Detection of ovarian follicles cancer cells using hybrid optimization technique with deep convolutional neural network classifier

Abstract

Having one’s life threatened by a disease like ovarian cancer is the single most crucial thing in the whole world. It is difficult to achieve high performance without sacrificing computational efficiency; the results of the denoising process are not as good as they could be; the proposed models are nonconvex and involve several manually chosen parameters, which provides some leeway to boost denoising performance; the methods generally involve a complex optimisation problem in the testing stage; Here at DnCNN, we’ve developed our own version of the deep ii learning model, a discriminative learning technique. The goal was to eliminate the need for the iterative optimisation technique at the time it was being evaluated. The goal was to avoid having to go through testing altogether, thus this was done. It is highly advised to use a Deep CNN model, the efficacy of which can be evaluated by comparing it to that of more traditional filters and pre-trained DnCNN. The Deep CNN strategy has been shown to be the best solution to minimise noise when an image is destroyed by Gaussian or speckle noise with known or unknown noise levels. This is because Deep CNN uses convolutional neural networks, which are trained using data. This is because convolutional neural networks, which are the foundation of Deep CNN, are designed to learn from data and then use that learning to make predictions. Deep CNN achieves a 98.45% accuracy rate during testing, with an error rate of just 0.002%.

Keywords

Ovarian follicles cancer deep learning hybrid optimization noise levels magnetic resonance imaging

1 Introduction

After a picture has been taken, the quality of the image may be improved by using a process called image denoising. This occurs after the shot has been captured by a camera. This is essential because the process of acquiring a picture produces a variety of sounds that lower the quality of the image. These noises must be removed before the image can be used. Disruptions in the Gaussian and speckle patterns are often to blame for the degradation of medical pictures. The technique of denoising MR images of ovarian cancer using the suggested deep learning model (DeepCNN) for various forms of noise with known and unknown noise levels is discussed in this proposed study. The generation of a noisy image can be accomplished in one of three ways: by supplementing the image with Gaussian noise at a level ranging from 5 to 50; by supplementing the image with speckle noise at a level ranging from 5 to 50; or by supplementing the image with a variety of noises at a specified noise level of 15. To accomplish the goal of denoising, a number of different deep learning models and traditional filters, such as Gaussian, adaptive, bilateral, and directed filters, as well as suggested deep CNN and pretrained DnCNN, are taken into consideration. With the use of metrics like as PSNR, SSIM, MSE, and MAE, it is possible to measure and compare the effectiveness of deep learning models with that of more conventional filtering procedures. The process of detecting ovarian cancer in a timely way while maintaining a high degree of accuracy is a challenging one for a doctor to undertake. Ovarian cancers are caused by the uncontrolled growth of cells in the ovary, and magnetic resonance imaging (MRI) is the technique of detection that is used most often for ovarian cancer. Ovarian cancers may be prevented by maintaining a healthy reproductive system (MRI). Even though there have been substantial breakthroughs in semiautomatic and fully automatic image processing approaches, there are still a lot of challenges to face. These include reducing noise from MR pictures, categorizing and segmenting ovarian cancer in MR images, and so on.

1.1 Existing works

Quantitative computed tomography (CT) image characteristics were studied [12] to determine their utility in predicting tumour responses to chemotherapy in ovarian cancer patients. The best features were chosen with the use of a feature selection technique, and then a new quantitative imaging marker was developed with the help of a fusion technique that gave equal weight to each feature. Comparison was made between RECIST (Response Evaluation Criteria in Solid Tumours) and quantitative imaging indicators for their predictive power. This method has a few drawbacks, such as a lack of scalability and difficulties with dataset segmentation.

Tissue microproteomics was employed [13] to categorise the local proteome in ovarian cancer into benign, tumour, and necrotic/fibrotic tumour areas. These protein-exhibiting areas were subjected to top-down tissue microproteomics. This method is effective for identifying the roles of altprots in ovarian health and illness. Clinical uses of CA-125 using a surfactant-stabilized nanobubble ultrasound image were shown by [14]. Increased tumour formation, strong signal intensity, and delayed washout rate were all results of CA-125 surface functionalization of nano bubbles. The use of CA-125 antibody-conjugated nano bubbles has been linked to a notable rise in the detection of EOC. Targeted NB-mediate needs further optimisation to become a more sensitive technique for early stage diagnosis of ovarian cancer.

To keep tabs on the amount of endogenous-gal in lysosomes, [15] developed the unique two-photon fluorescent probe FC-gal. Experiments showed that the probe FC-gal had less cytotoxicity when imaging endogenous-gal SKOV-3 cells of human ovarian cancer. Similar to Lyso Tracker, the probe proved successful in lysosome accumulation. To fully comprehend the pathogenic roles, biological activities, and endogenous-gal abnormalities associated with the two-photon probe, more research is necessary.

[16] thoroughly compared the accuracy and sensitivity of CT and US scans for the diagnosis of peritoneal carcinomatosis (PC) of ovarian cancer (OC). To identify PC patients before to surgery, US scans must be standardised, practical, helpful, and economically preferable over CT scans. For the purpose of detecting H2O2 in ovarian cancer cells in culture, [17] developed the fluorescent probe PEP-Npb1-Cy3 model. The probe is designed using 1, 8-naphthalimide boric acid to create fluorescence signals in response to H2O2’s oxide reactivity, leading to increased sensitivity and selectivity. Based on the outcomes of exogenous and endogenous H2O2imaging in ovarian cancer cells, PEP-Npb1-Cy3 was deemed an efficient tool for biomedical applications.

