Integration of adaptive segmentation with heuristic-aided novel ensemble-based deep learning model for lung cancer detection using CT images

Abstract

In recent days people are affected with lung cancer in, and the severe stage of this disease leads to death for human beings. Lung cancer is the second most typical cancer type to be found worldwide. Pulmonary nodules present in the lung can be used to identify cancer metastases because these nodules are visible in the lungs. Cancer diagnosis and region segmentation are the most important procedures because the prosperous prediction-affected area can accurately identify the variation in cancer and normal cell. By analyzing the lung nodules present in the image, the radiologists missed several useful low-density and small nodules, and this may tend to the diagnose process very difficult, and the radiologists needs more time to decide the prediction of affected lung nodules. Due to the radiologist’s physical inspection time and the possibility of missing nodules, automatic identification is needed to address these issues. In order to achieve this, a new hybrid deep learning model is developed for lung cancer detection with the help of CT images. At first, input images like CT images are gathered from the standard data sources. Once the images are collected, it undergoes for the pre-processing stage, where it is accomplished by Weighted mean histogram equalization and mean filtering. Consequently, a novel hybrid segmentation model is developed, in which Adaptive fuzzy clustering is incorporated with the Optimized region growing; here, the parameters are optimized by Improved Harris Hawks Optimization (IHHO). At last, the classification is accomplished by Ensemble-based Deep Learning Model (EDLM) that is constructed by VGG-16, Residual Network (ResNet) and Gated Recurrent Unit (GRU), in which the hyperparameters are tuned optimally by an improved HHO algorithm. The experimental outcomes and its performance analysis elucidate the effectiveness of the suggested detection model aids to early recognition of lung cancer.

Keywords

CT images lung cancer detection adaptive fuzzy clustering region growing ensemble-based deep learning model improved harris hawks optimization IHHO

1. Introduction

In recent days, the rapid enlarging cancer and other primary cause of death is lung cancer. Lung cancer is determined as a crucial health issue, and it is a highly diagnosed cancer in worldwide based on the reports of the World Health Organization’s (WHO’s) [1]. The method for identifying lung cancer is more important. Since, it gives more relevant data to practitioners and radiologists for making better decision about the detection of lung cancers. The study of the causes of the disease is useful for great management to save the lives of humans, animals and swarms [2]. Cancer is a terrible medical condition that encompasses more than 200 different dangerous medical problems and is not just one illness [3]. Cancer is generally defined as, any disease that is characterized by contaminated cells, unchecked growth and aberrant spread. If the spread of abnormal cells is not properly regulated, the anomaly identified in cells results in death and impairment [4]. Infectious agents, organisms, chemicals and tobacco are the external risk factors for the spread of cancer. Internal risk factors include immunological disorders, mutations and hereditary hormones. Various techniques are utilized to locate the categories and detect lung cancer at the early stages. The “Computed tomography (CT)” and “Magnetic resonance imaging (MRI)” are the two best approaches for detecting and diagnosing lung cancer [5]. The CT is an important to modal the image for diagnosing and monitoring the lung cancer. The benefit of CT is detecting lung cancer in its early stages and allowing doctors to recommend more efficient treatments [6].

Because of the compound anatomical area’s nodular position, tiny size and low contrast in the lung, the affected lung nodule is difficult to identify in its early stages [7]. The tumorous pulmonary nodules in CT thorax images are 3cm or less in size and round shaped. Therefore, detecting and diagnosing the tumorous pulmonary nodules are complicated. Automatic lung cancer prediction uses multiple techniques to identify the disease, including classification of cancer, selection of features, and extraction of cancer features, region segmentation and removal of image noise [8]. Some segmentation methods are self-organizing maps, fuzzy k-means clustering, fuzzy c-means, sobel detection, canny edge detection, distributed clustering and k-means clustering [9]. The likelihood of rectifying lung cancer is significantly higher with earlier observation and treatment. Monitoring of lung nodules is the first stage of classification, and other effective methods of identification lead to better outcomes. When compared to single view radiography, advanced recognition of lung cancer lessen the mortality rate by 20% [10]. However, diagnosing cancer from the image of a lung is based on the true negative and false positive results, burdening the patient’s using extra testing, expense, and stress as well as the doctor’s added workload [11]. The problem in the effectiveness of lung cancer treatment is inherited concerning forecast accuracy. In the field of biomedical, a sensitive process requires high qualification and time for the diagnosis and evaluation of the lung CT image, and it is done by field experts [12].

The screening techniques inspect the lung cells and cell variations that are used to forecast lung cancer, but maintaining the predictability is complex. CT is one of the screening technologies that efficiently identify the variations and alterations that are presented in the human body. These are discovered by exposing the body to X-rays [13]. Despite successfully predicting lung cancer, typical automatic systems still struggle with recognition accuracy rate [14] and take longer to analyze huge amounts of data. Furthermore, the system does not handle CT images of minimal quality, which may result in higher misclassification rates and incorrect lung features [15]. The enormous success of deep learning approaches in image identification and recognition is leads to various medical imaging issues. “Convolutional Neural Network (CNN)” excels the other modern techniques in a variety of medical image problems [16]. It is also better than human experts and other approaches in regards to accuracy and sensitivity. Therefore, to create reliable cancer identification, the clinical centre must implement Computer Assisted Automated Detection (CAD) technology [17]. Some of the drawbacks of the existing methods are could not be solved, and it is listed below. In existing approaches, the early detection of lung cancer is not possible for the ResNet model. Moreover, it fails to handle the huge amount of data while the training process takes place in the VGG-16 technique. Thus, it leads to cause overfitting issues. In GRU, the accurate position of the lung nodules is not detected exactly while screening takes place in the lung nodules. However, the existing model shows a high error rate thus, the misclassification problems occur while detecting lung cancer. To resolve the above-mentioned problems, a new IHHO-EDLM method is introduced to detect lung cancer better. Moreover, the offered model provides enhanced performance in terms of accuracy. Here, the classification performance can be increased due to the error rate reduction. Here, the pre-processing takes place; hence; it can minimize the background noise to improve the quality of the image to detect lung cancer. Moreover, the segmentation is performed using the VGG-16, ResNet and GRU to locate the exact position of the lung cancer effectively.

The contributions of the designed ensemble-based deep learning lung cancer detection model are mentioned as follows.

•
To adopt an ensemble-based deep learning model to detect and identify lung cancer at the early stage. In addition, it is helpful for patients who have breathing issues. The enhancement of the recommended technique is broadly applicable to healthcare applications.
•
To develop an IHHO algorithm for lung cancer detection. With the help of recommended IHHO algorithm, the parameters of maximum iterations and seed range are optimized to get better performance. Moreover, the offered IHHO algorithm is used to improve the feature propagation. However, the complexity issues while optimizing the complex parameters are resolved.
•
To segment the lung nodules, the adaptive fuzzy clustering and region-growing techniques are used for attaining effective segmentation, where the parameter is optimized to enhance the segmentation outcome. The segmented images make it easier to detect the affected lung nodules in an accurate manner. The segmentation takes place without losing valuable information.
•
To obtain the final outcome of lung cancer detection, an EDLM is used. In this EDLM, the VGG-16, GRU, and ResNet are used to develop EDLM, where the parameters are optimized from EDLM to get higher detection results with respect to accuracy and precision. Moreover, the recommended EDLM technique is performed to resolve the overfitting issues and underfitting issues.
•
The experimental outcome of the offered IHHO-EDLM-based lung cancer detection model is analyzed among several optimization techniques and other existing works to verify the performance of the recommended lung cancer detection model.

The upcoming sections of the ensemble learning-based lung cancer detection model using CT Images are explained in further sections. Here, Section 2 consists of a variety of existing lung cancer detection models with its features and drawbacks. Section 3 consists of a structural representation of the designed EDLM-based lung cancer detection model, and it describes about dataset and system model. Section 4 describes about the proposed algorithm and segmentation process in detail. Section 5 depicts about EDLM-based deep learning model and explains about the classifiers that are used to detect lung cancer. Section 6 includes the results and discussion of the suggested lung cancer detection model. In Section 7, the summarization of the developed EDLM-based lung cancer detection model using CT images is given.
2. Literature survey

In this section, several existing research work is focused and elaborated in Sub-section 2.1. Here, the diverse existing techniques are focused for detecting lung cancer using CT images. However, the challenges and benefits of the existing model are listed in the Sub-section 2.2. Owing to the limitations of the existing model, we developed a novel technique for the lung cancer detection model.

