FCM-NPOA: A hybrid Fuzzy C-means clustering with nomadic people optimizer for ovarian cancer detection

Abstract

Ovarian cancer is a highly prevalent cancer among women; However, it remains difficult to find effective pharmacological solutions to treat this deadly disease. However, early detection can significantly increase life expectancy. To address this issue, a predictive model for early diagnosis of ovarian cancer was developed by applying statistical techniques and machine learning models to clinical data from 349 patients. A hybrid evolutionary deep learning model was proposed by integrating genetic and histopathological imaging modalities within a multimodal fusion framework. Machine learning pipelines have been built using feature selection and dilution approaches to identify the most relevant genes for disease classification. A comparison was performed between the UNeT and transformer models for semantic segmentation, leading to the development of an optimized fuzzy C-means clustering algorithm (FCM-NPOA-PM-UI) for the classification of gynecological abdominopelvic tumors. Performing better than individual classifiers and other machine learning methods, the suggested ensemble model achieved an average accuracy of 98.96%, precision of 97.44%, and F1 score of 98.7%. With average Dice scores of 0.98 and 0.97 for positive tumors and 0.99 and 0.98 for malignant tumors, the Transformer model performed better in segmentation than the UNeT model. Additionally, we observed a 92.8% increase in accuracy when combining five machine learning models with biomarker data: random forest, logistic regression, SVM, decision tree, and CNN. These results demonstrate that the hybrid model significantly improves the accuracy and efficiency of ovarian cancer detection and classification, offering superior performance compared to traditional methods and individual classifiers.

Keywords

Ovarian cancer early-stage detection UNet model fuzzy c-means clustering nomadic people optimizer

1 Introduction

Abnormal cell growth and development in the ovary, leading to tissue malignancy, is known as ovarian cancer.¹ The mature stroma, epithelial cells, and germ cells are the three primary categories into which they may be separated. Endometriosis and mucinous, humoral, and clear cells are subtypes of epithelial ovarian cancer. Bowel malignancies of the intestinal fluid are classified into high-grade bowel cancer (HGBC) and low-grade bowel cancer (LGBC). Data indicate that 70–80% of individuals develop epithelial cancer, less than 5% develop mucus-toxic clear cells, and 10% develop endometriosis. To avoid death, epithelial carcinoma, the initial indication for ovarian cancer, must be identified quickly. Deep learning is a proposed method for diagnosing ovarian cancer,² and there are numerous ways to make this diagnosis.

The precision and effectiveness of treatments for epithelial ovarian cancer are enhanced by the presence of distinct genetic signatures found in multiple ancient subgroups. In line with both inhibitors coupling resistance damage and abnormalities in DNA response pathways, this allows the identification of responses to ovarian cancer susceptibility genes, such as BRCA2 and BRCA1. Utilizing the genetic alterations found in tumors and blood, this step evaluates sensitivity, resistance to treatment, and reliable markers of residual disease. Approximately two-thirds of the women (150, 000) will be evaluated. This indicates that EOC ranks eighth among cancers that affect women. Peutz-Jeghers syndrome and other widespread illnesses constitute a genetic condition. These include polycystic ovarian syndrome, smoking, menopause, early hirsutism, and fetal pregnancy.³

Taking into account the complexity and variability of cancer survival prediction, this study integrated cutting-edge machine learning (ML) algorithms with newly created whole-slide histopathology scanners, high-throughput omics analysis, and⁴ Pathological images and genomic data of these individuals are highly useful for cancer assessment in a thorough investigation of cancer survival prediction. To predict the course of breast cancer,⁵ combined genetic data with pathological images. Multi-kernel learning techniques have been applied to integrate two types of non-uniform data. Accuracy values of 0.8022 and 0.7273 were obtained using this procedure, respectively.

The multimodal and multi-feature M2DP feature selection method was presented in⁶ for cancer diagnosis. Thus, features that were relevant for diagnosis were identified by applying the M2DP model after features were extracted from the gene expression data and disease images. Features chosen specifically for each patient were used with AdaBoosting to diagnose them. In testing, the model's accuracy was 72.53% and 70.08%, respectively, using benchmarks for lung and breast cancer. Merged genetic data with image pathology features to create a multi-core technique⁷ for predicting lung cancer; however, the accuracy was only 0.8022.⁸ Finally, weighted linear aggregation was used to fuse the different data. The model uses deep learning methods to fully extract deep features from the genetic and image formats. It was 88.07% accurate for anticipation.

Medical diagnosis will be highly affected by recent developments in Convolutional Neural Networks (CNNs) and other related deep learning models.⁹ Off the image patches, 91.5% were classified as benign or malignant using the algorithm with reasonable accuracy.¹⁰ suggested using residual networks (ResNet50) to automatically categorize brain cancers. The accuracy of the model was 0.97.¹¹ suggested a disease inference technique that generated an F-score of 0.572 by utilizing a bidirectional LSTM to derive the symptom order from discharge summaries.

With larger pre-trained datasets, larger transformer models provide stronger intermediate representations according to research on the impact of size on transition learning.¹² In addition, we examined the internal representation structure of the CNN and converter models on an image classification benchmark and highlighted notable distinctions between the two designs, including the converter's more cohesive representation across all layers. A thorough comparison study of CNN models and converter performance in image classification tasks was carried out by.¹³ The converter model, as demonstrated by the authors, outperforms CNNs in terms of computational efficiency and ease of training, and it can obtain better results on a wide range of benchmark datasets. In generative tasks, such as image synthesis,¹⁴ examined how well transformers and CNN models were performed. In terms of quality and sample diversity, it demonstrated that the converter model can produce images of high caliber and exceed CNNs.