In order to choose features, the current system only had access to a limited dataset. Only tumours that the radiologist had indicated as a priority for follow-up were subjected to segmentation and analysis. Not all forms of chemotherapy were included, nor were all types of histology. Ultrasound pictures may be utilised to provide better outcomes than the CT scan images that were used in this study. The data has a problem with class inequality that has to be considered. Using deep learning models like ResNet, Inception, and tree-based models, we can improve CNN-based detection and classification systems. Moreover, machine learning methods are vulnerable to having too influential training characteristics. Manual tumour area segmentation was laborious.

1.2 Significance of noise

When shooting photographs, there is a possibility that noise may be captured in the final shots. This is due to a number of factors, including the low resolution of certain cameras as well as differences [6] in the lighting circumstances. This is because there are a variety of things that may contribute to the production of noise [7]. The noise is an unwanted component that contributes to the picture being contaminated in many different ways. When it comes to the process of picture denoising, having knowledge on the various forms of noise that were present in the original image is pretty much essential to have [8]. The filter is able to reduce the amount of noise that is present in a picture while preserving the information in its entirety. The type of the noise that is present in the picture may be either additive or multiplicative in character, depending on the method that was used in the creation of the image. The equation that may be used to convey an additive noise [9] is shown further down in this passage shown in Equation 1: $w (x, y) = s (x, y) + n (x, y)$ (1)

While the multiplicative noise [8] satisfies from Equation 2. $w (x, y) = s (x, y) \times n (x, y)$ (2)

Where the original picture is denoted by s(x, y), the noise that was injected to make the corrupted image is denoted by n(x,y), and the pixel position is represented by (x,y).

1.3 Noise may take in medical pictures

The Gaussian, salt and pepper, speckle, and Poisson noises are the many kinds of noises that are carried out for the purpose of denoising, and they are addressed in this section.

1.3.1 Gaussian noise

The probability density distribution of the kind of statistical noise known as Gaussian noise [10] is mathematically similar to that of the normal distribution. Gaussian noise is one type of statistical noise. The most frequent term for normal distribution is the name “normal distribution,” although it also goes by the name “normal distribution.” In addition to this, there is the prospect of mathematically describing Gaussian noise. Below is an equation that may be used to determine the probability distribution function of the normal distribution, which is denoted by the notation f(x). This equation can be found in Equation 3 [11]. $f (α) = e 2 σ^{2} \sqrt{2} 2 π σ^{2}$ (3)

Where, μ is the mean or expectation of the distribution, σ is the standard deviation, σ² is the variance.

1.4 Salt and pepper noise

In many cases, images that include salt and pepper noise feature pixels that are light in the dark areas and pixels that are dark in the light areas. The presence of both black and white dots in the photographs may be traced back to the noise that was present [12]. The rapid and unanticipated shifts in the visual signal, dead pixels, faults in the analog-to-digital converter, bit mistakes in transmission, and a range of other issues are the root causes of this noise [13]. Dark Frame Subtraction (DFS) and generating new data points around dark and bright pixels, both of which are created by either the median filter or the morphological filter, are two methods that can be used to get rid of this type of noise [14]. Both of these techniques are created by either the median filter or the morphological filter.

Cancer research is often either biological or clinical, with statistical research providing a counterpoint. The most alluring and difficult goal is illness prediction, which calls for solid data mining approaches. The field of medical data mining is expanding rapidly at the moment, with thousands of records being recorded every second. In general, it is challenging to acquire statistics on gynaecological cancers. As a result, this study’s primary objective is to provide a method for detecting and preventing illness at an earlier stage. Doctors may be able to rescue more people if they use this strategy to treat this kind of cancer. To improve ovarian cancer classification and diagnosis with more accuracy than current methodologies, the proposed study is being conducted on several ovarian cancer datasets utilising data mining technologies.

The innovation here is in the utilization of environmental variables and genetic identifiers to construct classification system approaches for preventing diseases. Ovarian cancer is highly preventable and treatable if caught and classified early disease. To present an algorithm for effective cancer diagnosis utilising a machine learning technique, employing unique annotated pictures of ovarian tissue. The goal of this study is to assess the ability to detect gynecological cancers in people. Improved classification performance has been shown using a deep neural network to address issues with examining the texture of ovarian cancer cells. The use of feature extraction and deep learning techniques enhances the accuracy and identification of ovarian cell classification. The organization of the paper is as follows; section 2 includes materials and methodology of proposed work, section 3 includes experimental analysis, section 4 includes conclusion and future work.

2 Materials and methods

Filtering methods are often the most effective method available to choose from when attempting to minimize the amount of noise in an image. In this section, we discuss the use of traditional filters as a means to reduce the volume of a wide range of sounds. Denoising is a crucial step in the preprocessing phase, which is used to improve the image’s quality as a whole and is one of the ways this is accomplished. This section makes use of a variety of different kinds of denoising filters, including Gaussian, adaptive, bilateral, and guided filters.

Denoising methods, such as adaptive and Gaussian filters, will provide an effect that smoothest out the harsher corners of the noise [17]. Research has shown that bilateral filtering is the nonlinear smoothing filter that is easiest to understand. It has the effect of inverting the gradient, and it estimates the weight by using an approximation based on a histogram. In spite of the fact that it has a high degree of computational complexity, an improved bilateral filter-based framework that is capable of efficiently decreasing universal noise was developed [18]. This was accomplished despite the fact that it was somewhat labour intensive. Additionally, the filter is able to achieve these results. One of the disadvantages of using a bilateral filter is the difficulty that comes with having the gradient in the other direction. Before processing the picture, the process of denoising the image involves applying a broad range of filters to it and playing a wide variety of sounds. Denoising an image may be accomplished by the use of a wide variety of filters, some of which are shown in Fig. 1. These filters include an adaptive filter, a Gaussian filter, a bilateral filter, and a directed Filter.

