Introduction

Diagnosing a lung disease [1–3] is one of the most demanding and crucial tasks in recent times for providing earlier treatment to the patients. The medical experts highly prefer an automated disease diagnosis tools for identifying and categorizing the types of diseases [4–6]. According to the recent reviews, it is observed that 1.61 million deaths happen at each year due to the lung cancer. Typically, the pulmonary lung cancer is considered as the dreadful and life-threatening disease highly affecting the people around the world [7–9]. Hence, it should be accurately diagnosed for saving the life of people. In medical sector, the restorative images like Computed Tomography (CT), Magnetic Resonance Imaging (MRI), X-ray, and etc are extensively used for disease diagnosis [10, 11]. Among others, the CT is the most suitable and accurate medical imaging tool that captures the human body organs with minimal noise effects. Due to the in homogeneity of lung nodules [12–14], developing an automated lung segmentation system is treated as one of the complicated tasks. So, most of the conventional works [15–18] developed the different types of segmentation and classification mechanisms for lung cancer identification and nodule segmentation. Yet, it faced various problems [13, 19–22] associated to the parameters of complexity, difficult to understand, over segmentation, reduced accuracy, high mis-prediction rate, and error rate. So, the proposed work motivates to develop an efficient and automated lung nodule segmentation and classification system for detecting the pulmonary diseases from the CT images. The major research objectives of this paper are as follows:

• To preprocess the input lung CT image, the tanh normalization mechanism is applied that eliminates the noise and smoothens the image for further processing.

• To exactly segment the lung nodules from the preprocessed image, the perceptual U-Net segmentation mechanism is utilized.

• To obtain the most relevant features from the segmented region, the Decked Dragonfly Optimization (DDO) technique is implemented, which helps to optimize the feature set.

• To predict the pulmonary lung cancer by using the optimal features, the Speculative Deceptive Network (SDN) based classification algorithm is deployed.

• To test the results and performance of the proposed lung nodule segmentation and disease classification system, various evaluation metrics are used during analysis.

The remaining portions of this paper are segregated into the following sections: Section 2 reviews the conventional segmentation, optimization, and prediction methodologies used for detecting the cancer from the lung CT images. Also, it individually analyzes the benefits and limitations of each model according to its key features and computational operations. Section 3 presents the description about the proposed lung nodule segmentation and classification methodology used for an earlier disease diagnosis. Section 4 validates the results of conventional and proposed lung cancer detection approaches by using various evaluation measures. Finally, the overall paper is summarized with the findings, challenges, and future scope in Section 5.

Related works

This section reviews the conventional image processing techniques used for an accurate segmentation and detection of nodules from the lung CT images. Also, it investigates the benefits and demerits of each technique according to its operating characteristics and principles.

Primakov et al. [23] implemented an automated segmentation system for detecting lung cancer from CT images. The purpose of this work was to accurately identify the tumor slice, location, and size by using an advanced U-Net based Convolutional Neural Network (CNN) classification mechanism. However, the suggested system was difficult to understand, which affects the efficiency of entire system. Wang et al. [24] deployed a Deep Self-paced Active Learning (DSAL) mechanism for detecting pulmonary nodules from 3D thoracic lung images. The purpose of this work was to accurately segment the tumor affected region with reduced error rate and increased segmentation accuracy. Yet, it has the major problems of increased dimensionality of features, reduced convergence speed, and time consumption. Shaziya et al. [25] deployed an automated lung segmentation system from CT images by using the U-Net Convolutional Network method. Also, it intends to attain an increased detection accuracy by accurately segmenting the lung regions. Haq et al. [26] utilized a multi-label Deep Learning Segmentation (DLS) technique for segmenting the cardio-pulmonary substructure from the CT images. Here, the image preprocessing was performed at the initial stage for normalizing the lung image with increased contrast and quality.