2.1 Related works

In 2019, Shakeel et al. [18] have introduced a lung cancer detection framework from CT images by utilizing deep learning and clustering approaches. Standard datasets were used for collecting the lung CT images. Histogram equalization has been used to remove the noise. Similarly, the profuse clustering technique was used for splitting the affected area in the image. For predicting lung cancer, those images were examined by utilizing Deep Neural Networks (DNN), and the simulation results were obtained using MATLAB. This approach gave better outcomes among lung cancer detection with minimum errors and better accuracy.

In 2021, Alzubaidi et al. [19] have investigated a feature extraction framework, that is used for identifying lung cancer. This implemented framework contained three main parts: i) collection of data’s ii) global and local training, and iii) testing. Image cropping and warping methods were used to preprocess the accumulated images during the testing and global training stage. Then the global features were retrieved from the accumulated images for representing every image with the feature vectors. The learning algorithms with the best performance in the global phase were used to build the detection model by the feature vectors. Every image was splitted into a group of local blocks in the testing and local training stage. From the results, the gabor filter features have obtained better specificity, sensitivity and accuracy rates.

In 2021, Sori et al. [20] have investigated a deep learning technique for detecting lung cancer using denoised CT scan images. The recommended “Denoising First Two-Path CNN (DFD-Net)” contained two parts for detecting and denoising. In the preprocessing phase, the noise was removed by using the denoising model. For detecting lung cancer, a two-path CNN was used, where the input was the noise removed images. The combination of global and local features was focused by two paths. Then the image contrast also balanced by utilizing the implemented retraining technique. Therefore the suggested model was useful in noise reducing, adapting different sizes and structures and balanced the size of the receptive field. This implemented method achieved competitive results.

In 2020, Karthiga and Rekha [21] have described a dynamic algorithm named “Accelerated Wrapper-based Binary Artificial Bee Colony (AWB-ABC)” for detecting lung cancer in its early stages. “Improved Naive Bayes classification (I-NBC)” of images was done to classify the different cancer stages. In the pre-processing phase, image enhancement and removing noise were done using Mini–mental state examination (MMSE). The pre-processed image was used for the retrieval of morphological features. By using the AWB-ABC algorithm, the features which have a great impact on causing lung cancer were chosen. For the final classification, an I-NBC algorithm was used. MATLAB was used for the simulation. Therefore, by using this proposed method, the classification accuracy has been improved.

In 2020, Shakeel et al. [22] have implemented an improved DNN and ensemble networks for the automatic identification of lung cancer. A dataset of CT scan images was collected to identify lung cancer. Deduction of the noise from the image and quality enhancement of the image was done by utilizing the multilevel brightness-preserving technique in the accumulated images. The affected area in the image was segmented, and also different features were retrieved by improved DNN. “Hybrid spiral optimization intelligent-generalized rough set technique” was used to choose the efficient features from the retrieved features. Then an ensemble network was used to classify the efficient features. This implemented method has given high lung cancer prediction results.

In 2019, Manickavasagam et al. [23] have proposed a “Neuro-Fuzzy Classifier” for detecting lung cancer automatically. This method contained two phases. During the first phase, the classification of abnormal or normal images was done using the Naive Bayes model. Similarly, in the second phase, various lung cancer stages have been classified. For the classification and detection process, 10 features were selected from the lung image. Therefore, the implemented algorithm classified lung cancer very effectively with the optimal selection of features. The results of the experiments were shown that the investigated model gave an improved classification accuracy rate.

In 2020, Halder et al. [24] have implemented an adaptive morphology-based segmentation method for identifying lung cancer. An adaptive morphological filter was used to identify the lung nodules using an “Adaptive Structuring Element (ASE)”. It was used to give higher accuracy outcome of the lung nodule detection process. For feature extraction, the detected candidate nodule regions have been processed. This method has reduced false positives. For detecting lung nodules, the “Support Vector Machine (SVM)” was used with intensity and texture-based features. This implemented method was accomplished with better outcome while detecting lung nodules automatically.

In 2020, Togacar et al. [25] have introduced a maximum relevance feature selection with a CNN-based lung cancer detection mechanism to give reduced redundancy results. In the classification and feature extraction process, CNN was used. The image augmentation approaches were applied to improve the classification rate among the collected CT images. By using the AlexNet model, the features were retrieved, and it was applied as an input to “softmax classifiers”, “SVM, Decision Tree (DT)”, “K-Nearest Neighbor (KNN)”, “Linear Discriminant Analysis (LDA)” and “Linear Regression (LR)” for checking the efficiency of the developed model. A better classification rate has been obtained by using both the NN classifier and the AlexNet model. This proposed model has given stable results while detecting lung cancer.

Table 1
Merits and demerits of existing lung cancer detection model

Author [citation]	Techniques	Merits	Demerits
Shakeel et al. [18]	IPCT	It acquires less classification error and high accuracy.	It consumes more time for training the features.
Alzubaidi et al. [19]	SVM with HOG	It maximizes sensitivity, accuracy and specificity.	It creates overlapping issues as it comprises multiple images.
Sori et al. [20]	CNN	It achieves a better detection rate.	Sometimes the data get lost during the implementation of the denoising process.
Karthiga and Rekha [21]	I-NBC	It obtains higher accuracy results. It reduces the error value.	It does not entail multiple classifiers to enhance the performance.
Shakeel et al. [22]	Ensemble	It increases detection accuracy. It mitigates the dimension.	The performance gets degraded since it contains a computational burden.
Manickavasagam et al. [23]	Neuro Fuzzy Classifier	It obtains more precision and accuracy.	It does not utilize other factors like location, growth rate and lung nodule size.
Halder et al. [24]	AMST	It yields desired results with distinct performance measures.	It does not provide effective segmentation results since it has low intensity variation among the nodules and structural complexity.
Togacar et al. [25]	CNN	It provides a high success rate in accordance with specificity, sensitivity and accuracy for cancer detection.	It does not consider other medical imaging techniques.

2.2 Problem statement

Lung cancer is the most disastrous disease rather than people get affected by breast or colon cancer. The detection and segmentation of cancer affected lung nodules becomes the major issue to treat patients. Also, various approaches are developed to the early detection of cancer but exist with some noteworthy challenges. The advantages and disadvantages of existing lung cancer detection techniques are given in Table 1. IPCT [18] acquire high accuracy and less classification error. But, for training the features, it consumes more time. SVM with HOG [19] maximizes sensitivity, accuracy and specificity. Yet, it creates overlapping issues as it comprises multiple images. CNN [20] achieves a better detection rate. However, sometimes the data get lost during the implementation of the denoising process. I-NBC [21] obtains higher accuracy results and reduces the error value. But, it does not entail multiple classifiers to enhance the performance. Ensemble [22] increases the detection accuracy and mitigates the dimension. On the other hand, the performance gets degraded since it contains a computational burden. Neuro Fuzzy Classifier [23] obtains more precision and accuracy. However, it does not utilize other factors like location, growth rate and lung nodule size. AMST [24] yields desired results with distinct performance measures. It does not provide effective segmentation results since it has low-intensity variation among the nodules and structural complexity. CNN [25] provides a high success rate in accordance with specificity, sensitivity and accuracy for cancer detection. But, it does not consider other medical imaging techniques for performance enhancement.

3. Motivation

Some limitations of the existing approaches are shown here. In the existing model, the training time is high. Thus, a timely diagnosis is impossible for radiologist to detect lung cancer. Here, the parameter optimization is not performed to get a better solution. On account of all aforementioned challenges and limitations, it is provoked to implement a novel method for the detection of lung cancer using CT images. Here, the developed model resolves the complex issues while the training the network. Moreover, it tends to manage a huge amount of data. Consequently, the offered model provides better intensity variations, which helps to identify the exact location of the lung nodules. The validation takes place with various evaluation measures for an effective outcome. Additionally, the accuracy rate is enhanced in the developed model.

4. Lung cancer detection using CT images: System model and dataset model

The section explores the developed model for lung cancer detection. Initially, the research gaps has been briefly expanded and then, the architecture of the developed ensemble-based deep learning model is visualized and evaluated in Subsection 3.1. In Subsection 3.2, the data that has been collected in this research work is explained. After the data collection, the preprocessing techniques are performed, and it is described in Subsection 3.3.

