Residual attention network based hybrid convolution network model for lung cancer detection

Abstract

Lung cancer is one of the dangerous diseases that cause shortness of breath and death. Automatic lung cancer disease identification is a challenging operation for researchers. This paper, presents an effective lung cancer diagnosis system using deep learning with CT images. It also decreases lung cancer’s misclassification. Initially, the input images are gathered from online resources. The collected CT images are given to the detection stage. Here, we perform the detection using a Multi Serial Hybrid convolution based Residual Attention Network (MSHCRAN). Using a deep learning framework lung cancer detection using CT images is effectively detected. The performance of the developed lung cancer detection system is compared to other conventional lung cancer detection models According to the analysis, the implemented deep learning-based detection of lung cancer system had a precision higher than 95.75% compared to CNN with 90.04%, ResNet with 89.62%, LSTM with 92%, and CRAN with 93.4% using dataset-1. The analysis with Dataset-2 shows a precision of 90.43% with CNN, ResNet with 90.12%, LSTM with 92%, and CRAN with 93.7%, with the proposed method precision of 95.8%.

Keywords

Lung cancer detection convolutional neural network residual attention network

1. Introduction

Lung cancer is one of the most dangerous cancer [1]. The mortality rate is highly associated with lung cancer since symptoms appear after the cancer has advanced. Computerized Tomography (CT) is a proposed and effective tool for screening [2]. When compared to radiography, CT screening could reduce lung cancer mortality rates. Cancer is the name given to a disease ,and it is caused by abnormal growth of tissue [3]. Cancer is a serious disease that threatens human health severely. In medical terms, there are more different forms of cancer presented. With rate, lung cancer is the most prevalent type of cancer worldwide [4]. Lung cancer is expected to have the highest incidence rate when the statistical data on the various cancer kinds in the world are studied [5]. The beating lung cancer is substantially higher with prompt detection and cancer disease cure. When compared to single-view radiography, early diagnosis of lung cancer via a CT image of the lung decreased the mortality rate [6].

Thus, it is crucial to get early and precise lung cancer detection. Experts medical area must examine the diagonised lung CT images, which is a delicate operation that takes time and great skill [7]. The subjective examination causes the observer’s perspectives to differ. These factors have increased the demand for computer-based systems. Using currently available technical tools and software can support the diagnosing procedure. As a result, the expense and diagnostic effort can be greatly diminished [8]. Deep learning is one of the popular models used recently, which is used to diagnose lung cancer. The radiologists to perform effective screening and create accurate datasets of large data of CT images is enormous. Automatic algorithmic solutions could lessen this strain, but interacting with clinicians through algorithms can be difficult because these systems are unable to consistently communicate their uncertainty. The method analyses scans of CT images directly and generates the scores that precisely describe uncertainty. Therefore, CT images include more detailed information and their use in lung cancer diagnosis has lately expanded.

The objectives of the deep learning-based lung cancer diagnosis system are described here.

•
To develop a deep learning-related lung cancer detection system that effectively identifies lung cancer from patients at an early stage.
•
To design an MSHCRAN-based lung cancer detection model to increase the performance of the lung cancer diagnosis system following the performance measure accuracy.
•
To evaluate the effectiveness of the investigated deep learning-related lung cancer diagnosis system among the previously developed lung cancer detection works in regards to various performance measures.

The proposed deep learning-based lung cancer detection system is explained in the remaining sections. The existing lung cancer detection model algorithms and methodologies with advantages and disadvantages are presented in Section 2. Section 3 presents the newly developed lung cancer detection model explanation. The lung cancer detection dataset description explanations are explained in Section 3. In Section 4, the lung cancer detection procedure is described. Section 5 provides a developed detection model for a lung cancer detection system.
2. Literature survey

2.1 Related works

A new automatic technique for lung cancer detection is proposed by Ozdemir et al. [13] that generated accurate probability estimated using CT images. For the openly available dataset like Kaggle and LUNA16, the developed system showed very accurate classification outcomes using neural networks. Also, it was compared to other models and identification techniques. Even though nodule detection systems were normally implemented and refined independently, they discovered that it was crucial to take the coupling between the components for detection and diagnosis. Using that linkage made it possible to create a conventional system that gives fewer false positive rates for predicting lung cancer. In the CT image analysis, they characterized the system using deep learning systems and showed better classification diagnosis results.

Togacar et al. [14] have investigated a new hybrid detection system for predicting lung cancer. It used some deep learning techniques such as LeNet, VGG-16, and AlexNet. An open dataset made up of CT images was used for the research. Here, deep learning techniques were utilized for classification and feature extraction. During the training of the models, the image augmentation techniques such as filling, zooming, cutting, and horizontal rotating were used on the model to improve the classification’s success rate. For the high classification accuracy outcomes, the NN classifier and AlexNet model combination were used. The feature selection approach was used to choose the effective details.

Shake et al. [15] implemented an Enhanced Profuse Clustering (IPCT)-based lung cancer segmentation model. It enhanced the quality of CT images and diagonised lung cancer. It reduced the misclassification rates. The input lung CT scans were collected from the dataset, and the histogram equalization technique was used to reduce the unwanted details from input images. The trained neural network was used in the lung cancer prediction stage to effectively predict lung cancer. So, the model was developed utilizing a MATLAB environment. The developed system maximized the accuracy and minimized the classification error rates.