Kiser et al. [27] deployed a Pleural effusion segmentation algorithm for detecting the pulmonary disease from the lung CT image. Here, the semi-automated segmentation techniques were used to detect the tumor with better accuracy. For validating the suggested approach, the similarity coefficients and hausdorff distance measures were computed in this analysis. ALzubi et al. [28] utilized a boosted NN ensemble classification mechanism for developing an accurate lung cancer diagnosis framework. Here, the Weight Optimized Neural Network (WONN) technique was utilized to improve the performance of detection system with minimized error and time. In addition to that, it objects to reduce the false positive rate and classification time by accurately categorizing the cancer. Yet, it has the major problems of high processing time, complex computational operations, and lack of scalability. Kriegsmann et al. [29] used a CNN based deep learning mechanism for accurately predicting the small-cell and non-small-cell lung cancer. The contribution of this work was to categorize the different types of lung cancers such as Pulmonary Adenocarcinoma (ADC), small cell lung cancer, and Pulmonary Squamous Cell Carcinoma (SqCC) with increased accuracy. Moreover, the hyper-parameter tuning was performed to tune the parameters for improving the detection performance. The major drawbacks of this work were high computational and cost complexity, which degrades the efficacy of entire cancer prediction system. Abe et al. [30] implemented a simple method for designing a real-time detection system for lung tumor identification and classification. Lin et al. [31] employed a Taguchi parametric optimization technique integrated with the 2D CNN mechanism for detecting lung cancer from the CT image. It was one of the statistical mechanism developed based on an orthogonal array, in which the control factor was calculated to improve the stability and minimize the quality loss. The key merits of this work were fast convergence speed and optimized time consumption.

According to this review, it is analyzed that the conventional works are highly focused on developing an accurate segmentation and classification techniques for automated lung disease diagnosis system. However, it faced some complications associated to the following factors:

1. Over segmentation

2. Complex mathematical operations

3. Reduced detection accuracy

4. High false positives

5. Difficult to understand the system model

Therefore, the proposed work objects to implement an effective medical image processing techniques for the design and development of lung nodule segmentation and classification system.

Proposed methodology

This section presents the detailed description about the proposed automated lung nodule segmentation and tumor classification system. The main contribution of this work is to develop a simple and efficient automated segmentation and classification system for accurately identifying the pulmonary lung cancer from the CT images. Also, it objects to reduce the error rate and mis-prediction outcomes by properly detecting the tumor region from the given input CT image. The main aim of this work is to accurately classify the nodule and non-nodule lung images by using an advanced medical image processing techniques. For this purpose, a combination of mechanisms such as segmentation, optimization, and classification are implemented in this framework, and the overall working flow of the proposed system is shown in Fig. 1. It includes the following stages:

• Tanh normalization and preprocessing

• Lung nodule segmentation using Perceptual U-Net algorithm

• Decked Dragonfly Optimization (DDO)

• Speculative Deceptive Network (SDN) classification algorithm

• Performance analysis

View Full Size

Working flow of the lung cancer detection framework

Citation: Imaging 15, 1; 10.1556/1647.2023.00106

• Download Figure
• Download figure as PowerPoint slide

Initially, the input CT lung image is obtained from the dataset and it is preprocessed by using the tanh normalization mechanism, which helps to improve the overall quality of more efficient of the image by suppressing the noise/artifacts. Then, an advanced Perceptual U-Net segmentation mechanism is implemented to segment the lung region from the filtered image, which helps to increase the accurate diagnosis of disease. After that, the DDO algorithm is employed to select the most relevant features based on the best optimal solution, which helps to reduce the complexity of classifier. The obtained features are properly trained for detecting the pulmonary lung cancer by using the SDN classification mechanism. At the end, the classifier accurately predicts the nodule and non-nodule classes according to the optimized features. The primary advantages of the proposed DDO-SDN based lung disease diagnosis system are follows: Increased convergence rate, reaches the optimal solution with minimum iterations, high accuracy, reduced false outputs, and reduced time consumption.

Decked Dragonfly Optimization (DDO)

After segmentation, the most relevant features are optimally selected from the segmented image by using an advanced DDO technique. The motivation of implementing this algorithm is to choose the optimal features from the lung image for the training the model of classifier. Also, it helps to reduce the false positives, error rate, and increase the detection accuracy with simple computational operations. Moreover, this technique is developed based on the migration and hunting behavior of dragonflies. It resembles the exploration and exploitation operations of the optimization process. The three important behaviors of this model are listed in below:

• Separation

• Alignment

• Cohesion

During separation, the individual particles are segregated from their neighbors to avoid collision as shown in below:

$X_i=-\sum _{k=1}^{N}S-S_k$(7)

Where, $S$ and $S_k$ are the current position of individual populations, $X_i$ indicates the separation parameter, k is the position, and N represents the total number of individuals. During alignment operation, the mean of velocities are computed by using the following model:

$Y_i=\frac{{\displaystyle \sum }_{k=1}^N{Ve}_k}{N}$(8)

Where, $Y_i$ is the alignment parameter, and ${Ve}_k$ represents the velocity of neighborhood individuals. Consequently, the cohesion is computed based on the attraction of individuals towards the center of neighborhood as shown in below:

$Z_i=\frac{{\displaystyle \sum }_{k=1}^N{S}_k}{N}-S$(9)

Where, $Z_i$ is the cohesion parameter. After that, the food source and enemies are identified based on the following models:

${Fs}_i=S^{+}-S$(10)

$E_i=S^{-}-S$(11)

According to the movement of dragonflies, its position is updated based on the factors of step vector and position vector as represented using the following equations:

$∆S_{m+1}=\left({aX}_i+b{Y}_i+{cZ}_i+d{E}_i+e{Fs}_i\right)+w∆S_m$(12)

$∆S_{m+1}=S_m+∆S_{m+1}$(13)

Where, a, b, c, d, and e are defined as the food vectors, m indicates the iteration, and w is the inertia weight factor. Based on the updated position, the best optimal solution is identified for parameter selection. The SDN classifier is trained for each solution using the training dataset, and the misclassification rate—the ratio of the number of incorrectly classified samples to all testing samples—is determined using the testing dataset. The place where the testing error rate is the lowest is where the best solution can be identified. The working flow of the DDO technique is represented in Fig. 2.

View Full Size

Flow of DDO

Citation: Imaging 15, 1; 10.1556/1647.2023.00106

• Download Figure
• Download figure as PowerPoint slide

Speculative Deceptive Network (SDN)

After feature selection, the SDN is used to accurately detect the pulmonary lung cancer from the input CT image according to the optimal features. Typically, various machine learning and deep learning based classification techniques are developed in the existing works for developing an automated disease diagnosis system. For instance, the Support Vector Machine (SVM), Neural Network (NN), Decision Tree (DT), Random Forest (RF), and other deep neural network are extensively used in many medical image processing systems. Still, it faced the problems based on the factors of computational complexity, time complexity, high false outputs, and misprediction rate. Thus, the proposed work aims to use an advanced and accurate classification mechanism for CT lung cancer detection. This classifier is properly trained by using the features obtained from the previous stage. In this model, the network is constructed at first with the set of p inputs and q output. Consider that, the weight values of the network could be w1, w2, and w3. Here, the input is represented by $x\left(d-1\right)$, output of hidden layer could be $r(d)$, structure layer could be $r_g(d)$, and output layer is indicated as q $(d)$. The architecture model of the SDN classification approach is shown in Fig. 3, in which the output is obtained based on the following model:

$r(d)=f\left(w_2{r}_g(d)+w_1(r(d-1))\right)$(14)

View Full Size

SDN architecture

Citation: Imaging 15, 1; 10.1556/1647.2023.00106

• Download Figure
• Download figure as PowerPoint slide

Then, the transfer function of hidden layer is computed by using the following model:

$r(d)={\left(1+e^{-r}\right)}^{-1}$(15)

The transfer function of this architecture is entirely correlated with the linear function as shown in below:

$q(d)=s\left(w_{3}r(d)\right)$(16)

Moreover, the error value $err$ of network is computed according to the obtained weight factor as shown in below:

$err={\displaystyle \sum }_{d=1}^n{\left(U_r-q_r\right)}^2$(17)

Where, $U_r$ is the final output vector that is used to categorize the normal and disease affected images with increased detection accuracy and reduced false positive rate.

Results and discussion

This section presents the performance analysis of the proposed perceptual U-Net segmentation based DDO-SDN classification techniques. The original contribution of this work is to develop an automated lung nodule segmentation and pulmonary disease classification system by using advanced image processing mechanisms. Here, the results of the proposed DDO-SDN technique is validated in terms of accuracy, sensitivity, specificity, precision, recall, f-measure, and time. These parameters are increasingly used in all types of medical image processing applications for assessing the performance and effectiveness of the mechanisms.

Dataset description

Here, the EL-CAP lung image dataset and Lung Cancer data obtained (https://data.world/cancerdatahp/lung-cancer-data) are used for system implementation and validation, which are the publicly available benchmark datasets. The ELCAP public lung imaging database is taken into consideration as an input dataset for nodule detection. There are many CT lung images (a collection of 50 low-dose images) for the purpose of diagnosing diseases. With i5 processors and 4GB RAM, the proposed lung nodule identification work is implemented in MATLAB 2016a. The outcome analysis compares several optimization strategies and discusses the isolated features for each lung image. Fig. 4 (a) to (d) depicts the sample input lung images obtained from the EL-CAP dataset, ROI extracted images, contrast enhanced portions, and segmented output images.

