNN is considered to be black box which is made to extract features according to pre-specified parameters and is proposed to predict overlain tumor for providing the desired results [24]. Figure 4 shows the structure of the CNN model used and the proposed structure of the CNN model in Table 1.

4.png

Figure 4. Structure of the used CNN model [25]

Table 1. CNN model proposed structure

In this section, we offer an empirical examination of the 1D-CNN on the ultrasound pictures dataset presented before. We describe the experimental paradigm before moving on to our discussion of the findings and results. An important but understudied topic is the accurate segmentation of ovarian cancers based on ultrasound images. The high rate of metastasis of ovarian cancer in particular results in a variety of tumor sizes, forms, and locations. Convolutional procedures were used in limited regions by conventional CNN-based tumor segmentation techniques to learn in-depth inherent properties of the input pictures, consequently, as a result of the local receptive field, restricted the prediction's accuracy. A novel graph convolutional module was developed to solve this problem.

At this stage, the CNN model is proposed and proven it’s efficient to classification ovarian tumors. The best subset of image obtained from the segmentation method which is important to get the highest possible accuracy. In general, CNN model for this paper is built using a python package called Keras. A high-level neural network toolkit called Keras was created to facilitate quick experimentation. According to high level, there is no requirement to write code in order to send output from one layer to another. All the requirements to know are the neural network architecture and input data with where the network can be trained. As a result, Keras is appropriate for research in which prototype may be quickly constructed and tested. Particularly, CNN model has been chosen and its architecture consists of sixteen layers, including three convolution layers, three activation function and three max-pooling layers for features extraction and one flatten layer. It is following by four fully-connected layers (three hidden and one output) and two dropout layers for prediction. The dropout layer is a regularization technique used to prevent overfitting in neural networks. It works by randomly dropping out (setting to zero) a certain percentage of input units during training. This forces the network to learn more robust features that are useful in making predictions on new data. The dropout rate is typically set between 0.2 and 0.5, with higher rates resulting in more regularization, which reduces the number of parameters in the model and helps prevent overfitting.

The first layer is the convolutional layer with 3×3 core size and 16 different filters. The second layer is a Rectified Linear Unit (ReLU), which is used as the nonlinear activation function, then followed by a max-pooling in layer three. A conv layer with 3×3 kernel size, 32 different filters, and ReLU activation function are implemented in the fourth and fifth layers. The maximum aggregation layer is at layer six.The seventh layer is the convolution layer with kernel size 3×3 and 64 different filters. The eighth layer is a ReLU activation function and also followed by a max-pooling in layer nine. The tenth layer is the flatten layer. Then, A completely connected layer with 256 neurons and a ReLU activation function makes up the eleventh layer. There is a layer dropdown in layer number twelve to overcome the problem of overfitting.

The 13 layer is a fully connected layer containing 512 neurons, with ReLU activation function.The fourteenth layer is the dropout layer. The fifteenth layer is a fully connected layer with 1024 neurons, and a ReLU activation function. The last fully connected layer has 3 neurons to classification for the ovarian tumors in layer number sixteen and uses a softmax layer as a prediction layer.

In addition, categorical cross entropy as the loss function to compute the loss, accuracy as a metric and the Adaptive Moment Estimation (ADAM) optimizer with learning-rate 0.0001 are selected for CNN model. This model is trained with 100 epochs. Both accuracy and loss are used to evaluate the classification model proposed. The performance of the classify model (Convolutional Neural Network) has been evaluated using the training and testing process. Generally, all the images that appeared as influencing images are taken. Then entered into the model to ensure their importance as well as the extent of accuracy that would be reached by using these images. Actually, the accuracy is dramatically increased after using the segmentation the images. Since the process of identifying the informative segment which passed through several stages and most of segmentation techniques are used. The CNN model's comparing accuracy and loss results before and after applying the segmentation the images method is illustrated in Table 2.

Table 2. The comparative results of accuracy and loss for the CNN model

5.png

Figure 5. Presents the outcomes

6.png

Figure 6. Percentage of accurately

The Figure 5 presents the outcomes. Pictures of benign and malignant tumors were both included when the threshold for the confidence score was set at 95%, at a rate of 61.71% and 58.57%, respectively. That instance, when the confidence score was adjusted to between 95 and 100%, approximately 40% of the photos among the accurately interpreted data needed manual reinterpretation; this defies the goal of employing to reduce work and time, use the auxiliary diagnostic system. Results from a ConvNN shows that the total classification performance is 98.79 percent accurate, as shown in Figure 6.