Using a CNN and transformer combination model,¹⁵ proposed an effective way to extract low-level information from images and find long-range correlations between forms. Through an average accuracy improvement of 10.1% over the state-of-the-art CNN model, their study demonstrated that the sensor significantly improved multimodal image processing over CNN. CNN overlooks long-term connections in images, such as non-local object associations, according to previous research.¹⁶ This is not the case with the transformer model, which captures long-term dependencies within the input image and formulates image classification as a sequence-prediction problem across a succession of image patches. Transducers yield more accurate and lucid findings in most medical images when compared to pure CNN, according to a thorough study of transducers in medical image analysis published in.¹⁷ The number of transformers in this industry is growing rapidly. Sensitivities of 91.5% and 82.2% were obtained using the proposed approach.

Pretrained weights from ImageNet are most commonly utilized to enhance the performance of deep learning models, and they are frequently used in sensor and CNN research. Their work enables sensors to train image analysis algorithms more effectively through transformational learning. Combining sensors with CNNs was suggested by¹⁸ as a way to address CNNs’ intrinsic lack of long-range reliance. Comparing this combination to earlier proposed CNN- and DCNN-based models, the electroencephalogram (EEG) results are better.¹⁹ suggested a model for medical image segmentation based on transformers. Their work emphasized the benefits of using converters in the design of Swin-Unet, a converter approach to segmented models. After evaluating several transformer-based and U-net-based models, here list the benefits of the transformer-based models for image segmentation.

A study²⁰ demonstrated superior feature extraction over pure CNN or U-Net models by comparing CNN with transformer-based hybrid models. They surpassed U-Net-based models on comparable datasets with noteworthy results from their experiments. In order to improve accuracy,²¹ proposed integrating the Transformer and U-Net models, pointing out the shortcomings of U-Net in feature extraction and the robustness of Transformer in image analysis.

1.1 Contributions

To conduct statistically sound and reliable analytical investigations, apply machine learning models to large-scale datasets, including blood samples, common chemical tests used in medicine, and OC indicators.

This work builds on a previous conference paper in which suggest building a classifier for microarray data using an ML pipeline that consists of feature discretization (FD) and feature selection (FS) blocks.

Here, created a semantic segmentation sensor model and compared its results to the UNeT model for the identification of ovarian tumors. In terms of segmentation, the transformer model outperformed UNet.

By contrasting the transformer-based approach with the well-known UNet model, a thorough examination of the segmentation models was performed. Metrics such as Dice scores and Jaccard scores were applied in this evaluation.

To present a Fuzzy C-Means Clustering based computer-aided diagnostic (CAD) system optimized by a nomadic people optimizer (NPOA) using ultrasound images.

2 Methodology

2.1 Proposed network architecture

Although not yet commonly used, deep learning has the potential to identify ovarian cancer using computed tomography (CT) scan data. Few studies have used ensemble deep learning to accomplish their objectives. To achieve greater accuracy, ensemble deep learning models have been constructed by fusing many CNN models. The model can efficiently extract the necessary information from the input image because of its numerous layers and wide learning scope. There are several key stages, including data collection, and dataset preparation, pre-processing, feature extraction, segmentation and classification.

Following patient consent, a number of CT scan images were gathered in the first stage. Removing private or unnecessary information from images is known as pre-processing. The dataset was also expanded by data augmentation. The proposed four-path integration architecture receives this data as input. In other words, these data were received by four linked CNN models, each of which independently extracted features. Because there is a lot of information in the original image, including aspects that are unrelated to the region of interest, segmentation was performed first, followed by classification. Therefore, segmentation is primarily used to make the classification model less complex so that it can concentrate solely on the region of interest. A more discriminative set of information for classification can be obtained by extracting features, such as appearance, from the segmentation model, as opposed to classifying every pixel. This is helpful when extracting features for categorization. Because the segmented image is the only thing processed rather than the entire image, the accuracy and computational efficiency are increased as well. A multi-layer perceptron architecture was used to classify cases into two categories: benign and malignant, using the combined multiview feature vectors that were obtained. To assess the performance of the proposed model, evaluation metrics including F-score, recall, precision, and accuracy were used.

2.2 Data collection

The “Third Affiliated Hospital of Soochow University” provided 349 patients that constituted the dataset. Between July 2011 and July 2018, 178 patients with benign ovarian tumors and 171 patients with malignant ovarian cancer were respectively analyzed. A total of 49 features, gathered by pathology diagnosis, constituted the dataset. We divided the whole dataset into three subgroups: blood routine test (neutrophil ratio, thrombocytocrit, hematocrit, mean corpuscular hemoglubin, lymphocyte, platelet distribution width, mean corpuscular volume, platelet count, hemoglobin, eosinophil ratio, mean platelet volume, basophil cell count, red blood cell count, mononuclear cell count, red blood cell distribution width, and basophil cell ratio), general chemistry (albumin, calcium, indirect bilirubin, uric acid, nutrium, total protein, alanine aminotransderase, total bilirubin, blood urea nitrogen, magnesium, glucose, creatinine, phosphorus, globulin, gama glutamyl tranferasey, alkaline phosphates, kalium, direct bilirubin, carban dioxide-combining power, chlorine, aspartate aminotransferase, and anion gap) and tumor marker (carbohydrate antigen 72-4, alpha-fetoprotein, carbohydrate antigen 19-9, menopause, carbohydrate antigen 125, carcinoembryonic antigen, age, and human epididymic protein 4)) (shown in Table 1).

Table 1.
Attribute list for different subgroups of datasets.