View Full Size

a) Input lung images; b) Region of interest (ROI) extracted lung images; c) Contrast enhanced lung images; d) Segmented region

Citation: Imaging 15, 1; 10.1556/1647.2023.00106

• Download Figure
• Download figure as PowerPoint slide

Table 1 and Fig. 4 validates the precision of conventional [32] and proposed classification methodologies with respect to varying number of training samples. Similarly, Tables 2–4 shows the recall, f-measure, and accuracy of conventional and proposed techniques with respect to different number of training samples, and its corresponding graphical illustrations are presented in Figs 5–8 respectively. Typically, the precision, recall, and f-measure are the most essential parameters used for determine the accurate detection performance of classifier. Then, the values of these measures should be improved for assuring the better system functioning and efficiency, which are calculated by using the following models:

$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\frac{TP}{TP+FP}\times 100\%$(18)

$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\frac{TP}{TP+FN}\times 100\%$(19)

$F-\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}=\frac{2\times \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{S}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{S}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}}\times 100\%$(20)

Where, TP – True Positives, TN – True Negatives, FP – False Positives, and FN – False Negatives. According to the estimated results, it is analyzed that the proposed DDO-SDN technique outperforms the other approaches with increased precision, recall, and f-measure values. Due to the proper segmentation and optimization operations, the exact lung region is segmented and, its associated features are selected from the CT image. It helps to train the classifier for exactly identifying the lung disease, hence the overall performance of the proposed DDO-SDN is highly improved than the other approaches.

Table 1.

Precision analysis

No of training samples	Precision
	KSCC with linear kernel	SVM-NN	Proposed
100	95.48	92.4	96.5
200	95.86	93.54	96.98
300	97.15	94.91	98.68
400	98.15	95.48	99.12
500	98.62	95.98	99.47

Table 2.

Recall analysis

No of training samples	Recall
	KSCC with linear kernel	SVM-NN	Proposed
100	92.84	91.51	93.51
200	92.99	91.85	94.43
300	94.19	93.45	95.68
400	95.97	94.51	96.98
500	97.8	96.89	98.45

Table 3.

F-Measure analysis

No of training samples	F-Measure
	KSCC with linear kernel	SVM-NN	Proposed
100	95.6	94.2	96.2
200	96.1	94.5	96.7
300	97	95.7	97.8
400	98.6	97.1	98.7
500	99.6	98.3	99.2

Table 4.

Accuracy analysis

No of training samples	Accuracy
	KSCC with linear kernel	SVM-NN	Proposed
100	95.45	94.15	95.6
200	96.98	94.98	96.8
300	97.46	95.68	97.9
400	97.89	96.87	98.5
500	98.85	98.54	99.4

View Full Size

Precision vs no of training samples

Citation: Imaging 15, 1; 10.1556/1647.2023.00106

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Recall vs no of training samples

Citation: Imaging 15, 1; 10.1556/1647.2023.00106

• Download Figure
• Download figure as PowerPoint slide

View Full Size

F-measure vs no of training samples

Citation: Imaging 15, 1; 10.1556/1647.2023.00106

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Accuracy vs no of training samples

Citation: Imaging 15, 1; 10.1556/1647.2023.00106

• Download Figure
• Download figure as PowerPoint slide

Table 5 and Fig. 9 validates the accuracy and time of conventional [33] and proposed classification methodologies. Typically, the time consumption of disease detection system is estimated based on the amount of time required to train and test the samples by the classifier. Moreover, the reduced time consumption ensures the improved performance and efficacy of the classifier. Based on the estimated results, it is analyzed that the proposed technique outperforms the other classifiers with increased accuracy and reduced time consumption in terms of seconds. Because, the DDO technique could efficiently reduce the dimensionality of features by selecting the optimal parameters based on the best optimal solution. Hence, the training and testing operations of the classifier are improved with reduced processing time consumption.

Table 5.

Accuracy and time analysis

Techniques	Accuracy (%)	Processing Time (s)
Decision tree	95.40	0.0159
KNN – Euclidean distance (k = 6)	99.8	0.0313
KNN – Euclidean distance (k = 10)	96.4	0.0312
GA-KNN	100	0.0156
Proposed	100	0.0098

View Full Size

Comparative analysis based on accuracy and time

Citation: Imaging 15, 1; 10.1556/1647.2023.00106

• Download Figure
• Download figure as PowerPoint slide

Table 6 and Fig. 10 shows the overall performance analysis of the proposed DDO-SDN mechanism, where the results indicate that the proposed technique provides an improved performance values by accurately detecting the disease.