Fig. 1

Block diagram illustrating the process of picture denoising using a variety of conventional filters.

Figure 1 shows a block diagram that depicts the process of executing a single ovarian tumours image. The picture that is now being read in is a random-sized RGB image that has three channels, and its dimensions are 705 pixels wide by 981 pixels height. During the stage of preparation, it is also changed into a grayscale image that has dimensions of 256 pixels by 256 pixels and a single channel. In the first framework, which can be seen in Fig. 1, the noisy image is given a filtering treatment. During this process of filtering, a number of different filters are used, including an adaptive filter, a Gaussian filter, a bilateral filter, and a guided filter, among others.

2.1 The underlying structure of the pretrained DnCNN model

There are a multitude of pretrained convolutional neural networks available for usage, and they may be used to regression issues as well as images. In order to complete the task of image denoising, the DnCNN, which was created by [19], is a model that has been pretrained with 20 convolutional layers of depth. It has a total of 59 layers, which consists of one input layer, twenty convolutional layers, nineteen ReLU layers, eighteen batch normalization levels, and one regression output layer. It also has one batch normalization level. The graphical depiction of the DnCNN model’s internal structure is shown in Fig. 2, which can be found here.

Fig. 2

Architecture of DnCNN model.

The pooling layers are not used in this specific variant of the architecture that is being discussed. Figure 2 demonstrates that the model is made up of three unique types of layers. These may be observed in the model. Following the input layer is what the neural network refers to as the “Conv+ReLU,” which is the name of the first convolutional layer in the network (patch size, patch size, channels). The activation layer, which in this specific architecture is made up of ReLU, is placed in the position of precedence following the convolutional layer. Each convolutional layer employs filters with dimensions of 3 pixels by 3 pixels and channels of size 64 filters in order to construct a total of 64 feature maps. In this scenario, the number of channels is equal to one for a grayscale image but increases to three for an RGB picture. The Conv+BN+ReLU layer is constructed from of its individual components, which include the convolutional layer, the batch normalization layer, and the ReLU layer. In the model that has been shown, this combination is used 18 times between the first Conv+ReLU layer and the last convolutional layer, which is referred to as Conv20. After the Convolutional layer and before the ReLU layer, a batch normalising layer is added, and sixty-four filters with a size of three by three by sixty-four are used in it. The convolutional is the last layer of the processing that involves convolution, and in order to reproduce the output, a collection of filters with dimensions of 3 3 64 is used. The mean squared error is computed with the help of the regression layer, which is also the component that acts as the model’s finishing touch.

2.2 Image denoising with DnCNN

The second framework offers detailed information on the use of DnCNN for the purpose of image denoising. In the second architecture, shown in Fig. 3, denoising is carried out with the assistance of one of the pretrained convolutional neural networks (DnCNN). The DnCNN has undergone rigorous training with the ImageNet database, which consists of millions of photographs taken in natural settings. After loading well-optimized weights into the DnCNN network or model, proceed to the next step of minimizing noise by passing the DnCNN network along with a noisy 2D image. The computation of performance metrics begins with the denoised image as the starting point. Utilizing a pre-trained DnCNN network allows for the removal of noise from an image.

Fig. 3

Framework of image denoising with proposed Deep CNN.

The built-in pretrained denoising neural network, commonly known as DnCNN, is the quickest, easiest, and most promptly evaluated method for noise reduction. A built-in pretrained denoising neural network is the moniker for this technique. Since the relevant network or model was not trained on the BraTS dataset, it is important to remember that the pretrained network does not provide a great deal of leeway when dealing with known sources of noise. This is an issue that needs to be considered.

To eliminate Gaussian noise without the time-consuming and error-prone procedure of training a network, the pretrained DnCNN network is used. Using the pretrained network to eliminate noise is fraught with complications, such as the ones listed below. The network can only identify Gaussian noise, which has a constrained standard deviation, hence noise reduction is only applicable to two-dimensional pictures with a single channel. It is possible to reduce the amount of noise in a three-dimensional picture with many colour channels or planes by focusing on each one separately. To overcome the limitations of this framework, to obtain more flexibility, and to enhance the denoising performance, it is required to train the network with the mentioned layers and a tailored denoising neural network. The network may be trained to achieve these aims.

2.3 Denoising of an MR image of an ovarian tumour using a suggested DeepCNN model

In this section, we will discuss the proposed deep learning model, which is called DeepCNN, as well as the process of image denoising that makes use of the proposed model.

2.4 The basic structure of the suggested machine learning technique for deep learning (DeepCNN)

The architecture of the DeepCNN model that has been suggested is precisely the same as the architecture of the DnCNN model. However, it is a 17 convolutional layers depth model and has a total of 50 layers, comprising one input layer, 16 ReLU layers, 15 batch normalization layers, and one regression output layer. Additionally, it contains 17 convolutional layers. In addition to that, there is one regression output layer included inside it. In DnCNN, an image with a patch size of 50 x 50 is used for the input layer, however in DeepCNN, an image with a patch size of 61 x 61 may be used for the input layer. DnCNN’s input layer uses an image with a patch size of 50 x 50. Deep CNN’s input layer uses an image with a patch size of 61 x In addition to this, it is capable of producing first-rate outcomes in a wide range of traditional image noise reduction applications, which is a considerable benefit.

2.4.1 Performance evaluation metrics

Metrics like as PSNR in dB, SSIM, MSE, and MAE may be used to provide an analysis of the success of the solutions that have been discussed. PSNR: The peak signal-to-noise ratio, also known as the peak signal-to-noise ratio (PSNR), of an image reflects the amount of noise immunity that the image has. A high PSNR score indicates that there is less interference produced by noise in the MRI imaging of the ovary. This may be inferred from the fact that the value is positive. The following term may be used to characterize the PSNR in the vast majority of situations.