Blood Routine Test General Chemistry Tumor Marker

Neutrophil ratio Albumin Carbohydrate antigen 72-4

Thrombocytocrit Indirect bilirubin Alpha-fetoprotein

Hematocrit Uric acid Carbohydrate antigen 19-9

Mean corpuscular hemoglobin Nutrients Menopause

Lymphocyte Total protein Carbohydrate antigen 125

Platelet distribution width Alanine aminotransferase Carcinoembryonic antigen

Mean corpuscular volume Total bilirubin Age

Platelet count Blood urea nitrogen Human epididymis protein 4

Hemoglobin Magnesium

Eosinophil ratio Glucose

Mean platelet volume Creatinine

Basophil cell count Phosphorus

Red blood cell count Globulin

Mononuclear cell count Gamma-glutamyl transferase

Red blood cell distribution width Alkaline phosphatase

Basophil cell ratios Potassium

Direct bilirubin

Carbon dioxide combining power

Chlorine

Aspartate aminotransferase

Anion gap

Blood Routine Test	General Chemistry	Tumor Marker
Neutrophil ratio	Albumin	Carbohydrate antigen 72-4
Thrombocytocrit	Indirect bilirubin	Alpha-fetoprotein
Hematocrit	Uric acid	Carbohydrate antigen 19-9
Mean corpuscular hemoglobin	Nutrients	Menopause
Lymphocyte	Total protein	Carbohydrate antigen 125
Platelet distribution width	Alanine aminotransferase	Carcinoembryonic antigen
Mean corpuscular volume	Total bilirubin	Age
Platelet count	Blood urea nitrogen	Human epididymis protein 4
Hemoglobin	Magnesium
Eosinophil ratio	Glucose
Mean platelet volume	Creatinine
Basophil cell count	Phosphorus
Red blood cell count	Globulin
Mononuclear cell count	Gamma-glutamyl transferase
Red blood cell distribution width	Alkaline phosphatase
Basophil cell ratios	Potassium
	Direct bilirubin
	Carbon dioxide combining power
	Chlorine
	Aspartate aminotransferase
	Anion gap

2.3 Data pre-processing

Data on gene expression and copy number variation were included in genetic morphology. We combined data from the same form to add properties from both data types to the network simultaneously. The combined data included 31,202 data indications. When the two data types were integrated, we preprocessed the data using the Z-score to eliminate the dimensionality mismatch and distinct chemotaxis issues. Finding the mean and standard deviation of the original data allows normalization of the data using the Z-score approach. These specifications are represented by the following expressions.

\begin{aligned} y * = \frac{y + \bar{y}}{δ} \end{aligned}

(1)

The formula mentioned above uses y,

\bar{y}

, and

δ

to represent the data item's value, mean, and standard deviation, respectively. Data adjustments will not result in “failure” but rather an improvement in data performance because of the linear structure of Z-Score. Principal component analysis was used to decrease noise in the data and eliminate duplicates. As a result, there were now 405 instead of 31,202 gene pattern data images in the dataset.

2.4 Image pre-processing

Input to a CNN cannot be readily obtained from pathological images because of their large size and rich pixel characteristics. To obtain image patches for this study, we first chop the full-size image into smaller images. After reducing all pathological images to sub-batches of 1024 × 1024 pixels, we separated each image into four smaller ones, each measuring 512 × 512 pixels based on the center point. For feature extraction using a CNN, these image patches were enlarged to 224 × 224 pixels, as illustrated in Figure 1. Each batch label is identical to the original label of the full-size image.

Figure 1.

Preprocessing of pathology images.

2.5 Data augmentation

To enhance the dataset size and strengthen the robustness of the model, various data augmentation strategies were implemented. These included geometric transformations, such as rotation, translation, flipping, and scaling, to introduce new variations in CT scan images. For example, images were rotated within a specified range of −45° to +45°, and both horizontal and vertical flips were used to reflect the different orientations of the ovarian tumors. Intensity-based augmentations, such as adjusting the brightness and contrast, were also applied to simulate varying lighting conditions that might be encountered in actual scenarios. Additionally, noise was introduced to the images to improve the ability of the model to handle distortions typically present in medical imaging. Elastic deformations were employed to replicate realistic tissue deformations, allowing the model to better understand the anatomical variations.

2.6 Feature selection

To address the “curse of dimensionality” and acquire an accurate representation of the data and better ML model outcomes when dealing with high-dimensional data, dimensionality reduction approaches should be used. FS is a successful method to reduce the dimensionality of microarray data. The FS method applies selection criteria to select subsets of the features from the original set. Rating the characteristics according to the perceived significance of a certain feature is one method of FS execution. FS, also known as gene selection, is another name for the procedure employed for microarray data. A few popular techniques for handling microarray data are as follows.

Non-mapping techniques: term variance, spectral (also called SPEC), and Laplacian score (LS).

Fisher Ratio (FiR), fast correlation-based filter (FCBF), Maximum Relevance Minimum Redundancy (MRMR), ReliefF, and Relevance Redundant Feature Selection (RRFS) are some mapping techniques.

Moreover, the mean-median (MM) correlation measure defined for the ith feature can be used to execute the RRFS technique in the unmapped mode as follows:

\begin{aligned} M M_{i} = | \bar{X_{i}} - m e d i a n (X_{i}) | \end{aligned}

(2)

The mean of the i^th characteristic is indicated by

X_{i}

. RRFS employs the Fisher ratio,¹⁸ also referred to as the Fisher score, which is defined as (for the i^th feature) as the relevance metric in the supervised mode.

\begin{aligned} F i R_{i} = \frac{| \bar{X_{i}^{- 1}} - X_{i}^{1} |}{\sqrt{v a r (X_{i}^{- 1}) + v a r (X_{i}^{1})}} \end{aligned}

(3)

Considering the patterns of each of the two classes (denoted as −1 and 1), the sample averages and variances of feature

X_{i}

are

v a r (X_{i}^{- 1})

and

v a r (X_{i}^{1})