4.1 System model: Proposed lung cancer detection

Among the variety of cancer types, lung cancer is the major death causing cancer. Smoking is the main cause of affecting lung nodules. Therefore, the starting stages are necessary, and the initial stage detection may helpful to cure or control lung cancer [38]. Different approaches are used to identify lung cancer. But, the predicted results are not accurate all the time, and false identification of cancers is also possible during the detection of lung cancer. The prediction models require more cost to maintain and analyze the results effectively. For performing lung cancer detection, high contrast CT images are required. High radiation exposure is one of the main disadvantages of the existing lung cancer detection models. Automatic identification of lung cancer is necessary for avoiding the small and low density missing nodules. Moreover, the precise and accurate detection of affected lung nodules is essential to diagnose lung cancer at the initial stage. The deep learning techniques have the ability to learn the features very effectively from the CT images that may be useful for identifying the affected lung nodules. Therefore, to overcome those difficulties, an ensemble learning-based deep learning model is implemented. The architecture of the implemented ensemble-based deep learning model for detecting lung cancer is given in Fig. 1.

Figure 1.

The architecture of the implemented ensemble-based deep learning model for detecting lung cancer.

An ensemble-based deep learning model for lung cancer detection model is performed using CT Images to reduce the chance of people dying from lung cancer. In order to implement this, the CT images are collected from the benchmark data links. Then the collected raw images are preprocessed using the Weighted mean histogram equalization and mean filtering. These techniques magnify the image quality. Then the lung nodule segmentation is performed using the Adaptive fuzzy clustering and region growing approaches. In adaptive fuzzy clustering, maximum iterations are optimized using IHHO. Similarly, from the region growing approach, seed ranges are optimized using IHHO. This optimization of maximum iterations and seed range during the segmentation stage may improve the detection accuracy of the investigated framework. The segmented images are given to the ensemble learning network, where the images are detected using VGG-16, ResNet and GRU. Moreover, epochs from VGG-16, hidden neuron count and epochs from GRU, hidden neuron count and epochs from ResNet are tuned by the IHHO. The predicted scores from the used ensemble-based deep learning model such as vgg-16, ResNet and GRU are averaged and the final prediction score is obtained and this score is used for detecting lung cancer. The prediction results obtained from the implemented ensemble learning-based model for identifying lung cancer with the help of CT images are compared with other lung cancer detection models to ensure the effectiveness of the IHHO-EDLM-based lung cancer detection.

4.2 Dataset model

Dataset 1 (“Chest CT-Scan images Dataset”): The required CT images are gathered from Chest CT-Scan images that are taken from the dataset link of https://www.kaggle.com/datasets/mohamedhanyyy/chest-ctscan-images: access date: 2023-02-09. This website provides CT images of various types of chest cancer. The data set consists of 3types of chest cancer. The validation set, testing set and training set are present inside the data folder. In the dataset, 10% data are validation sets, 20% data are testing set, and 70% of data are training set. The size of the dataset is 124 MB. Totally 1000 files are present in this dataset. The sample input images from the dataset to detect lung cancer detection are given in Fig. 2.

Figure 2.

Sample images gathered from the dataset for lung cancer detection.

The collected sample CT images are indicated by $\textit{IM}_{P}^{\textit{Ro}}$ , where $p=1,2,\ldots,P$ and $P$ defines the overall count of collected images.

4.3 CT Image pre-processing

The CT image pre-processing is performed in the implemented lung cancer detection model for enhancing the quality of images without losing any information in the original image. The sample image $\textit{IM}_{P}^{\textit{Ro}}$ is pre-processed to magnify the image quality. The image pixels can be altered into any contrast and density that is needed for noise and blur deduction. The stored images can be extracted easily. Weighted mean histogram equalization and mean filtering approaches are used for the image pre-processing in the implemented ensemble-based deep learning lung cancer detection model.

4.3.1 Weighted mean histogram equalization [26]

The most beneficial method for improving the brightness of the images is weighted mean histogram equalization. This approach is used for adjusting contrast in image. In this process, the information in images is denoted by close contrast values for improving the global contrast. The input image $\textit{IM}_{P}^{\textit{Ro}}$ is given to the histogram equalization to increase the image contrast. The intensities on the image can be spread evenly based on this histogram values. This approach works well in images with dark and bright foregrounds and normal backgrounds. Histogram makes it possible for areas with low local contrast and provides more contrast to the applied image.

The statistical procedure of weighted mean is computed by multiplying each product’s weight by the quantitative result. A weighted mean is produced by dividing the resultant by the total weights attached to the observations. The arithmetic mean is identical to the weighted mean if all observation weights are the same. The formula for the weighted mean is given below in Eqs (1) and (2), respectively.

$\displaystyle\textit{WM}=\Sigma{a^{l}=1D_{a}}.G_{a}/\Sigma{a^{l}}=G_{a}$ (1) $\displaystyle\textit{WM }=(D_{1}G_{2}+D_{1}G_{2}+D_{1}G_{2}+\ldots+D_{l}G_{l})% /(G_{1}+G_{2}+G_{3}+\ldots+G_{l})$ (2)

Here, WM defines the weighted mean, $D_{l}$ defines the data, and the term $W_{l}$ defines the weight. The resultant contrasted image is $\textit{IM}_{P}^{\textit{WE}}$ .

4.3.2 Mean filtering [27]

After histogram equalization, the resultant image $\textit{IM}_{P}^{\textit{WE}}$ is mean filtered for further pre-processing. The most fundamental smoothing filter is mean filter. Typically, mean filtering is viewed as a convolution operation, here the mask is dragged across the image until every pixel is covered. It is built around a kernel, which symbolizes the size and structure of the neighborhood samples while determining the mean. When more intense smoothing is desired, more kernels are employed. There are several variations on the mean filter, such as threshold averaging, where smoothing is only applied if the dissimilarity between the centre pixel’s average value and the original value is larger than a predetermined threshold. Therefore, it eliminates the blurs from the images, and the noise is also cleaned. The smoothing procedure is identical to low-pass filtering because it replaces the neighborhood average with the centre pixel’s grey level and eliminates the edges. Thus, the blurring and noise reduction have been done via the smoothing filters. Removal of blurs is the frequently used preprocessing step, where the undesirable details are removed sensitively and to bridge gaps in between the curves as well as lines. The mean filter is used as a better image processing tool that is used for a variety of tasks, including reducing Gaussian noise, blurring images to remove blemishes and fine lines, and thresholding the images to remove small details. In the process of segmenting images, the mean is also employed as a textural or derived feature. After the mean filtering process, the resultant pre-processed image obtained is denoted as $\textit{IM}_{P}^{\textit{MN}}$ .

5. Soft computing-based lung nodule segmentation using adaptive fuzzy clustering and optimized region growing

In this section, the brief explanation of the developed IHHO algorithm is further explored in sub-Section 4.1. Moreover, the algorithm for the IHHO is provided, and also the flowchart for the suggested IHHO algorithm is shown in Sub Section 4.1. Then, the lung module segmentation takes place using the pre-processed image. Moreover, the adaptive fuzzy clustering, as well as the region growing approach, is utilized for segmenting the lung nodules in sub Section 4.2. The output images that are obtained by the preprocessing and segmentation are visualized in Subsection 4.3.

5.1 Improved harris hawks optimization

An IHHO is proposed in the implemented ensemble-based deep learning model to detect lung cancer. It is used for optimizing the parameters such as seed size from adaptive region growing and maximum iteration from adaptive fuzzy clustering are optimized during segmentation to improve the segmentation performance. Moreover, the parameters like epochs from VGG16, ResNet, and GRU are tuned, and also the hidden neuron count from ResNet and GRU are optimized by the IHHO for maximizing the accuracy and precision of the lung cancer detection model. The HHO approach can manage high dimensional data. When compared with other optimization techniques, HHO provides high performance and provide better optimization results. Still, HHO has some drawbacks like population diversity, local optima and feature selection problems. To rectify these issues, the IHHO algorithm is proposed. By using this IHHO algorithm, the precision and accuracy of the suggested lung cancer identification using ensemble learning framework is maximized. The IHHO is developed in accordance with the current fitness function, worst fitness function and also the current fitness function. The value of $j$ is updated by Eq. (3).