Where maxi refers to the pixel in the input ovarian tumour MR picture that has the greatest value, and MSE refers to the average squared error in the image. The usual values for the peak signal-to-noise ratio (PSNR) in a noisy picture and the original image are between 30 and 50 dB, assuming that the bit depth is 8 bits. When looking at a picture, a greater number suggests a better overall quality. When working with 16-bit data, the PSNR values will often lie within a range that spans from 60 to 80 dB.

The structural similarity index method (SSIM) is a technique that may be used to speculate on the perceived characteristic of digital TV and photographic pictures, in addition to a wide variety of other forms of digital images and videos. It is closer to the human level of comprehension due to the fact that it displays the similarity in the structural information of image pairs is shown in Fig. 3. The computation of a P test picture in respect to a Q reference image is carried out so that it may be determined how visually comparable two different pictures are to one another. The two photos, P and Q, that will be compared and contrasted are: p= “pi|i=1,2,...N” and q= “qi|i=1,2,...N” are the notations that are used to represent the local square pairs of P and Q, respectively shown in Equations 7. $I (p, q) = (2 μ_{p} μ_{q} + C 1) (μ_{p}^{2} + μ_{q}^{2} + C 1)$ (4) $c (p, q) = (2 σ_{1} σ_{q} + C_{2}) (σ_{p}^{2} + σ_{q}^{2} + C_{2})$ (5) $I (p, q) = (σ_{p,} + C 3) (σ_{p} σ_{y} + C 3)$ (6)

Where, the constants C1, C2, and C3 are considered to obstacle. When (μ²+μ²), (σ_p²+σ_q²) or σ_p σ_q is set to zero. The l (p,q) index stands for the luminance differences, c(p,q) stands for contrast differences, and r (p,q) stands for the structural variations between x and y.

Now the SSIM index is stated as shown in Equation 7 $SSIM (\underline{μ}, = [1 (p, Θ)]^{α} \cdot [c (p, g)]^{β} \cdot [r (\underline{μ}, Θ)]^{γ}$ (7)

Where, p and q are the parameters that characterize the comparative significance of each component and SSIM (p,q) is a number that may vary from 0 (totally diversified) to 1 (unbalanced), where 0 indicates that the component is entirely diversified and 1 indicates that it is unbalanced (identical patches).

If the value of SSIM is 1, it implies that the reconstructed image is structurally comparable to the original one. SSIM values may range from 0 to 1, and if the value is 1, it shows that SSIM is equal to 1. In general, good quality reconstruction methods will have SSIM values of 0.97, 0.98, or 0.99. This is because these values correspond to the following:

MSE, which stands for “mean squared error,” is an abbreviation that refers to a measurement used to determine the overall quality of the picture. It is feasible to explain it by saying that it is the cumulative squared error value between the input photos R (a, b) and the input images of the segmented image. This explanation is correct.

Where “m” refers to the number of rows, and “n” refers to the number of columns that were included in the photo that was submitted. The value of the PSNR compared to the value of the MAE has an inversely proportional relationship.

The mean-absolute error (MAE), the mean-square error (MSE), and the peak signal-to-noise ratio (PSNR) are three relevant metrics to look at when assessing the quality of image denoise. The MSE is the cumulative squared error that occurs between the denoised version of the image and the original version of the image, while the MAE is the absolute error that exists between the denoised version of the image and the original version of the image. When the value of MSE is decreased, the number of mistakes made likewise falls to a lower level. Calculated by employing the peak signal-to-noise ratio (PSNR), the peak signal-to-noise ratio (PSNR) of two images is expressed in decibels. The quality of the original picture compared to the denoised version of the image may be evaluated using this ratio as a measuring stick. A higher quality image will result from either denoising or rebuilding it, and this improvement will be proportional to the PSNR (picture signal to noise ratio).

3 Experimental results and discussions

In this section, the images that will later be cleaned up by denoising are deformed in one of three distinct ways. The image is given Gaussian noise throughout a wide range of noise levels, starting at 5 and going all the way up to 50 as the first phase in the process. Second, the image is destroyed because of the superimposition of speckle noise with a wide range of noise levels, which may be anywhere from 5 to 50. Third, the image was degraded by the addition of a range of sounds that had a specified noise level of 15. These noises included Gaussian noise, salt and pepper noise, Poisson noise, and speckle noise. In order to accomplish the experimental findings of each of these cases, traditional filters, DnCNN, and the recommended DeepCNN model were used.

Figure 4a 4b illustrates the process of removing noise from MR images of ovarian tumours after the pictures had been distorted by Gaussian noise with a noise level ranging from 5 to 50. Original photos and noisy images are shown in Fig. 4 respectively. Beginning with Sl. No. 1 and continuing through Sl. No. 10 in Table 1. In this table denoised images that were obtained by using a variety of models and filters, such as the pretrained DnCNN model, the proposed DeepCNN model, the Gaussian filter, the bilateral filter, the adaptive filter, and the guided filter.

Fig. 4

(a) Utilizing a pre-trained DnCNN network allows for the removal of noise from an image.

Fig. 4

(b) Process of removing MR image of ovarian tumours.