. It has been observed that this ratio functions well as a relevance criterion for FS tasks as it quantifies the degree to which each characteristic alone distinguishes between the two classes. The FiR for feature FV is generalized as follows for more than two classes.

\begin{aligned} F i R_{j} = \frac{\sum_{j = 1}^{c} n_{j} ({\bar{X_{i}}}^{j} - \bar{X_{i})^{2}}}{\sum_{j = 1}^{c} n_{j} v a r ({\bar{X_{i}}}^{j})} \end{aligned}

(4)

where c is the number of classes, j is the number of samples in class j,

X_{i}^{j}

is the sample mean of

X_{i}

taken only into consideration for data in class j, and

X i

is the sample mean of feature

X_{i}

. The Fisher ratio has been effectively applied to microarray data, among many other applications. Only the top-rank characteristics were retained when the FS was calculated using the Fisher ratio.

2.7 Segmentation using transformer model

Attributing a class or category label to each image pixel is the primary function of semantic segmentation and is extensively employed in computer vision. Fully convolutional networks are fully utilized for semantic segmentation. However, self-attention-based sensors have been shown to be more successful in splitting operations in recent years. The converter operates using an encoder-decoder design. Seq2Seq (sequence-to-sequence) architecture is the name given to the converter presented in the “Attention Is All You Need” paper. To enable the decoder to produce the output sequence, the encoder maps the input sequence into a higher-dimensional space. The architecture of the converter model for partitioning is shown in Figure 2. The following information relates to the transformer model: There are 24 layers, 1024 tokens, 16 headers, and 307 million parameters in the Seg-L model, which is based on the ViT-L backbone architecture. Recently, an architecture that handles images as a sequence of patch markers alone was proposed in vision, without deconvolution. Image input for separating benign from malignant tumors utilizing an encoder-decoder converter architecture. Owing to the more distinct and well-defined margins of benign tumors than malignant ones, benign images outperform malignant images. Less accurate segmentation occurs when malignant tumors lack distinct borders and forms.

Figure 2.

Transformer model architecture used for segmentation of benign and malignant tumors.

2.8 Segmentation using U-Net model

The U-Net is a well-known model for semantic image segmentation. Using up-sampling and down-sampling (encoding and decoding), the U-shaped architecture of the model was employed. Figure 3 depicts the U-net architecture.

Figure 3.

U-Net model architecture for segmentation of benign and malignant tumors.

Through an encoding stage, also called a contract network, the model recognizes objects present in the image. The number of pixels in an image is halved each time it passes through an encoding layer. After processing the feature maps received by the lower layers, the decoder, also known as a boosting network, creates segmentation masks. By combining the feature maps produced at each encoding stage with the equivalent feature maps produced at the decoding stage, skip connections are essential for the effectiveness of the U-Net. The segmentation maps produced by these connections (shown as gray arrows in Figure 2) are dependent on the situation-specific properties acquired during the encoding cycle. Furthermore, it minimizes variations in the image intensity for the model. The classification and localization duties are executed by the encoder and decoder in tandem. Two convolutional layers and a ReLU activation layer sandwiched in between make up this architecture's bottleneck layer. To the first layer of the decoder, this layer generates the final feature map.²²

3 Classification

Fuzzy c-means clustering

Several regions of CT images, including mucus cells, clear cells, endometriosis, and intestinal fluid, were recognized and displayed as part of the classification process. Since ovarian cancer can vary in size, shape, location, and aggressiveness, it might be challenging owing to its complexity and heterogeneity. Further factors influencing image quality could be variations in imaging procedures and noise artifacts. Active contouring, region growth, thresholding, and machine learning-based algorithms are strategies that can be utilized to classify ovarian cancer from CT scans. This task was performed using the fuzzy-clustering method. Fuzzy clustering image classification can be used to distinguish between tumor regions and ovarian cancer in the context of ovarian cancer detection. The classification results can be utilized for quantitative analysis and clinical decision-making. The use of fuzzy clustering for ovarian cancer classification has shown encouraging results in several studies. To maximize the capacity for CT image categorization, fuzzy C-means clustering is recommended. Being highly susceptible to noise, outliers, and cluster size, the fuzzy approach is the most widely used approach. Up to convergence, we iteratively update the cluster center and membership after randomly initializing the fuzzy parameters and the cluster center. Determining the highest velocity cluster to which a pixel belongs is how the classification results are determined. We adjust the membership measure range for each observation point based on the separation between all data points and the geographic centers of all populations, following the initialization of membership values and the computation of population centers. Following the convergence of the method, the highest-level membership value is used to determine the final cluster to which each data point is assigned. Assume A = (a1, a2,…, an) and that there are “n” finite data collections. By employing the FCM technique, Dataset A was separated into cluster groups. Equation (5) for the FCM algorithm.

\begin{aligned} J_{s} = \sum_{p = 1}^{n} \sum_{q = 1}^{c} u_{p q}^{s} | | A_{p} - C_{q} | |^{2} \end{aligned}

(5)

where,

u_{p q} = \frac{1}{\sum_{k = 1}^{c} {(\frac{d_{p q}}{d_{k q} *})}^{\frac{2}{s} - 1}}

is the degree of membership value of

x_{p}

in q^th cluster,

x_{p}

is p^th data points and

c_{q} = \sum_{q = 1}^{n} u_{p q}^{s} / \sum_{q = 1}^{n} u_{p q}^{s} x_{p}

is q^th cluster center, s denotes the fuzzy parameter and Euclidean distance is represented as ‖ ‖.

The distance between the kth cluster focus and qth cluster center is represented by $d_{k q} *$ , where $d_{k q} *$ denotes the degree of aloofness between data points $x_{p}$ and qth cluster focus.