$\displaystyle j=\frac{\textit{wfit}-\textit{fit}[j]}{\textit{bfit}}$ (3)

The variable $j$ represents the random number, wfit is the worst fitness and bfit is the best fitness. The term fit defines fitness. With the help of updated random value, the developed IHHO provides a higher optimization solution than the conventional HHO.

5.1.1 HHO

HHO [28] is a population-based optimization approach, and it is formulated based on the behavior of nature. It is a gradient-free optimization technique. The surprise pounce chasing technique and cooperative attitude of Harris’ hawks act as an important source of inspiration for HHO. Based on the static natural situations and the prey’s flying movements, Harris hawks can pursuit a numerous strategies.

5.1.2 Exploration phase

By the powerful eyes of Harris’ hawk, it can find and identify the prey. But, the prey cannot be seen periodically. In order to find prey, the hawks will observe, detect and wait for the prey, in desert areas for hours. The Harris’ hawks rest at various locations, and it will find prey based on the two strategies. For all resting strategies, an equal chance for $j$ is given. The Harris’ hawks may consider the position of family members and rest accordingly. For the condition of $j<0.5$ , a rabbit is considered in Eq. (4) and also for the condition of $j\geqslant 0.5$ , and the hawks resting in tress are considered in Eq. (4).

$\displaystyle Z(s+1)=\left\{{{\begin{array}[]{ll}\textit{Zrly}(s)-p_{1}|% \textit{Zrly}(s)-2p_{2}Z(s)|&{j\geqslant 0.5}\\ (\textit{Zrab}(s)-\textit{Zn}(s))-p_{3}(\textit{HE}+P_{4}(\textit{RE }-\textit% {HE}))&{j<0.5}\\ \end{array}}}\right.$ (4)

In the above equation, $Z(s+1)$ defines the position of hawks in the upcoming iteration $s$ , the location of the rabbit is $\textit{Zrab}(s)$ . The term $p_{1},p_{2}p_{3,}p_{4}$ defines the Hawks with a position vector $Z(s)$ . The term $j$ defines the random numbers inside $(0,1)$ . The variables HE and RE is defined as upper and lower bounds respectively. The term $\textit{Zrly}(s)$ represents a random hawk. The average position of a hawk is defined by $Z_{n}$ .

Based on the other hawks and different locations, the solutions are generated by the first rule. In the second rule, a random scaling coefficient is considered to find different locations. By using Eq. (5), the average position of the hawk is achieved, and it is given in Eq. (5).

$\displaystyle Z_{n}(s)=\frac{1}{M}\sum\limits_{a=1}^{M}{Z_{a}}(s)$ (5)

Here, $Z_{a}(s)$ defines the position of all hawks in iteration $s$ . Total number of hawks is defined as $M$ .

5.1.3 The transition of exploration into exploitation phase

The exploration phase has the ability to change into the exploitation phase in the HHO algorithm. Based upon the prey’s energy to run away from the hawk will lead to a change the functioning of the explorative phase. When the prey tries to run away from the hawk, the prey’s energy level will decrease. Therefore, the energy level of prey is considered in Eq. (6).

$\displaystyle G=2G_{0}\left(1-\frac{s}{S}\right)$ (6)

In the above equation, $G$ defines the prey’s energy level to runaway, $G_{0}$ defines the initial energy level of the prey. The term $S$ is represented by the number of iterations. For all iterations, the value of $G_{0}$ alters inside the range of $\left[{-1,1}\right]$ . Strength of the rabbit decreases, when the $G_{0}$ decreases from range $\left[{0,1}\right]$ . Similarly, the strength of rabbits decreases when the value of $G_{0}$ increases from the range $\left[{0,1}\right]$ . According to the iterations, the active runaway energy level $G$ decreases. The exploration phase takes place when $\left|G\right|\geqslant 1$ . Likewise, the exploitation phase takes place when $\left|G\right|<1$ .

5.1.4 Exploitation phase

The prey that is found from the exploration phase faces the surprise pounce from the hawk’s attack. When the prey tries to escape, various chasing modes will occur. There are four chasing modes are used in the proposed HHO algorithm.

Assume the prey’s chance of runaway is $k$ . If the prey can able to run away from the surprise pounce is $(k<0.5)$ but the prey cannot able to escape means $(k\geqslant 0.5)$ . Depending on the prey’s energy level, the hawks will surround the prey from various directions by a soft or hard manner. Hawks will execute the surprise pounce by getting closer to the prey and also kill the rabbit at the same time. In HHO, to process the hard and soft manner of the hawk variable $G$ is used. When $\left|G\right|<0.5$ then, surrounding the prey in a hard manner will occur. Similarly, if $\left|G\right|\geqslant 0.5$ then, surrounding the prey in a soft manner will occur.

5.1.5 Soft encircling

When the rabbit attempt to escape from the hawk but, it cannot able to escape and it still has the energy to escape, then consider $\left|G\right|\geqslant 0.5$ and $(k\geqslant 0.5)$ , respectively. In this situation, the hawks do the surprise pounce by surrounding the rabbit in a soft manner. The following Eqs (7) and (8) defines the behavior of hawk.

$\displaystyle Z(s+1)=\Delta Z(s)-G\left|{LZ_{\textit{rab}}(s)-Z(s)}\right|$ (7) $\displaystyle\Delta Z(s)=Z_{rab}(s)-Z(s)$ (8)

Here, the difference between the rabbit and position vector is represented in the term $\Delta Z(s)$ .

$\displaystyle L=2(1-p_{5})$ (9)

The variable $L$ represents the jumping power of the rabbit, and the values alter at all the iterations in this stage. The term $p_{5}$ refers to a random value between $(0,1)$ .

5.1.6 Hard encircling

When the prey does not have the energy to escape and the same time, the hawk will handle the hard procedure to surround the prey, then $k\geqslant 0.5$ and $\left|G\right|<0.5$ , respectively. Then the current positions are upgraded by using Eq. (10).

$\displaystyle Z(s+1)=Z_{rab}(s)-G\left|{\Delta Z(s)}\right|$ (10)

5.1.7 Soft encircling with fast dives

At the moment $\left|G\right|\geqslant 0.5$ , but $k<0.5$ then the rabbit has a sufficient amount of energy to escape from the hawk. Yet, the hawk surrounds the prey in a soft manner. When compared with the hard encircling approach, soft encircling with fast dives is the better method.

In the HHO algorithm, the Levy Flight (LF) concept, various run away methods of prey and leapfrog movements are used. The hawks do more dives in order to catch the prey. They choose the best dive to surround the prey in terrific situations. Hence, the hawks calculate the next move to encircle the prey in a soft manner. This calculation is shown in Eq. (11).

$\displaystyle X=Z_{\textit{rab}}(s)-G\left|{LZ_{\textit{rab}}(s)-Z(s)}\right|$ (11)

The hawks evaluate their previous dive whether it is good enough to catch prey, if not it start doing twisted dives to catch the prey. Therefore, the hawk’s dive using the LF-based model. It is given in the following Eq. (12).

$\displaystyle U=X+T\times\textit{LFN}(E)$ (12)

Here, the term $E$ defines the problem’s dimension and a vector $T$ by the size $1\times E$ is defined. The levy flight function is referred as LFN, and it is computed by Eq. (13).

$\displaystyle\textit{LFN}(z)=0.01\times\frac{b\times\omega}{\left|c\right|^{% \frac{1}{\alpha}}},\omega=\left({\frac{\Phi(1+\alpha)\times\sin(\frac{\pi\beta% }{2})}{\Phi(\frac{1+\alpha}{2})\times\alpha\times 2\left(\frac{\alpha-1}{2}% \right)}}\right)^{\frac{1}{\alpha}}$ (13)

In the above equation $b, c$ is, the random values inside $(0,1)$ , $\alpha$ defines the constant value that is set to $1.5$ in default. Consequently, the plan of action for upgrading the hawk positions throughout the soft encircling phase can be carried out by Eq. (14).

$\displaystyle Z(s+1)=\left\{{{\begin{array}[]{ll}X\hfill&{\textit{if}\ H(X)<H(% Z(S))}\hfill\\ U\hfill&{\textit{if}\ H(U)<H(Z(s))}\hfill\\ \end{array}}}\right.$ (14)

Here, the value of $Z$ and $X$ are acquired from Eqs (11) and (12).