Table 1

Presents the performance assessment metrics of the proposed DeepCNN, the pretrained DnCNN, the Gaussian filter, the adaptive filter, the bidirectional filter, and the guided filter with Gaussian noise ranging from level 5 to level 50

Sl. No	Noise level	Method	PSNR (dB)	SSIM	MSE	MAE
1	σ = 5	DeepCNN	28.9274	0.2385	8242	7.9241
		DnCNN	25.3215	0.1931	192.5861	12.122
		Gaussian Filter	22.0212	0.1021	408.2814	11201
		Adaptive Filter	25891	0.121	284.5587	12.9042
		Guided Filter	22.5777	0.0948	359.1772	13766
		Bilateral Filter	19.376	0.0702	750.7282	14.0564
2	σ = 10	DeepCNN	25.3826	0.1978	192.6814	12.112
		DnCNN	22.4189	0.1631	375.0291	16.8145
		Gaussian Filter	19.0666	0.0711	806.1604	18.3615
		Adaptive Filter	20.6993	0.0842	555446	18.0257
		Guided Filter	18.7139	0.0566	874.3521	18.9955
		Bilateral Filter	16.4364	0.0466	1.48E+03	19.6528
3	σ = 15	DeepCNN	22.39	0.1587	375.0389	16.7951
		DnCNN	20.6479	0.1391	559.042	20.6937
		Gaussian Filter	17.3626	0.0581	1.19E+03	22.6497
		Adaptive Filter	19.0365	0.0698	811.7672	22.2023
		Guided Filter	16.5214	0.0432	1.45E+03	25897
		Bilateral Filter	14.7473	0.0372	2.18E+03	24.2206
4	σ = 20	DeepCNN	20.6564	0.1412	559.042	20.6937
		DnCNN	19.439	0.119	738.3465	26042
		Gaussian Filter	16.2135	0.0487	1.56E+03	25.9013
		Adaptive Filter	17.9776	0.0605	1.04E+03	25.2583
		Guided Filter	15.1004	0.0343	2.01E+03	27.0262
		Bilateral Filter	16152	0.03	2.83E+03	27.6357
5	σ = 25	DeepCNN	19.448	0.125	738.3745	27842
		DnCNN	18.5059	0.1099	911.2204	26.3927
		Gaussian Filter	15.3035	0.0445	1.92E+03	28.8271
		Adaptive Filter	17.1165	0.0554	1.26E+03	28.0908

The performance metrics of suggested DeepCNN and DnCNN techniques, in addition to those of other filtering methods, are shown in Table 1, along with variable degrees of Gaussian noise for denoising. It should come as no surprise that the PSNR and SSIM values would drop as the amount of noise in the signal increases. Deep CNN’s de-noising performance measures, such as PSNR in dB, SSIM, MSE, and MAE, are as follows: 20.6564, 0.1412, 559.042, and 20.6937 correspondingly, when applied to a single instant with a noise level sigma equal to 20. According to Table 1, the proposed DeepCNN model demonstrates improved performance when contrasted with DnCNN as well as all filtering techniques with an unknown noise level of Gaussian noise. Therefore, Deep CNN seems to be the most appropriate way for decreasing noise when the picture is ruined by a known or unknown degree of Gaussian noise. This is because Deep CNN uses convolutional neural networks.

Figure 5 depicts the fluctuation of the peak signal-to-noise ratio (PSNR) with varying noise levels from 5 to 50 for a variety of filters and models, including the Gaussian Filter (GAF), the Adaptive Filter (AF), the Guided Filter (GUF), and the Bilateral Filter (BF) (pretrained DnCNN and proposed DeepCNN). When compared with other filters and models, the suggested DeepCNN has a PSNR that is maximum and declines with rising noise level. This effect is caused by the addition of Gaussian noise.

Fig. 5

Plot of PSNR values in dB versus Gaussian noise level for various denoising methods.

Figure 6 depicts, how the SSIM changes depending on the noise level, which ranges from 0 to 50. This is done for a variety of filters and models, including the Gaussian Filter (GAF), the Bilateral Filter (BF), the Adaptive Filter (AF), and the Guided Filter (GUF) (pretrained DnCNN and proposed DeepCNN). When compared with other filters and models, the suggested DeepCNN has a superior SSIM, which is 0.2385, and its performance worsens as the degree of noise becomes higher. Gaussian noise was used in this comparison.

Fig. 6

Plot of SSIM values versus Gaussian noise level for various denoising methods.

Figure 7 depicts the fluctuation in MSE for a variety of filters and models, including the Gaussian Filter (GAF), the Bilateral Filter (BF), the Adaptive Filter (AF), and the Guided Filter (GUF). The noise level ranges from 0 to 50. (pre-trained DnCNN and proposed DeepCNN). When compared with other filters and models, the suggested DeepCNN has the lowest MSE value possible, and its performance improves along with an increase in the amount of Gaussian noise used.

Fig. 7

Plot of SSIM values versus Gaussian noise level for various denoising methods.

Figure 8 depicts the fluctuation of the Mean Absolute Error (MAE) with varying levels of noise sigma, ranging from 5 to 50, using a variety of filters and models, including the Gaussian Filter (GAF), the Bilateral Filter (BF), the Adaptive Filter (AF), and the Guided Filter (GUF) (pretrained DnCNN and proposed DeepCNN). When compared with other filters and models, the suggested DeepCNN has an MAE value that is the lowest possible and grows as the amount of noise becomes higher. This effect is caused by the addition of Gaussian noise.

Fig. 8

Plot of MSE values versus Gaussian noise level for various denoising methods.

The performance metrics that were achieved by the proposed DeepCNN, the pretrained DnCNN model, and other filtering approaches are shown in Table 2, along with the various noise levels that were used for the denoising job. It is undeniable that the PSNR and SSIM values improve along with a reduction in the amount of noise content. Denoising performance measures of proposed DeepCNN such as PSNR in dB, SSIM, MSE, and MAE are 39.4316, 0.9705, 7.4118, and 0.5178 accordingly for a single moment with a noise level equal to 10. Table 2 makes it clear that the proposed DeepCNN model has greater performance when compared with pretrained DnCNN and all filtering techniques with an unknown noise level of speckle noise. This can be seen by examining the results of these comparisons. It is possible to draw the conclusion that the DeepCNN approach that was developed is a more appropriate way for decreasing noise when the picture is ruined by an unknown noise level of speckle noise than when the noise was caused by Gaussian noise.