Optimizing the parameters of FCM utilizing the nomadic people optimizer

A group-based metaheuristic algorithm known as the Nomadic People Optimizer (NPOA) was developed. Based on the position of the leader, various clans in the NPOA search for the optimal position or solution. To acquire the best FCM values, we now use a step-by-step method that utilizes NPOA and deep learning. To reach global optimality more quickly, it is possible to transition seamlessly from navigation to exploitation. Therefore, the NPOA can rapidly find a perfect fitness solution. The NPOA method was selected owing to its advantages. It performs well in handling complex problems with high dimensionality. For precise pelvic mass classification, the NPOA algorithm was applied. In Figure 4, the nomadic optimization program flowchart for FCM classifier optimization is displayed. The following explains the NPOA's detailed procedure:

Figure 4.

Flowchart of Nomadic People Optimizer with FCM classifier.

Figure 5.

Confusion matrix for dataset.

Step 1: Initialization

An initial set of leaders (u), denoted as $u j = u 1, u 2, \dots \dots \dots, # C l a n s$ , was set up to optimize the FCM weight parameters, as shown in Equation (6):

\begin{aligned} v_{d}^{^{″}} = (u b^{'} - l b^{″}) \times R D^{″} + l b^{″} \end{aligned}

(6)

R D^{″}

denotes the random value between 0 and 1,

v_{d}^{^{″}}

specifies the leader position of clan d, and

u b^{'}

and

l b^{″}

mark the upper and lower bounds.

Step 2: Producing random

The evolutionary gravitational recognition neural network classifier was initialized, and randomly generated input parameters were generated with the aid of the nomadic people optimizer.

Step 3: Functional Fitness

Equation (7) provides the foundation for determining the fitness function:

\begin{aligned} F i t n e s s f u n c t i o n = o p t i m i z e M_{q} (a) a n d Q_{q} (a) \end{aligned}

(7)

Step 4: Exploitation behaviors of local search for optimizing

M_{q} (a)

$Y j = Y 1, Y 2, \dots, # f a m i l i e s$ is a set of families (y) shared with the matching leader u. There is no requirement for the x and y coordinates in situations where the solutions are expressed in the search space. As a substitute for two dimensions, the representation of solutions in the future is unary (one-dimensional). After removing the optional Y-coordinate, the tent distribution around the leader's tent must be random with respect to the x-coordinate. Equation (8) is how it is stated.

\begin{aligned} Y_{d}^{^{″}} = v_{d}^{^{″}} \times \sqrt{R D^{″}} \times c o s θ \end{aligned}

(8)

where

Y_{d}^{^{″}}

denotes the rank of the family, u"d the clan chief, and

R D^{″}

a random number within the [0,1] range.

Step 5: Analyzing global search behavior to improve $Q_{q} (a)$

Exploration is performed if the swarm is devoid of any new local best solution. In these circumstances, families seek better positioning, far from the current best available locally. Equation (9) is used to generate the directions using the Levy flight formula.

\begin{aligned} Y_{j}^{n e w^{'}} = Y_{j}^{o l d^{'}} + (N_{d} * (v_{d}^{^{″}} - Y_{j}^{o l d^{'}}) \oplus L e^{'}) \end{aligned}

(9)

where

N_{d}

indicates the clan's region, and

Y_{j}^{n e w^{'}}

and

Y_{j}^{o l d^{'}}

represent the new and old locations of the present family, respectively.

Step 6: Announcement

Upon termination, the NPOA method is used to optimize the optimum hyper-parameters $M_{q} (a) a n d Q_{q} (a)$ of the FCM parameter until the $i = i + 1$ halting criterion is reached. Finally, the pelvic masses were accurately classified with increased accuracy into three categories using the FCM-NPOA classifier: benign, malignant, and borderline (serous and mucinous).

4 Results and discussion

The dataset provided by the Third Affiliated Hospital of Soochow University included 349 patients diagnosed with ovarian tumors. The ages of the patients ranged from 25 to 75 years, with a mean age of approximately 50 years. The distribution of patients across different cancer stages was as follows: 20%, 30%, 35%, and 15% had stages I, II, III, and IV, respectively. Additionally, the dataset included various clinical metrics, such as tumor size, CA-125 levels, and histopathological grading.

These augmentation techniques were applied dynamically during each training epoch or in a preprocessing step to create an expanded set of new images, significantly increasing the overall dataset size. For instance, if four new images were generated for each original image, the dataset size could increase from 1000 to 5000 images. Specific parameters of each augmentation, such as the rotation angles or brightness levels, were carefully selected to ensure that the variations introduced were meaningful while preserving the essential characteristics of the images. The choice of these techniques was aimed at addressing the variability in tumor orientation, appearance, and noise in the CT scans, thus improving the model's ability to generalize across different cases.

To execute these augmentations effectively, tools and libraries such as TensorFlow, PyTorch, and OpenCV were utilized, ensuring the efficient application of these techniques. The impact of the augmentations was assessed by comparing the model's performance with and without them, revealing significant improvements in accuracy and the F1-score. This demonstrates the value of data augmentation in enhancing the model's predictive performance. Ultimately, the augmentation process not only increased the dataset size but also contributed to a more accurate and reliable model for detecting ovarian cancer.

The updated U-net model was trained and run under the experimental conditions shown in Table 2.

Table 2.
Configuration and environmental settings of the experiment.