5.1.8 Hard encircling with fast dives

The rabbit cannot escape at $\left|G\right|<0.5$ , and $k<0.5$ , thus, a hard encircling is performed before the surprise attack to capture and kill the prey. Hawks attempts to close the gap between their usual location and the prey. Hence, in a heavy encircling situation, the following Eq. (15) is followed:

$\displaystyle Z(s+1)=\left\{\begin{array}[]{ll}X\hfill&{\textit{if}\ H(X)<H(Z(% S))}\hfill\\ U\hfill&{\textit{if}\ H(U)<H(Z(s))}\hfill\\ \end{array}\right.$ (15)

Here, the value of $Z$ and $X$ are acquired by implementing new rules in Eqs (16) and (17), respectively.

$\displaystyle X=Z_{\textit{rab}}(s)-G\left|{\textit{LZ}_{\textit{rab}}(s)-Z_{n% }(s)}\right|$ (16) $\displaystyle U=X+T\times\textit{LFN}(E)$ (17)

In the above equation $Z_{n}(s)$ is computed using Eq. (5). For the upcoming iterations $X$ and $U$ are the new locations. The pseudocode of the proposed IHHO is given in Algorithm 1. The flowchart for the proposed IHHO is shown below in Fig. 3.

Algorithm 1: Designed IHHO
Set the population size $M$ and maximum number of iterations $S$
Set the random population $Z_{j}\left({j=1,2,.....M}\right)$ with the improved concept
While (ending condition is not met)
Compute the hawks fitness value
For (every hawk $\left({Z_{j}}\right))$
Set the rabbit location as $Z_{\textit{rab}}$
Upgrade the initial energy $G_{0}$ and jumping strength $L$
$G_{0}=2$ and $(-1)$ , $L=2(1-))$
Upgrade the value of $G$ by Eq. (6).
If $\left({\left\|G\right\|\geqslant 1}\right)$ then
Upgrade the position by Eq. (4)
If $\left({\left\|G\right\|<1}\right)$
If $(k\geqslant 0.5)$ , and $\left({\left\|G\right\|\geqslant 0.5}\right)$
Position is upgraded by. (7)
Else If $(k\geqslant 0.5)$ , and $\left({\left\|G\right\|<0.5}\right)$
Position is upgraded by Eq. (9)
Else If $(k<0.5)$ , $\left({\left\|G\right\|\geqslant 0.5}\right)$
Position is upgraded by Eq. (13)
Else If $(k<0.5)$ and $\left({\left\|G\right\|<0.5}\right)$ then
Position is upgraded by Eq. (14)
Return the value of $(X_{\textit{rab}})$

Figure 3.

Flowchart of the proposed IHHO.

5.2 Adaptive Lung Nodule Segmentation

Lung Nodule Segmentations are done in the developed lung cancer detection models for splitting the lung CT images into several image segments. Segmentation is used to rationalize the portrayal of an image to analyze it easier. It is also used to identify the curves and lines in images. In this ensemble learning-based deep learning lung cancer detection model, adaptive fuzzy clustering and region growing techniques are used for segmentation.

5.2.1 Adaptive fuzzy clustering [29]

Splitting the data into equivalent areas according to the likelihood of the items is known as clustering. Data that is logically and physically comparable is combined in order to decrease the number of disc accesses and maximize the efficiency of the database system. The clustering approach involves grouping the $b$ feature vectors into $F$ clusters, with $L^{F}$ serving as the centre of each cluster. The pre-processed image $\textit{IM}_{P}^{\textit{MN}}$ is given to the adaptive fuzzy clustering process for segmentation.

Domains like fuzzy identification and pattern recognition have adopted in the fuzzy clustering method. Many different fuzzy clustering techniques are initiated, and the majority of them are based on distance standards. Computation of fuzzy weights using reciprocal distance is known as FCM, and it is the most popular approach. It requires a set of clusters as input, denoted by $f$ . The FCM cluster’s goal is to lessen the total weighted mean square error, and it is given in Eq. (18).

$\displaystyle A=(X^{b^{f}},L^{\left(f\right)})=\sum\limits_{(b=1,B)}{\sum% \limits_{(f=1,F)}{(X_{\textit{{bf}}})^{i}||t^{\left(b\right)}-l^{(f)}}}||^{2}$ (18)

Each feature vector can belong to many clusters using the FCM, each with a different fuzzy membership value. Final categorization will be done by choosing the highest feature vector’s weight across all clusters. The comprehensive algorithm is given below:

Input is the image pixels that contain three RGB dimensions and $F$ is the amount of clusters. A collection of $F$ clusters that lessen the total distance error is the output.

Every pixel is assigned with a random weight. It makes use of fuzzy positive weight $\{X^{\textit{bf}}\}$ that lies between $[0,1]$ .

The initial weights are equalized using Eq. (19).

$\displaystyle X_{\textit{bf}}\left/\sum\limits_{n=1,F}{X_{\textit{bn}}}\right.$ (19)

For obtaining $X^{\textit{bf}}$ , equalize the weights over $f=1,\ldots,F$ for all $b$ that is indicated in Eq. (20).

$\displaystyle X_{\textit{bf}}=X_{\textit{bf}}\left/\sum\limits_{(n=1,B)}{X_{% \textit{nf}},b=1,\ldots,B}\right.$ (20)

Calculate the new centroid value using Eq. (21).

$\displaystyle L^{\left(f\right)}=\sum\limits_{(b=1,B)}{X_{\textit{bf}}T^{\left% (b\right)}},f=1,\ldots,F$ (21)

Upgrade the weights using Eq. (22).

$\displaystyle X_{bf}=(1/||t^{b}-l^{f}||^{2})^{1/(i-1)}\left/\sum\limits_{(n=1,% F)}{(1/||t^{b}-l^{n}||^{2})}^{1/(i-1)},F=1,.F,b=1,...B\right.$ (22)

Repeat the process from step 3, if any change in input or else stop.

Every cluster is assigned with pixels based upon the highest weight.

Figure 4.

Lung segmentation model using adaptive fuzzy clustering and region growing technique.

Figure 5.

Resultant images after the pre-processing and segmentation process.

After the adaptive fuzzy clustering, the segmented image $\textit{IM}_{f}^{\textit{MN}}$ is obtained.

5.2.2 Region growing [30]

The resultant images $\textit{IM}_{f}^{\textit{MN}}$ are further segmented by region growing technique. In the region growing technique, the regions are combined iteratively, starting with individual pixels or another starting segmentation, and grows until every pixel has been analyzed. It is explained in the following steps:

1)
Create pattern cells from the entire image.
2)
Using a similarity metric, every pattern cell is compared to the neighboring cells to check if they are similar. If the cells are similar, combine the cells to create a fragment and upgrade the property used for comparison.
3)
Examine every neighbor of the fragment until there are no more joinable regions. Therefore, the completed segment of the fragment is tagged.
4)
Continue in this manner until all cells have tags, and then go to the next blank cell.

The resultant image from the region’s growing process is defined as $\textit{IM}_{r}^{\textit{MN}}$ . The diagrammatic representation for the lung segmentation model using adaptive fuzzy clustering and region growing technique is shown in Fig. 4.
5.3 Preprocessing and lung nodule segmentation output

The resultant images obtained after the pre-processing and segmentation process is shown in below Fig. 5.

6. Intelligent lung cancer detection by optimized ensemble-based deep learning model

In Section 5, the detailed description that is related with deep learning models like VGG-16, ResNet, and GRU are discussed in Subsection 5.1. However, the developed ensemble based deep learning model is explained in Subsection 5.2.

6.1 Description of all ensemble models

Ensemble models uses various modeling algorithms to provide a final outcome. It collects the results of each algorithmic models and gives the final predicted results. Ensemble-based deep learning models are used to predict lung cancers in the developed lung cancer identification model, which incorporates VGG-16, ResNet and GRU models to formulate the ensemble learning model. The segmented image $\textit{IM}_{r}^{\textit{MN}}$ is given to the input of the EDLM-based detection process.