Table 2

Performance evaluation metrics of proposed DeepCNN, pretrained DnCNN, Gaussian, adaptive, bilateral, and guided filters with speckle noise from various noise level 5 to 50

Sl. No.	Noise level	Method	PSNR (dB)	SSIM	MSE	MAE
1	σ = 5	DnCNN	36.625	0.9547	14.1444	0.556
		DeepCNN	39.4316	0.9705	7.4118	0.5178
		Gaussian Filter	36.6841	0.9544	1953	0.6766
		Bilateral Filter	37.1663	0.9667	12.4868	0.5652
		Adaptive Filter	35.5314	0.9557	18.1945	0.7181
		Guided Filter	35.7749	0.9454	17.2023	0.7506
2	σ = 10	DnCNN	34.4705	0.9368	22292	0.7523
		DeepCNN	37.6716	0.9611	11.1152	0.6611
		Gaussian Filter	3839	0.9298	26.8644	0.9738
		Bilateral Filter	34552	0.9405	29.3469	0.9022
		Adaptive Filter	32.2648	0.9258	38.6014	1.0963
		Guided Filter	38835	0.9316	26.5905	0.9357
3	σ = 15	DnCNN	32209	0.9256	30.9734	0.7869
		DeepCNN	36.5237	0.9568	14.4782	0.6599
		Gaussian Filter	32.2283	0.9154	38.9267	1.0902
		Bilateral Filter	31.3022	0.9189	48.1795	1.1019
		Adaptive Filter	30.4614	0.9064	58.4714	1.2945
		Guided Filter	32.2292	0.916	38.9189	1.046
4	σ = 20	DnCNN	32.1186	0.9131	39.9225	0.9878
		DeepCNN	35.7782	0.9512	17.1894	0.7507
		Gaussian Filter	31.0491	0.9038	51.071	1.2896
		Bilateral Filter	29.5764	0.9	71.6874	1.418
		Adaptive Filter	29.1363	0.892	79.3326	1.5609
		Guided Filter	30.8004	0.9001	54.08	1.2768
5	σ = 25	DnCNN	31.3743	0.9066	47.386	0.9893
		DeepCNN	35.2188	0.9476	19.5525	0.7425
		Gaussian Filter	30.1705	0.8946	62.5216	1.3831
		Bilateral Filter	28.3862	0.8861	94.2891	1.6035
		Adaptive Filter	28.2123	0.8815	98.1417	1.6957
		Guided Filter	29.725	0.8866	69.2762	1.413
6	σ = 30	DnCNN	30.7113	0.8992	55.2009	1.1387
		DeepCNN	34.5764	0.9436	22.6695	0.8119
		Gaussian Filter	29.4205	0.8871	74.3079	1.5296
		Bilateral Filter	27.4065	0.8756	118.1493	1.8488
		Adaptive Filter	27.4384	0.8735	117.2847	1.8869
		Guided Filter	28.7844	0.8761	86.028	1.6051

Figure 9 illustrates the change in PSNR that occurs when the noise level is changed from 5 to 50 using a variety of filters and models, including the Gaussian Filter (GAF), the Bilateral Filter (BF), the Adaptive Filter (AF), and the Guided Filter (GUF). When compared with alternative filters and models for the addition of speckle noise, Fig. 10 demonstrates that the proposed DeepCNN yields the greatest PSNR in dB of 39.4316 and declines with increasing noise intensity.

Fig. 9

Plot of PSNR values versus speckle noise level for various denoising methods.

Fig. 10

Plot of SSIM values versus speckle noise level for various denoising methods.

Figure 10 illustrates how the SSIM changes depending on the noise level, which ranges from 0 to 50. This is done for a variety of filters and models, including the Gaussian Filter (GAF), the Bilateral Filter (BF), the Adaptive Filter (AF), and the Guided Filter (GUF) (pretrained DnCNN and proposed Deep CNN. When compared with alternative filters and models for the addition of speckle noise, Fig. 12 demonstrates that the Deep CNN provided by the authors produces a superior SSIM value, which is 0.9705, and that this value falls as the noise level rises.

Figure 11 illustrates how the mean squared error (MSE) varies throughout a range of noise levels from 5 to 50 for a variety of filters and models, including the Gaussian Filter (GAF), the Bilateral Filter (BF), the Adaptive Filter (AF), and the Guided Filter (GUF) (pretrained DnCNN and proposed DeepCNN). In contrast to the previous filters and models for the addition of speckle noise, the suggested DeepCNN has a minimum MSE value of 7.4118, and this value goes up as the noise level becomes higher. This is something that can be seen from the figure that was just shown

Fig. 11

Plot of MSE values versus speckle noise level for various denoising Methods.

Fig. 12

Plot of MAE values versus speckle noise level for various denoising Methods.

Figure 12 depicts the fluctuation of the Mean Absolute Error (MAE) with varying levels of noise sigma, ranging from 5 to 50, using a variety of filters and models, including the Gaussian Filter (GAF), the Bilateral Filter (BF), the Adaptive Filter (AF), and the Guided Filter (GUF) (pretrained DnCNN and proposed DeepCNN). When compared with several different filters and models for the addition of speckle noise, it is clear from the illustration that the suggested DeepCNN produces the lowest MAE value, which is 0.5178, and that this value grows when the sigma parameter is increased. Consider the third scenario, in which the picture is ruined by the addition of a variety of sounds with a noise level being provided. In this particular instance, the noisy picture is produced by independently applying several types of noises, such as Gaussian noise, salt and pepper noise, Poisson noise, and speckle noise, each with their own set of defined parameters. Figure 13 illustrates the process of removing noise from MR images of ovarian tumours when such images have been affected by known Gaussian noise at level 15.