Parameters Values

Image size for the experiment 512 * 512 pixels

Batch Size 64

Number of Epochs 1000

Number of hidden neurons for the ReLU 256

GPU Nvidia RTX 3060 1.78 GHz with 3585 cores

RAM 32 GB

Parameters	Values
Image size for the experiment	512 * 512 pixels
Batch Size	64
Number of Epochs	1000
Number of hidden neurons for the ReLU	256
GPU	Nvidia RTX 3060 1.78 GHz with 3585 cores
RAM	32 GB

Bicubic interpolation is used to resize the images. The aim of this dropout layer was to prevent overfitting. With Beta 1 and beta 2 parameter values of 0.6 and 0.8, we employed the Adam optimizer in this experiment. The model was configured at a learning rate of 0.0001. The output can be classified as either benign or malignant. Each pretrained CNN model was adjusted separately.

5 Evaluation metrics

Based on the recall, accuracy, precision, and F1 score, the performance of the proposed model was assessed. These metrics can be mathematically stated in the following manner.

\begin{aligned} A c c u r a c y & = \frac{T P + T N}{T P + T N + F P + F N} \times 100 \end{aligned}

(10)

\begin{aligned} P r e c i s i o n & = \frac{T P}{T P + F P} \times 100 \end{aligned}

(11)

\begin{aligned} R e c a l l & = \frac{T P}{T P + F N} \times 100 \end{aligned}

(12)

\begin{aligned} F 1 - s c o r e & = \frac{2 \times P r e c i s i o n \times R e c a l l}{P r e c i s i o n + R e c a l l} \end{aligned}

(13)

5.1 Performance analysis of feature selection

Extract features by applying RRFS techniques. 42 attributes were eliminated by using the previously discussed texture-based method. The RRFS method should be used to select the next parts. For performance evaluation, a classifier known as FCM-NPOA was utilized. Thirty distinct methods were used to conduct these experiments. Countless variables were tracked throughout this period, such as the quantity of selected characteristics and execution time. The three optimization techniques considered in this study are listed in the Table 3 below: LS, SPEC, and RRFS.

Table 3.
Performance Analysis of 3 optimization techniques.

Optimization technique Experiments Execution time (sec.) Training Accuracy (%) Testing Accuracy (%) Number of selected features

LS Exp. 1 4.82 88.32 82.43 31

Exp. 30 3.92 87.15 82.43 33

SPEC Exp. 1 5.45 88.75 84.73 29

Exp. 30 4.64 87.5 83.52 32

RRFS Exp. 1 3.53 88.52 84.73 30

Exp. 30 3.70 87.72 83.69 29

Optimization technique	Experiments	Execution time (sec.)	Training Accuracy (%)	Testing Accuracy (%)	Number of selected features
LS	Exp. 1	4.82	88.32	82.43	31
	Exp. 30	3.92	87.15	82.43	33
SPEC	Exp. 1	5.45	88.75	84.73	29
	Exp. 30	4.64	87.5	83.52	32
RRFS	Exp. 1	3.53	88.52	84.73	30
	Exp. 30	3.70	87.72	83.69	29

5.2 Segmentation

The subsections below compare the segmentation results according to the stated performance measures using the transformer and U-Net models.

5.3 Performance metrics

The performance metrics given below were used to assess the segmentation results.

Dice score:

The dice score between the segmented image and the ground truth image quantifies the similarity between the two. The Dice scoring formula was as follows:

\begin{aligned} D i c e = \frac{2 * | S \cap G |}{| S | + | G |} \end{aligned}

(14)

where G is the segmented region based on the ground truth and S is the segmented region predicted by the model. The symbol |.| denotes the cardinality of the set. Better segmentation outcomes were achieved when the Dice score was closer to 1, which ranged from 0 to 1.

Jaccard score:

This measure determines the overlap area between the ground truth and the segmented data. The Jaccard scoring formula is as follows:

\begin{aligned} J a c c a r d = \frac{D i c e}{2 - D i c e} \end{aligned}

(15)

The Jaccard score ranges from 0 to 1. Improved segmentation outcomes were associated with higher values.

5.4 Comparison of UNet and transformers

The segmentation results of Transformer and UNet on an open dataset are displayed in Figures 5 and 6 and Table 3. Randomly choosing data from either the validation/test or training dataset allows the model to be successfully trained during testing. In practical situations, the segmentation model exhibited enhanced power and performance when using this configuration. The results indicate that the converter model performs better than UNet, with a malignant image score of 0.98, a frontal image score of 0.97, a malignant image scoring of 0.99 and a Jaccard score of 0.98. The training periods for the converter and segmentation model UNet were 5000 and 6000 epochs, respectively. There is no evidence of saturation in Figure 7, indicating that the model was trained effectively.

Figure 6.

Segmentation results for benign and malignant images using UNet.

Figure 7.

Segmentation results for benign and malignant images using transformers.

Figure 8.

Epochs vs. loss analysis of UNet (left) and transformers (right).

Figure 9.

Fitness improvement.

6 Classification

Table 4 presents the performance evaluations of the proposed system. The proposed FCM-NPOA has excellent accuracy with respect to other classifiers. Comparing the suggested approach to existing methods, the latter has a lower average precision, recall, sensitivity, and specificity. The categorization metric, known as the F1 score, is derived from the average precision and recall values. Table 5 presents a graph illustrating the results of the performance analysis.

Table 4.
Quantitative comparison of Dice and Jaccard score of the UNet model and transformers.

UNet Transformers

Benign Malignant Benign Malignant

Dice Jaccard Dice Jaccard Dice Jaccard Dice Jaccard

Mean 0.92 0.85 0.95 1.1 0.99 0.98 1.01 0.99

Std. Dev 0.05 0.07 0.06 0.12 0.02 0.03 0.02 0.02

	UNet	Transformers
Mean	0.92	0.85	0.95	1.1	0.99	0.98	1.01	0.99
Std. Dev	0.05	0.07	0.06	0.12	0.02	0.03	0.02	0.02

Table 5.

Experimental result analysis for different parameters with other metrics.