6.1.1 VGG-16 [31]

The VGG-16 network contains 16 various convolution layers, and it visualizes a tiny 3 $\times$ 3 respective fields. The layers present in the VGG-16 are arranged in a sequential manner. It has one single layer for both input and output. In the input layer, the features, signals and images are sent as an input and changed into 3 dimensional and 2 dimensional. The input is applied to the convolutional layer, which segregates the images by different convolutional filters. Convolution of the images is done by transferring or moving the filters vertically or horizontally. The Rectified Linear Unit (ReLU) executes the threshold function in all the elements that are present in the image, and also it activates all the hidden layers. Different rectangle shaped pooling segments are classified from the input images by the max pooling layer. It is a $2\times 2$ size layer. The dropout layer assigns the value of zero to the elements present in the input images. A bias vector and weight matrix are included for the input images to multiply the inputs in the fully connected layer. In VGG-16, the softmax layer is a final layer; a softmax classifier is applied to all the input signals or images. Finally, the prediction score based on lung cancer is obtained.

6.1.2 Residual Network (ResNet) [32]

The production of every residual layer is combined with the input of the next layer. In ResNet, the gradients can flow back, enables fast training and add many more layers. The convolutional and identity blocks make up the two fundamental building components of the ResNet model. The identity block has equal proportions. Identity blocks have three elements. The initial element is a 2D convolutional layer, and the filter size is $\left({1\times 1}\right)\left({1,1}\right)$ . The ReLU function is utilized for nonlinear function, while BN is used to normalize the channel axis. The second uses filters with a large size. The third element lacks ReLU’s activation function. The ReLU activation function is then applied after combining the inputs and shortcut connections.

Figure 6.

Developed EDLM-based lung cancer detection model.

Figure 7.

Performance evaluation of the developed ensemble learning-based deep learning lung cancer detection model with diverse algorithms regarding “(a) Accuracy, (b) F1-Score, and (c) Precision”.

Figure 8.

Performance evaluation of the developed ensemble learning-based deep learning lung cancer detection model among different existing classifiers in terms of “(a) Accuracy, (b) F1-Score, and (c) Precision”.

Figure 9.

5-fold analysis of the designed ensemble learning-based deep learning lung cancer detection model with existing algorithms in terms of “(a) Accuracy, (b) F1-Score, and (c) Precision”.

In a convolutional block, the proportions of output and input are not the same. Linear projections are performed by shortcut connections to alter the proportions. The structure of the convolutional block is similar to the identification block but in the convolutional block, a 2D convolutional layer is added in a shortcut way. In this way, the size of the input changes in order to match the main path. At last, the altered shortcut is merged with the resultant output from the main path. The disappearing gradients are controlled in the modified shortcut, and it is the most beneficial process of the convolution block. Finally, the lung cancer prediction score is obtained at the final layer.

6.1.3 Gated Recurrent Unit (GRU) [33]

The upgraded version of the Long Short-Term Memory (LSTM) neural network is known as the GRU neural network model. It optimizes the LSTM’s network structure, and it reduces three gating units into two gating units of LSTM. The reset and update gates are the two gating units. Moreover, the characterization of the GRU model is low, and also it maintains the information reliance over a long distance. GRU model is composed of three layers such as implicit layer, input layer and output layer. The collection of GRU neurons forms the hidden layer. Time series data are the input of the GRU model. If the input is $(u\_1,u\_2,\ldots,u\_T)$ , then the formulas for standard GRU with reset and update gate in regards to time $T$ is calculated by the following Eq. (23).

$\displaystyle p_{T}=\lambda(Y_{p}\times[f_{T-1},u_{T}])$ (23)

$\displaystyle v_{T}=\lambda(Y_{v}\times[f_{T-1},u_{T}])$ (24)

$\displaystyle k_{T}=\tan f(Y_{\times}[p_{T}\times f_{T-1},u_{T}])$ (25)

$\displaystyle f_{T}=(1-v_{T})\times f_{T-1}+v_{T}\times k_{T}$ (26)

$\displaystyle w_{T}=\lambda(Y_{0}\times f_{T})$ (27)

Here, at the time $T$ the output of the reset gate is $p_{T}$ , the term $v_{T}$ defines the output in accordance with time $T$ . The terms $Y_{p}$ and $Y_{v}$ represents the weight between $f_{T-1}$ and input in the reset and update gate, respectively. The term $u_{T}$ defines the input. New value vector created in the $\tan f$ layer is defined as $k_{T}$ . The term $f_{T}$ defines the output of standard GRU unit. The variable $v\_T$ defines the output gate at a time. Moreover, the variable $Y$ represents the output of the update gate. However, the term $Y_{0}$ is represented as the weight, and also the sigmoid activation function is termed as $\tan f$ .

Hence, the GRU model reduces the gating units to preserve the main information and deduct the unnecessary information. For storing the information, it uses the hidden layer. Then, the lung cancer detection score is attained at the output layer.

6.2 Developed ensemble-based deep learning model

The proposed ensemble-based deep learning model is used to detect lung cancer that may prevent the people from death. Some deep learning approaches that are used in the suggested lung cancer detection model are is VGG-16, ResNet and GRU. In this ensemble approaches, the parameters like epochs in VGG-16, hidden neuron count and epochs from ResNet, hidden neuron count and epochs from GRU are tuned by the IHHO algorithm. VGG16 is useful for detecting diseases from images. It consumes more time to get trained, and also it needs more bandwidth and space. Object separation and identification of images is done by using ResNet. More number of layers can be trained easily. It consumes more power. GRU takes less amount of time to get trained. Modifying GRU is simple. Learning efficiency is low in GRU. The objective function of this parameter optimization is the maximization of accuracy and precision. Here, the objective function is given in the Eq. (28).

$\displaystyle\textit{OB}=\mathop{\arg\min}\limits_{\left\{{\textit{MX}_{a}^{% \textit{FCM}},\textit{SD}_{b}^{\textit{RG}},\textit{ES}_{i}{}^{\textit{VGG}},}% \right.\left.{\textit{HC}_{j}^{\textit{RS}},\textit{ES}_{k}^{\textit{RS}},% \textit{HC}_{l}^{\textit{CRU}},\textit{ES}_{m}^{\textit{GRU}}}\right\}}\left({% \frac{1}{\textit{Acy}}+\frac{1}{\Pr}}\right)$ (28)

In the above equation, accuracy is defined as Acy, and precision is defined as $\Pr$ . The term $\textit{MX}_{a}^{\textit{FCM}}$ defines the maximum iterations in FCM in the range between $\left[{50,\,\,100}\right]$ . The term $\textit{SD}_{b}^{\textit{RG}}$ indicates the seed range in the region growing technique that lies between $\left[{200,\,\,255}\right]$ . The term $\textit{ES}_{i}^{\textit{VGG}}$ represents the epochs from VGG-16 that lies between $\left[{50,\,\,100}\right]$ , $\textit{HC}_{j}^{\textit{RS}}$ is the hidden neuron count from ResNet in the interval of $\left[{5,\,\,255}\right]$ , $\textit{ES}_{k}^{\textit{RS}}$ represents the epochs from ResNet between the ranges of $\left[{50,\,\,100}\right]$ , the term $\textit{HC}_{l}^{\textit{CRU}}$ indicates the hidden neuron count in GRU that lies in the range of $\left[{5,\,\,255}\right]$ and the term $\textit{ES}_{m}^{\textit{GRU}}$ defines the epochs in GRU that lies between $\left[{50,\,\,100}\right]$ . The accuracy is expressed in Eq. (29).

$\displaystyle\textit{Acy}=\frac{\textit{TPS}+\textit{TNG}}{\textit{TPS}+% \textit{TNG }+\textit{FPS }+\textit{FNG}}$ (29)

Here, the term TPS defines the true positive value, the term TNG refers the true negative, FPS defines the false positive, and the term FNGdefines the false negative value. The formula for precision is given in Eq. (30). The schematic diagram of the designed EDLM-based lung cancer detection model is depicted in Fig. 6.

$\displaystyle\textit{pr}=\frac{\textit{TNG}}{\textit{TPS}+\textit{FPS}}$ (30)

7. Results and discussions

The result section shows the description related with the implemented platform, and also the various evaluation metrics are evaluated and described in all sections.

7.1 Experimental setup

Figure 10.

5-fold evaluation of the developed ensemble learning-based deep learning lung cancer detection model among various existing techniques in terms of “(a) Accuracy, (b) F1-Score, and (c) Precision”.