Table 6.

Performance analysis of DDO-SDN

Parameters	Performance value
Sensitivity	98.7
Specificity	98.9
Accuracy	99.25
Precision	99.2
Recall	98.9
F-Measure	98.7
FPR	0.006
TPR	99.1

View Full Size

Overall performance analysis

Citation: Imaging 15, 1; 10.1556/1647.2023.00106

• Download Figure
• Download figure as PowerPoint slide

Table 7 presents the comparative analysis of existing tensorflow and proposed tanh normalization models based on the parameters of error rate (%) and time (s). According to the results, it is observed that the tanh normalization mechanism outperforms the existing tensorflow model with reduced time 15s and error rate 6.54%. Due to its efficient filtering and noise removal operations, the performance of the tanh normalization model is highly improved than the tensorflow model. Table 8 and Fig. 11 assess the performance of the perceptual U-Net lung nodule segmentation mechanism used in the proposed work. Overall, the obtained results indicate that the proposed U-Net segmentation algorithm provides an improved results by accurately segmenting the given lung image.

Table 7.

Comparative analysis between tensorflow and tanh normalization models

Techniques	Error (%)	Time (s)
Tensorflow normalization	8.25%	21
Tanh normalization	6.54%	15

Table 8.

Performance analysis of segmentation

Performance measures	Values (%)
Sensitivity	99
Specificity	98.8
Accuracy	99
Dice	98.2
Jaccard index	98

View Full Size

Performance analysis of the perceptual U-Net segmentation algorithm

Citation: Imaging 15, 1; 10.1556/1647.2023.00106

• Download Figure
• Download figure as PowerPoint slide

Discussion

The paper primarily addressed lung image segmentation and nodule classification in the context of medical image processing. Finding the best subset of features from the large number of input images in the ELCAP public dataset is the primary problem of image classification. We extracted the lung and nodule portions of the dataset and identified the nodule impacted and normal lung images. For this purpose, an advanced image processing techniques such as tanh normalization, perceptual U-Net segmentation, DDO based feature selection, and SDN classification. Here, the lung nodules and non-nodules are exactly segmented and classified with increased accuracy and reduced processing time.

Conclusion

This paper presents a new Perceptual U-Net segmentation based DDO-SDN classification system for detecting the pulmonary lung cancer from CT images. The original contribution of this work is to implement a simple and effective detection framework for an accurate diagnosis of pulmonary lung cancer. In this framework, an advanced preprocessing, segmentation, optimization, and classification methodologies are used to design an automated lung disease detection system. Here, the tanh normalization mechanism is applied at first for preprocessing the input CT images by eliminating the noise/artifacts, and improving the quality of image. Typically, the noisy image can disrupt the overall detection performance of the classifier, hence it should be properly filtered for further operations. After that, the perceptual U-Net segmentation algorithm is used to accurately segment the lung nodules from the preprocessed image, which helps to avoid over-segmentation by accurately cropping the regions. After that, the most relevant features are extracted from the segmented portion based on the optimal solution provided by the DDO algorithm. It mainly helps to obtain an increased detection accuracy and reduce the error rate by suppressing the dimensionality of features. Finally, the SDN based classification mechanism is used to exactly detect the pulmonary lung cancer from the given CT image according to its features. During performance analysis, the results of the proposed segmentation based DDO-SDN technique are validated and compared using various measures. According to the comparative results, it is evident that the proposed DDO-SDN technique overwhelms other approaches with increased accuracy and reduced false positives.

In future, this work can be enhanced by implementing a new optimization and deep learning based classification methods for lung disease diagnosis.

Authors Contribution

The study's inception and design involved input from all authors. T.Chitra prepared the material, collected the data, and carried out the analysis. All writers provided feedback on earlier draughts of the paper after T.Chitra wrote the original draught. The paper's author, P.Jaganathan, contributed to the background research and assisted with the mathematical derivations. The author C.Sundar Technically participated, contributed a factual evaluation, and assisted with manuscript editing. The final manuscript was read and approved by all writers.

Conflict of interest

The authors declare that they have no conflict of interest.

Funding sources

The authors have no relevant financial or non-financial interests to disclose.

Ethical statement

The study submitted to IMAGING have been conducted in accordance with the Declaration of Helsinki and according to requirements of all applicable local and international standards.

Acknowledgements

I am grateful to all of those with whom I have had the pleasure to work during this and other related Research Work. Each of the members of my Dissertation Committee has provided me extensive personal and professional guidance and taught me a great deal about both scientific research and life in general.