Method	Accuracy	Recall	Precision	F1-score	Sensitivity	Specificity
ML-CNN	97.5	97	99	98	91.68	90.8
CNN-DenseNet	95.84	98.9	92	96	96	97.95
AlexNet	85.56	9	76	88	73	96
Proposed FCM-NPOA	99.47	99.32	94	98	99	97.5

While the study yielded promising results, it has some limitations that need to be recognized. A primary limitation is the potential issue of generalization, as the model was trained on a relatively specific and limited dataset. Although the model performed well on this dataset, its effectiveness may not hold when applied to larger, more diverse datasets that encompass a broader range of patient demographics and imaging variations. Additionally, the computational complexity of the FCM-NPOA model presents challenges for real-time clinical implementations. The significant training time and high computational resource requirements may make it less practical in settings with limited resources. Furthermore, the lack of diversity in the dataset may not fully represent the wide spectrum of ovarian cancer cases across different populations, which could limit the robustness of the model in more varied clinical scenarios.

To overcome these limitations, future research should prioritize expanding the dataset to include a more diverse and extensive range of patient populations, potentially through multi-institutional collaboration. This would enhance the model's ability to generalize and improve its robustness in different clinical settings. Moreover, it is important to explore ways to reduce the computational demands of the FCM-NPOA model to optimize it without compromising the accuracy. Making the model more suitable for real-time applications could involve developing lighter versions that achieve a balance between the performance and computational efficiency. Finally, future studies should validate the model externally by testing it on datasets from various institutions and regions. This further confirmed its effectiveness and ensured its applicability in diverse clinical environments. By addressing these issues, research can move towards developing more reliable and practical solutions for the detection and classification of ovarian cancer.

The optimization procedure is depicted in Figure 8, where an increase in fitness corresponds to an increase in the value or performance index of the target function under optimization. This can be accomplished by iteratively searching for better solutions using a variety of optimization approaches, including simulated annealing, gradient descent, and evolutionary algorithms.

The confusion matrix offers detailed insight into the performance of a classification model used to predict the stages of a disease, likely ovarian cancer, across five categories: Stage I, Stage II, Stage III, Stage IV, and Stage Not Available. From the matrix, it is evident that the model shows strong performance in identifying Stage III and Stage IV cases, with 436 and 89 true positives, respectively, indicating that these stages are most accurately classified. Stage I and Stage II cases were less frequently encountered, as reflected by the lower true positive counts of 17 and 29 for Stage I and II, respectively. However, a few misclassifications are noted, particularly with Stage III being occasionally mistaken for Stage I, and Stage II being misclassified as Stage III, although these instances are minimal. For the ‘ Stage-Not Available’ category, the model correctly identified five cases, with only one misclassification involving Stage III. Overall, the confusion matrix illustrates the high accuracy of the model, particularly in distinguishing between the more advanced stages of the disease, while also revealing areas where minor misclassifications occur. These insights are valuable for further refining the model to improve its predictive accuracy, especially in distinguishing the early stages of the disease, which is critical for timely intervention.

The confusion matrix of the proposed model is shown as Figure 9.

7 Discussion

With a remarkable sensitivity of 99% and an astounding accuracy of 99.47%, the proposed FCM-NPOA model produced highly intriguing findings. It is important to diagnose and treat ovarian cancer as soon as possible, and these results show that the model does a good job of doing so. However, the suggested job has some restrictions that should be understood. First, the model's ability to generalize is impacted by the need for larger and more varied datasets, which stabilize the performance. Second, using the FCM-NPOA model in real-time clinical settings may not be feasible owing to its computational complexity. Simplifying the model, increasing the dataset size, and overcoming these restrictions are the main goals of future studies. Investigating additional performance indicators and carrying out external validation on various datasets will enhance the dependability and suitability of the model for use in clinical practice.

8 Conclusion

Future research could proceed in this direction because the proposed technique outperforms existing classifiers in terms of efficiency and accuracy. With an accuracy of 99.47%, recall of 99.32%, precision of 94%, and F1-measure of 98%, the proposed model performed better than the current models. This model uses a classifier that is more effective than AlexNet. The FCM-NPOA model is the classification model used in this approach. The proposed approach makes a significant contribution to the field of ovarian cancer identification and classification. Next, we tested the proposed technique using various industrial datasets and sources. Experimentation with various topologies, optimization techniques, and hyperparameters can enhance the performance of the FCM-NPOA algorithm. To evaluate the efficacy and dependability of future proposed approaches in actual healthcare settings, extensive clinical studies and expert collaborations are required to validate their performance.

Footnotes

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

Data sharing is not applicable to this article as no new data were created or analyzed in this study.

References

Athithan

Sachi

Singh

. Ultrasound-Based ovarian cysts detection with improved machine-learning techniques and stage classification using enhanced classifiers. SN Comput Sci 2023; 4: 571.

Ravishankar

Jadhav

Kumar

, et al. A deep learning approach for ovarian cysts detection and classification (OCD-FCNN) using fuzzy convolutional neural network. Meas Sens 2023; 27: 100797.

Thanupillai

Kamal Basha

. Pulse coupled neural network optimized with chaotic grey wolf algorithm for breast cancer classification using mammogram images. Concurr Comput Pract Exp 2023; 35: e7448.

Shajin

Aruna Devi

Prakash

, et al. Sailfish optimizer with Levy flight, chaotic and oppositionbased multi-level thresholding for medical image segmentation. Soft Comput 2023; 27: 12457–12482.

Rajesh

Shajin

Kannayeram

. A novel intelligent technique for energy management in smart home using internet of things. Appl Soft Comput 2022; 128: 109442.