Table 2

Performance analysis of the designed ensemble learning-based lung cancer detection model

Comparison among various algorithms
Terms	DHOA-EDLM [34]	TSA- EDLM [35]	SSA- EDLM [36]	HHO- EDLM [28]	IHHO-EDLM
Precision	95.49669	95.56869	95.63492	96.04222	96.68874
F1-Score	95.43349	95.5629	95.63492	96.16909	96.62475
Accuracy	95.4	95.5	95.6	96.13333	96.6
Specificity	95.43011	94.43011	95.56452	95.96774	96.9978
Sensitivity	95.37037	94.37037	95.63492	96.2963	96.56085
MCC	90.79959	90.9123	91.19944	92.26633	93.19972
FDR	4.503311	4.32152	4.365079	3.957784	3.945788
NPV	95.30201	95.6920	95.56452	96.22642	96.51007
FNR	4.62963	4.56251	4.365079	3.703704	3.939153
FPR	4.569892	4.369891	4.435484	4.032258	4.360215
Comparison among various classifiers
Measures	CNN [20]	SVM [19]	VGG16 [20]	VGG_Resnet_GRU [37]	IHHO-EDLM
Precision	80.52632	86.18858	85.31746	85.99736	96.6578
F1-Score	80.73879	86.01723	85.31746	86.0542	96.72475
Accuracy	80.53333	85.93333	85.2	85.93333	96.6
Specificity	80.10753	86.02151	85.08065	85.75269	96.6880
Sensitivity	80.95238	85.84656	85.31746	86.11111	96.7024
MCC	61.06338	71.86634	70.39811	71.86463	93.1986
FDR	19.47368	13.81142	14.68254	14.00264	3.42190
NPV	80.54054	85.67604	85.08065	85.8681	96.5214
FNR	19.04762	14.15344	14.68254	13.88889	3.46705
FPR	19.89247	13.97849	14.91935	14.24731	3.5428

The ensemble learning-based deep learning model for lung cancer identification was implemented using Python. The outcome of the designed ensemble-based deep learning model for the identification of lung cancer was evaluated ‘among various algorithms to ensure the progress of the recommended algorithm. The experiments are organized with a population rate of 10, chromosome length of 7 and maximum iteration of 25. The optimization algorithms like Deer Hunting Optimization Algorithm (DHOA) [34], Tree–Seed Algorithm (TSA) [35], Salp Swarm Algorithm (SSA) [36] and HHO [28] and techniques like CNN [20], SVM [19], VGG16 [31], and VGG_ResNet_GRU [37] are used for the evaluation process.

7.2 Evaluation metrics

The evaluation metrics are performed to improve the performance of the developed model as follows.

$\displaystyle\text{F1-measure}=\frac{2\textit{TPS}}{2\textit{TPS}+\textit{FPS}% +\textit{FNG}}$ (31) $\displaystyle\text{Specificity}=\frac{\textit{TPS}}{\textit{TPS}+\textit{FNG}}$ (32) $\displaystyle\text{Sensitivity}=\frac{\textit{TPS}}{\textit{TPS}+\textit{FPS}}$ (33) $\displaystyle\textit{NPV}=\frac{\textit{TNG}}{\textit{TNG}+\textit{FNG}}$ (34) $\displaystyle\textit{FPR}=\frac{\textit{TPS}}{\textit{FPS}+\textit{TNG}}$ (35) $\displaystyle\textit{FDR}=\frac{\textit{FPS}}{\textit{FPS}+\textit{TNG}}$ (36) $\displaystyle\textit{FNR}=\frac{\textit{TNG}}{\textit{TNG}+\textit{FNG}}$ $\displaystyle\textit{MCC}=\frac{\textit{TPS}\times\textit{TNG}-\textit{FPS}% \times\textit{FNG}}{\sqrt{\begin{array}[]{l}\left({\textit{TPS}+\textit{FPS}}% \right)\left({\textit{TPS }+\textit{FNG}}\right)\\ \left({\textit{TNG}+\textit{FPS}}\right)\\ \end{array}}\left({\textit{TNG}+\textit{FNG}}\right)}$ (37)

7.3 Effectiveness analysis using learning percentage

The performance evaluation of the designed IHHO-ELDM-based lung cancer detection model is shown in below Fig. 7. As well as, Fig. 8 shows the performance evaluation of the implemented ensemble learning-based deep learning lung cancer detection model among various existing techniques. The accuracy of the proposed EDLM-based detection of lung cancer model has improved than HHO, SSA, TSA, and DHOA with 0.9%, 1.36%, 1.89% and 2.0% at a learning percentage of 60. Moreover, the graph analysis shows equivalence performance. Here, the existing HHO-EDLM algorithm attains second better performance concerning accuracy. However, it tends to enhance the system performance. Additionally, the TSA-EDLM algorithm achieves low performance in terms of accuracy, precision and F1 score. Moreover, it degrades the system’s performance. Thus, the training of the network becomes high, and also it is unsuitable for effectively detecting lung cancer. While considering the classifiers, the SVM model provides a second better performance. On the contrary, CNN does not provide effective performance other than the existing models. Therefore the proposed model has provided a better performance rate in accordance with precision, accuracy and f1-score when comparing with other algorithms.

7.4 Effectiveness analysis using 5 fold

The 5-fold evaluation of the offered ensemble learning-based lung cancer detection model among various algorithms is given below in Fig. 9. As well as, Fig. 10 shows the 5-fold evaluation of the suggested ensemble learning-based lung cancer detection model with diverse existing approaches. The precision of the developed EDLM-based lung cancer detection model is enhanced than VGG16-ResNet-GRU, VGG16, SVM and CNN with 2.7%, 3.9%, 3.09% and 5.02% with the 5-fold value of 5. In K-fold analysis, the given data is splitted into five datasets. While considering the accuracy based analysis, the TSA-EDLM shows a second better performance in terms of 1-fold. The developed method is performed well than the conventional prediction model by analyzing the precision, F1-score and accuracy.

7.5 Overall analysis of the developed lung cancer detection model

The overall computational analysis of the developed ensemble learning-based deep learning detection for lung cancer is shown in Table 2. The suggested ensemble learning-based lung cancer detection model has given improved F1-Score of 0.5%, 1.0%, 1.77% and 1.2% than HHO-EDLM, SSA- EDLM, TSA- EDLM and DHOA-EDLM. Furthermore, the developed method attains high performance than prior algorithms.

8. Conclusion

An ensemble learning-based lung cancer detection using CT images was implemented to identify lung cancers for humans at an early stage. For performing this, a set of sample CT images were collected from a traditional database. Here, the input images were pre-processed to enhance the image quality. In this model, preprocessing has been done by using two approaches like, weighted mean histogram equalization and mean filtering. Then the preprocessed images were segmented using adaptive fuzzy clustering and region growing techniques, where the parameter optimization taken place via the IHHO. The outputs of the segmented images were then detected using the ensemble-based classifiers such as VGG-16, ResNet and GRU. Moreover, parameters like epochs in VGG-16, hidden neuron count and epochs in ResNet, hidden neuron count and epochs in GRU were tuned using IHHO. These classifiers gave the predicted outcome, and the values were averaged to get the final output. The result of the developed model has been ensured among various classifiers, and the outcome was shown that the proposed method attained with 12.3%, 13.3%, 12.4% and 19.79% higher F1-score than VGG_Resnet_GRU, VGG16, SVM and CNN. Therefore, the experimental outcomes show the developed ensemble-based lung cancer detection model has high performance with respect to accuracy and precision among various algorithms and classifiers. The scientific explanation for attaining the best performance of the developed model is shown below. The designed IHHO-EDLM model tends to optimize complex parameters in order to improve system performance. Thus, it can to increase the accuracy rate in the lung cancer detection model. Owing to these, the radiologist can provide the appropriate treatment for the patients who are suffering from lung cancer. Moreover, the result shows that the developed model provides the significant performance when compared with the existing approaches. In future work, a novel deep learning technique will be adopted to evaluate lung cancer detection using PET/MRI images.

References

Zheng

Guo

Cui

Veldhuis

RNJ

Oudkerk

van Ooijen

PMA

. Automatic Pulmonary Nodule Detection in CT Scans Using Convolutional Neural Networks Based on Maximum Intensity Projection. IEEE Transactions on Medical Imaging. 2020; 39(3): 797-805.

Sun

Bauer

Beichel

. Automated 3-D Segmentation of Lungs With Lung Cancer in CT Data Using a Novel Robust Active Shape Model Approach. EEE Transactions on Medical Imaging. 2012; 31(2): 449-460.