Fig. 13

Ovarian tumor MR image denoising with Gaussian noise of level 15.

Figure 14(a), 14(b), 14(c) shows how MR images of ovarian tumours, which had been polluted by a salt-and-pepper noise level of 15, were cleaned up. Figure 14a and 14b show the unedited pictures and the noise-enhanced versions of the same image, respectively. Denoised images obtained using several models and filters, including the pretrained DnCNN model, the recommended Deep CNN model, the Gaussian filter, the bilateral filter, the adaptive filter, and the guided filter, are shown in Fig. 16(c–h). In terms of PSNR and SSIM performance, the proposed Deep CNN is superior to all existing filters and pretrained DnCNNs up to salt and pepper noise level 15. Figure 16 shows that the recommended Deep CNN is not effective at reducing noise when the image is corrupted by salt and pepper noise with a noise level greater than 15.

Fig. 14

Ovarian tumor MR image denoising with salt and pepper noiselevel15. a) Original image b) Noisy image c) Pretrained DnCNN denoised images.

Noise reduction in MR images of ovarian tumours is shown in Fig. 15(a), 15(b) for an image corrupted by Poisson noise at the level of 15. In Fig. 15(c) and 15(d), we see the unaltered photographs and the noisy versions, respectively. Different models and filters were used to acquire the denoised images in Fig. 15(e–h), including the pretrained DnCNN model, the recommended Deep CNN model, the Gaussian filter, the bilateral filter, the adaptive filter, and the guided filter. The proposed Deep CNN outperforms the state-of-the-art DnCNN and all other filters in terms of peak signal-to-noise ratio (PSNR) and signal-to-noise ratio (SSIM).

Fig. 15

Ovarian tumor MR image denoising with Poisson noise level of 15. a) Original image, b) Noisy image, c) Pretrained DnCNN denoised images, d) Proposed Deep CNN denoised images, e) Gaussian filter denoised images, f) Bilateral filter denoised images, g) Adaptive filter denoised images, h) Guided filter denoised images.

Fig. 16

Ovarian tumor MR image denoising with speckle noise level of 15.

Figure 15(a–h) illustrates the process of removing noise from MR images of ovarian tumours when the pictures have been damaged by a known speckle noise level of 15. Table 3 provides the findings of several filters applied to the combination of a number of distinct sounds with a noise level of 15. The DeepCNN model that was suggested offers improved PSNR and SSIM values for images that are distorted by known levels of a variety of noises. These values apply to images that are noisy. The most impressive findings are shown in bold in the table below. When the picture is ruined by salt and pepper noise with a noise level of 15 or higher, the DeepCNN is not an appropriate tool for noise reduction. Figure 16 shows the feature extracted from the candidate of the follicle in the ovarian cell.

Table 3

PSNR (dB)/SSIM of proposed DeepCNN, pretrained DnCNN, Gaussian, bilateral, adaptive, and guided filters for various noises with noise level 15

S1. No	Noise	Gaussian		Poisson		Speckle		Salt and pepper
	type filters	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
1	DeepCNN	22.39	0.1587	40.3205	0.98	39.4316	0.9705	19.6042	0.2957
2	DnCNN	20.6564	0.1412	37.9084	0.9707	38085	0.9493	19.5151	0.2944
3	Gaussian Filter	17.3626	0.0581	36.6818	0.9593	28.9099	0.8918	19.2175	0.3074
4	Bilateral Filter	15.9876	0.0391	32.6595	0.9504	26.2759	0.8589	15.8919	0.2215
5	Adaptive Filter	19.0365	0.0698	37.4753	0.9618	27.2489	0.8783	17.7985	0.2483
6	Guided Filter	16.5214	0.0432	34.1424	0.9353	28.3088	0.8758	15.8473	0.2329
7	Median Filter	22.2424	0.1414	38.4352	0.9684	32793	0.9258	39.5647	0.9806

Figure 17 is a bar chart that displays the variation of PSNR when different noises such as Gaussian noise, salt and pepper noise, Poisson noise, and speckle noise are added to the models (proposed DeepCNN and pretrained DnCNN) and four different filters (Gaussian filter (GAF), Bilateral filter (BF), Adaptive filter (AF), and Guided filter (GUF)). It has been discovered that the suggested DeepCNN model produces PSNR values in dB that are superior to those produced by pretrained DnCNN and other filters. These values include 40.3205, 39.4316, 22.39, and 19.6042.

Fig. 17

PSNR performance comparison of various denoising methods.

Figure 18 is a bar chart that illustrates the variance of the SSIM. The figure is located in the middle of the page. The models (proposed DeepCNN and pretrained DnCNN) and four distinct filters (Gaussian filter (GAF), Bilateral filter (BF), Adaptive filter (AF), and Guided filter (GUF)) are presented, as well as the addition of a variety of noises including Gaussian noise, salt and pepper noise, Poisson noise, and speckle noise. In addition, the models are shown with the addition of various noises. When contrasted with the pretrained DnCNN and other filters, the SSIM values that the proposed DeepCNN model generates are 0.98, 0.9705, 0.2957, and 0.1587, which are superior to those generated by the other filters. These results were obtained using the suggested DeepCNN model. Figure 19 shows the PSNR in dB performance comparison of proposed Deep CNN for denoising of image corrupted with either Gaussian or speckle noise.