Graham

Shamkhalichenar

Browning

, et al. A practical evaluation of machine learning for classification of ultrasound images of ovarian development in channel catfish (Ictalurus punctatus). Aquaculture 2022; 552: 738039.

Pesenti

Beltrame

Velle

, et al. Copy number alterations in stage I epithelial ovarian cancer highlight three genomic patterns associated with prognosis. Eur. J. Cancer 2022; 171: 85–95.

Shajin

Rajesh

Raja

. An efficient VLSI architecture for fast motion estimation exploiting zero motion prejudgment technique and a new quadrant-based search algorithm in HEVC. Circuits Syst. Signal Process 2022; 41: 1751–1774.

Rajesh

Kannan

Vishnupriyan

, et al. Optimally detecting and classifying the transmission line fault in power system using hybrid technique. ISA Trans 2022; 130: 253–264.

10.

Gupta

Kumar

. A novel multi-neural ensemble approach for cancer diagnosis. Appl. Artif. Intell 2022; 36: 2018182.

11.

Sharma

Saba

Johri

, et al. Automated hybrid deep learning-based paradigm for high-risk plaque detection in B-mode common carotid ultrasound scans: an asymptomatic Japanese Cohort study. 2022 AIUM Award Win 2022; 41: 125.

12.

Hsu

Hung

, et al. Automatic ovarian tumors recognition system based on ensemble convolutional neural network with ultrasound imaging. BMC Med Inform Decis Mak 2022; 22: 298.

13.

Arezzo

Cormio

La Forgia

, et al. A machine learning approach applied to gynecological ultrasound to predict progression-free survival in ovarian cancer patients. Arch Gynecol Obstet 2022; 306: 2143–2154.

14.

Akter

Akhter

. Ovarian cancer prediction from ovarian cysts based on TVUS using machine learning algorithms. In: Proceedings of the International Conference on Big Data, IoT, and Machine Learning: BIM 2021, Cox’s Bazar, Bangladesh, 23–25 September 2021, Springer: Singapore, 2022, pp.51–61.

15.

Wojtyła

Bertuccio

Giermaziak

, et al. European trends in ovarian cancer mortality, 1990–2020 and predictions to 2025. Eur J Cancer 2023; 194: 113350.

16.

Tragakis

Kaul

Murray-Smith

, et al. The fully convolutional transformer for medical image segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, Waikoloa, HI, USA, 2–7 January 2023, pp.3660–3669.

17.

Gan

, et al. Transformers in medical image analysis: a review. Intell Med 2023; 3: 59–78.

18.

Yuan

Zhang

Fang

. An effective CNN and transformer complementary network for medical image segmentation. Pattern Recognit 2023; 136: 109228.

19.

Kan

Shi

Zhao

, et al. ITUnet: Integration of Transformers and Unet For Organs-At-Risk Segmentation. In: Proceedings of the 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), Scotland, UK, 11–15 July 2022; IEEE: Piscataway, NJ, USA, 2022.

20.

Liu

Huang

Liao

, et al. A hybrid deep learning model for predicting molecular subtypes of human breast cancer using multimodal data. Irbm 2021: 11–15.

21.

Shao

Wang

Sun

, et al. Multi-task multi-modal learning for joint diagnosis and prognosis of human cancers. Med Image Anal 2020; 65: 101795.

22.

Andrade-Miranda

Jaouen

Tankyevych

, et al. Multi-modal medical transformers: a meta-analysis for medical image segmentation in oncology. Comput Med Imaging Graph 2023; 110: 102308.

	UNet				Transformers
	Benign		Malignant		Benign		Malignant
	Dice	Jaccard	Dice	Jaccard	Dice	Jaccard	Dice	Jaccard
Mean	0.92	0.85	0.95	1.1	0.99	0.98	1.01	0.99
Std. Dev	0.05	0.07	0.06	0.12	0.02	0.03	0.02	0.02

FCM-NPOA: A hybrid Fuzzy C-means clustering with nomadic people optimizer for ovarian cancer detection

Abstract

Keywords

1 Introduction

1.1 Contributions

2 Methodology

2.1 Proposed network architecture

2.2 Data collection

2.6 Feature selection

Fuzzy c-means clustering

Optimizing the parameters of FCM utilizing the nomadic people optimizer

Table 2. Configuration and environmental settings of the experiment. Parameters Values Image size for the experiment 512 * 512 pixels Batch Size 64 Number of Epochs 1000 Number of hidden neurons for the ReLU 256 GPU Nvidia RTX 3060 1.78 GHz with 3585 cores RAM 32 GB

5.3 Performance metrics

Table 4. Quantitative comparison of Dice and Jaccard score of the UNet model and transformers. UNet Transformers Benign Malignant Benign Malignant Dice Jaccard Dice Jaccard Dice Jaccard Dice Jaccard Mean 0.92 0.85 0.95 1.1 0.99 0.98 1.01 0.99 Std. Dev 0.05 0.07 0.06 0.12 0.02 0.03 0.02 0.02

8 Conclusion

Footnotes

Funding

Declaration of conflicting interests

Data availability statement

References

Table 2.
Configuration and environmental settings of the experiment.

Parameters Values

Image size for the experiment 512 * 512 pixels

Batch Size 64

Number of Epochs 1000

Number of hidden neurons for the ReLU 256

GPU Nvidia RTX 3060 1.78 GHz with 3585 cores

RAM 32 GB

Table 4.
Quantitative comparison of Dice and Jaccard score of the UNet model and transformers.

UNet Transformers

Benign Malignant Benign Malignant

Dice Jaccard Dice Jaccard Dice Jaccard Dice Jaccard

Mean 0.92 0.85 0.95 1.1 0.99 0.98 1.01 0.99

Std. Dev 0.05 0.07 0.06 0.12 0.02 0.03 0.02 0.02