Zhou

Wang

. Deep Learning Assisted Predict of Lung Cancer on Computed Tomography Images Using the Adaptive Hierarchical Heuristic Mathematical Model. IEEE Access. 2020; 8: 86400-86410.

Venkatesh

Bojja

. A Dynamic Optimization and Deep Learning Technique for Detection of Lung Cancer in CT Images and Data Access Through Internet of Things, Wireless Pers Communication, 2022.

Halder

Dey

Sadhu

. Lung Nodule Detection from Feature Engineering to Deep Learning in Thoracic CT Images: a Comprehensive Review. J Digit Imaging. 2020; 33: 655-677.

Chamberlin

Kocher

Waltz

, et al., Automated detection of lung nodules and coronary artery calcium using artificial intelligence on low-dose CT scans for lung cancer screening: accuracy and prognostic value. BMC Medicine. 2021; 19(55).

Kavitha

Shanthini

Sabitha

. ECM-CSD: An Efficient Classification Model for Cancer Stage Diagnosis in CT Lung Images Using FCM and SVM Techniques. J Med Syst. 2019; 43(73).

Ren

Cao

, et al., Diagnostic accuracy of computed tomography imaging for the detection of differences between peripheral small cell lung cancer and peripheral non-small cell lung cancer. Int J Clin Oncol. 2017; 22: 865-871.

Wang

Shen

. Lung Nodule Detection in CT Images Using a Raw Patch-Based Convolutional Neural Network. J Digit Imaging. 2019; 32: 971-979.

10.

Makaju

Prasad

PWC

Alsadoon

Singh

Elchouemi

. Lung Cancer Detection using CT Scan Images. Procedia Computer Science. 2018; 125: 107-114.

11.

Ruan

Meng

. Development of Deep Learning-based Automatic Scan Range Setting Model for Lung Cancer Screening Low-dose CT Imaging, Academic Radiology, 2022.

12.

Indumathi

Vasuki

. Segmentation of Lung Cancer from CT Image- A Comparative Analysis. Material Today: Proceedings. 2021; 47: 56-58.

13.

Akter

Moni

Islam

. Lung cancer detection using enhanced segmentation accuracy. Appl Intell. 2021; 51: 3391-3404.

14.

Zhang

Yang

Gong

. An Appraisal of Nodule Diagnosis for Lung Cancer in CT Images. J Med Syst. 2019; 43: 181.

15.

Jose

Chithara

Nirmal Kumar

. et al., Automatic Detection of Lung Cancer Nodules in Computerized Tomography Images. National Academy Science Letters. 2017; 40: 161-166.

16.

Muthazhagan

Ravi

Rajinigirinath

. An enhanced computer-assisted lung cancer detection method using content based image retrieval and data mining techniques. Journal of Ambient Intelligence and Humanized Computing, 2020.

17.

Majidpourkhoei

Alilou

Majidzadeh

. et al., A novel deep learning framework for lung nodule detection in 3d CT images. Multimedia Tools and Applications. 2021; 80: 30539-30555.

18.

Mohamed Shakeel

Burhanuddin

Desa

. Lung cancer detection from CT image using improved profuse clustering and deep learning instantaneously trained neural networks. Measurement. 2019; 145: 702-712.

19.

Alzubaidi

Otoom

Jaradat

. Comprehensive and Comparative Global and Local Feature Extraction Framework for Lung Cancer Detection Using CT Scan Images. IEEE Access. 2021; 9: 158140-158154.

20.

Sori

Feng

, Godana, DFD-Net: lung cancer detection from denoised CT scan image using deep learning. Front. Comput. Sci. 2021; 15: 152701.

21.

Karthiga

Rekha

. Feature extraction and I-NB classification of CT images for early lung cancer detection. Material Today Proceedings. 2020; 33: 3334-3341.

22.

Shakeel

Burhanuddin

Desa

. Automatic lung cancer detection from CT image using improved deep neural network and ensemble classifier, Neural Comput & Application, 2020.

23.

Manickavasagam

Selvan

. Automatic Detection and Classification of Lung Nodules in CT Image Using Optimized Neuro Fuzzy Classifier with Cuckoo Search Algorithm. J Med Syst. 2019; 43(77).

24.

Halder

Chatterjee

Dey

Kole

Munshi

. An adaptive morphology based segmentation technique for lung nodule detection in thoracic CT image. Computer Methods and Programs in Biomedicine. 2020; 197.

25.

Togacar

Ergen

Comert

. Detection of lung cancer on chest CT images using minimum redundancy maximum relevance feature selection method with convolutional neural networks. Biocybernetics and Biomedical Engineering. 2020; 40(1): 23-39.

26.

https://www.geeksforgeeks.org/histograms-equalization-opencv/ https//www.geeksforgeeks.org/how-to-calculate-the-weighted-mean/.

27.

Rakshit

Ghosh

Uma Shankar

. Fast mean filtering technique (FMFT). Pattern Recognition. 2007; 40: 890-897.

28.

Heidari

Mirjalili

Faris

Aljarah

Mafarja

Chen

. Harris hawks optimization: Algorithm and applications. Future Generation Computer Systems. 2019; 97: 849-872.

29.

Taher

Sammouda

. Lung cancer detection by using artificial neural network and fuzzy clustering methods. IEEE Acess, 2011; 295-298.

30.

Shehab

Fahmy

Gasser

El-Mahallawy

. Satellite Imagery Segmentation: a region growing approach. IEEE Acess. 2021; 33: 404-412.

31.

Jahangeer

GSB

Rajkumar

. Early detection of breast cancer using hybrid of series network and VGG-16. IEEE Acess, 2021; 80: 7853-7886.

32.

Shehab

Fahmy

Gasser

El-Mahallawy

. An Efficient Brain Tumor Image Segmentation Based on Deep Residual Networks (ResNets). Journal of King Saud University-Engineering Sciences, 2020.

33.

Wang

Shi

Liu

. OGRU: An Optimized Gated Recurrent Unit Neural Network. IEEE Acess, 2019; 1325.

34.

Sengupta

Kulkarni

Godbole

. Analytical application of DHOA for the determination of trace metallic constituents in U based fuel materials by ICP-AES. IEEE Acess. 2011; 289: 961-965.

35.

Duh

Sowa

Roberts

Karnes

. Traditional serrated adenomas (TSA) and the company they keep: TSA presence predicts advanced neoplasm states. 2020; 35: 1351-1354.

36.

Gharavi

Skopouli

Patel

Hughes

GRV

Moutsopoulos

. Rheumatoid factors in Ro (SSA) positive and Ro (SSA) negative rheumatoid arthritis patients. IEEE Acess, 1986; 6: 61-62.

37.

Ahmad

Ullah

Ahmad

AL-Sharabi

Ullah

Khan

Rasheed

Ullah

Md. Nasir

Uddin

Md. Sadek

Ali

. A Novel Hybrid Deep Learning Model for Metastatic Cancer Detection. Computational Intelligence and Neuroscience, 2022.

38.

Jose

Chithara

Kumar

Kareemulla

. Automatic detection of lung cancer nodules in computerized tomography images. National Academy Science Letters. 2017; 40: 161-166.

Integration of adaptive segmentation with heuristic-aided novel ensemble-based deep learning model for lung cancer detection using CT images

Abstract

Keywords

1. Introduction

2.1 Related works

Table 1 Merits and demerits of existing lung cancer detection model

3. Motivation

4. Lung cancer detection using CT images: System model and dataset model

4.1 System model: Proposed lung cancer detection

4.3.1 Weighted mean histogram equalization [26]

5. Soft computing-based lung nodule segmentation using adaptive fuzzy clustering and optimized region growing

5.1 Improved harris hawks optimization

5.1.2 Exploration phase

5.1.5 Soft encircling

5.2.1 Adaptive fuzzy clustering [29]

6. Intelligent lung cancer detection by optimized ensemble-based deep learning model

6.1 Description of all ensemble models

6.1.1 VGG-16 [31]

6.1.2 Residual Network (ResNet) [32]

7.1 Experimental setup

7.4 Effectiveness analysis using 5 fold

7.5 Overall analysis of the developed lung cancer detection model

8. Conclusion

References

Table 1
Merits and demerits of existing lung cancer detection model