Fig. 18

SSIM performance comparison of various denoising methods.

Fig. 19

PSNR (dB) performance comparison of proposed Deep CNN for Gaussian and speckle noise.

5 Conclusion

The PSNR performance of a Gaussian noise damaged image is improved by the proposed Deep CNN model to 14.24% at a noise level of 5. The proposed model improves performance by 7.66 percentage points while dealing with speckle noise that also includes noise at the level 5. When the Deep CNN model is applied to Gaussian noise with a known noise level of 15, the PSNR improves by 8.39 percentage points. As a consequence, the proposed Deep CNN model seems to be better at reducing noise, whether the noise in the image is speckle noise, Gaussian noise, or a mixture of the two with either known or unknown levels. Furthermore, the created Deep CNN algorithm stores the SSIM values that are produced to recall the structural information of the picture. Our next suggested study will focus on ovarian cancer MR image categorization through transfer learning and hybrid models. These images might be benign LGG or very suspicious HGG.

References

Mahajan

, Mittal

and Pandit

A.K.

, Image segmentation approach based on adaptive flower pollination algorithm and type II fuzzy entropy, Multimedia Tools and Applications 82(6) (2023), 8537–8559.

Lucy

M.C.

, Staples

C.R.

, Michel

F.M.

and Thatcher

W.W.

, Energy balance and size and number of ovarian follicles detected by ultrasonography in early postpartum dairy cows, Journal of Dairy Science 74(2) (1991), 473–482.

Visintin

, Feng

, Longton

, Ward

D.C.

, Alvero

A.B.

, Lai

,... and Mor

, Diagnostic markers for early detection of ovarian cancer, Clinical Cancer Research 14(4) (2008), 1065–1072.

Badgwell

and Bast

R.C.

Jr , Early detection of ovarian cancer, Disease Markers 23(5-6) (2007), 397–410.

Ong

S.Q.

, Nair

, Dabbagh

R.D.A.

, Aminuddin

N.F.

, Sumari

, Abualigah

,... and Elminaam

D.S.A.

, Comparison of pre-trained and convolutional neural networks for classification of jackfruit artocarpus integer and artocarpus heterophyllus, In Classification Applications with Deep Learning and Machine Learning Technologies (pp. 129–141), Cham: Springer International Publishing, (2022).

Conrads

T.P.

, Fusaro

V.A.

, Ross

, Johann

, Rajapakse

, Hitt

B.A.

,... and Veenstra

T.D.

, High-resolution serum proteomic features for ovarian cancer detection, Endocrine-Related Cancer 11(2) (2004), 163–178.

Gyorffy

, Discovery and ranking of the most robust prognostic biomarkers in serous ovarian cancer, Geroscience (2023), 1–10.

Havrilesky

L.J.

, Whitehead

C.M.

, Rubatt

J.M.

, Cheek

R.L.

, Groelke

, He

,... and Berchuck

, Evaluation of biomarker panels for early stage ovarian cancer detection and monitoring for disease recurrence, Gynecologic Oncology 110(3) (2008), 374–382.

Ong

S.Q.

, Comparison of Pre-trained and Convolutional Neural Networks for Classification of Jackfruit Artocarpus integer and Artocarpus heterophyllus, In: L. Abualigah, (eds) Classification Applications with Deep Learning and Machine Learning Technologies, Studies in Computational Intelligence, vol. 1071. Springer, Cham, (2023). https://doi.org/10.1007/978-3-031-17576-3_6

10.

, Graybill

W.S.

and Zhu

, Detection and monitoring of ovarian cancer, Clinica Chimica Acta 415 (2013), 341–345.

11.

, Graybill

W.S.

and Zhu

, Detection and monitoring of ovarian cancer, Clinica Chimica Acta 415 (2013), 341–345.

12.

Yurkovetsky

, Skates

, Lomakin

, Nolen

, Pulsipher

, Modugno

,... and Lokshin

A.E.

, Development of a multimarker assay for early detection of ovarian cancer, Journal of Clinical Oncology 28(13) (2010), 2159.

13.

Deb

, Uddin

and Chakraborty

, miRNAs and ovarian cancer: An overview, Journal of Cellular Physiology 233(5) (2018), 3846–3854.

14.

Shaaban

and Rezvani

, Ovarian cancer: detection and radiologic staging, Topics in Magnetic Resonance Imaging 21(4) (2010), 247–259.

15.

Elias

K.M.

, Guo

and Bast

R.C.

, Early detection of ovarian cancer, Hematology/Oncology Clinics 32(6) (2018), 903–914.

16.

D.J.

, Chui

M.H.

, Vanderbilt

C.M.

, Jung

, Robson

M.E.

, Park

C.S.

,... and Fuchs

T.J.

, Deep Interactive Learning-based ovarian cancer segmentation of H&E-stained whole slide images to study morphological patterns of BRCA mutation, Journal of Pathology Informatics 14 (2023), 100160.

17.

Sharma

, Chakraborty

, Singh

D.P.

, Mahajan

and Pandit

A.K.

, Image-based automatic segmentation of leaf using clustering algorithm, International Journal of Nanotechnology 19(6-11) (2022), 539–553.

18.

Singh

, Tiwari

, Band

S.S.

, Rehman

A.U.

, Mahajan

, Ding

,... and Pandit

A.K.

, Impact of quarantine on fractional order dynamical model of Covid-19, Computers in Biology and Medicine 151 (2022), 106266.

19.

Rein

B.J.

, Gupta

, Dada

, Safi

, Michener

and Agarwal

, Potential markers for detection and monitoring of ovarian cancer, Journal of Oncology 2011 (2011).