Ruan et al. [16] developed a new automatic lung cancer screening system to effectively detect lung cancer. The study examined to differentiate between the real lung boundaries and the scout landmarks. They developed a deep learning-based developed model using the CT scan data set. A validation and training set was created from the data set in a ratio. The scan ranges and original lung boundaries were adopted in the test set to increase the effectiveness of the developed system.

Sori et al. [17] have investigated a new novel based on lung cancer detection using a deep learning framework. The newly introduced model had a complete diagnosing and detection component. The first step was residual learning DR-Net-based preprocessing. It was used to reduce the noise from input images. At last, the suggested model solves the challenges like imbalance and high cost of conventional models. They discovered that kind of model may effectively minimize image noise, balance field size, provide more representative characteristics, and be quickly modified to account for variations in nodule shape and size. The experimental outcomes proved high accuracy than other conventional models.

2.2 Problem statement

Table 1
Advantages and disadvantages of lung cancer detection using deep learning with ct images

Author [citation]	Methodology	Features	Challenges
Ozdemir et al. [13]	3DCNN	• It eliminates the nodule false positive reduction rates. • The high robustness performance helps to increase the results.	• It takes more time for the classification process. • It does not detect lung cancer in its early stage.
Togacar et al. [14]	CNN	• It achieved generalized results. • It increases the prediction rate.	• It needs more data to predict cancer effectively. • It does not give accurate solutions.
Shakeel et al. [15]	IPCT	• It easily reduces balance field size, noise in an image, and other features. • It automatically detects important features.	• It minimizes the prediction error. • It provides poor generalization capability.
Ruan et al. [16]	CNN	• It provides flexibility. • It is easily adaptable to nodule size and shape.	• It provides high computational complexity. • It increases healthcare costs.
Sori et al. [17]	DR-Net	• It provides quick prediction after training. • It highly supports complex non-linear problems.	• It provides a low accuracy ratio because of smaller data. • It requires a long training process.

Most lung cancer detection may experience side effects, including long-term neuropathy, radiation to the lungs, depression, and anxiety. Also, it independently affects the prognosis. Some of the deep learning approaches are implemented for the detection of lung cancer. Some are listed in Table 1. 3DCNN [13] eliminates false positive reduction rates effectively. Also, the high robustness performance helps to increase the results. But, it takes more time for the classification process and does not detect lung cancer in its early stage. CNN [14] achieved generalized results. Also, increases the prediction rate. Yet, it needs more data to predict cancer effectively and does not give accurate solutions. IPCT [15] easily reduces balance field size and noise in an image. Also, it automatically detects important features. But, it minimizes the prediction error and also provides poor generalization capability. CNN [16] provides flexibility. Also, it is very effectively adaptable to nodule size and shape. Yet, it provides high computational complexity and also increases healthcare costs. DR-Net [17] provides quick prediction after training. Also, It highly supports complex non-linear problems. But, it provides a low accuracy ratio because of smaller data and it requires a long training process. Hence, these disadvantages help to develop efficient detection of lung cancer with advanced deep learning-based techniques.

3. Proposed adaptive deep learning-based lung cancer detection model with CT Images

3.1 Proposed lung cancer detection model

Lung cancer detection is hampered by some obstacles for both patients and clinicians. Patients face obstacles such as not knowing about detecting programmers, perception barriers like fear of cancer diagnosis and perceived stigma, financial worries, and difficulties accessing imaging sites. The problems of the lung cancer detection approach include selecting high-risk people, standardizing nodule categorization and measurement, training radiologists specifically, optimizing screening intervals and length, and handling ancillary findings. To identify and recruit high-risk and difficult-to-reach persons for screening and to remove identified barriers to participation, the best and most affordable approaches must be established. More study is required to improve risk prediction models for the target group and add life expectancy rather than just mortality. As well as using life expectancy rather than risk stratification at a particular time point, future research must determine how to optimize risk prediction models for the target population. It will also be necessary to handle the issues of manpower requirements, training, and quality assurance. Regarding the number of radiologists needed to complete a screening program, serious concerns emerged. Before clinical implementation, more studies on performance will be required to meet the existing requirements. The structural representation of the deep learning-related lung cancer detection system is shown in Fig. 1.

Figure 1.

Structural representation of deep learning-based lung cancer detection system.

The newly developed deep learning-related lung cancer detection system is used to identify lung disease in its early stage and reduces lung cancer mortality rates because the severe stage of lung cancer leads to affect the patient’s life. The collected images of CT scans are gathered from online resources. The inputs of CT scans are given to the detection phase. Here, lung cancer detection can be done using the MSHCRAN technique. The CT images of the lungs can be used to locate abnormal spots using the lung cancer detection method. It effectively detects lung cancer using this detection technique. It reduces the possibility of patients’ mortality rate due to lung cancer. The performance of the newly suggested MSHCRAN detection of lung cancer system is compared to that of previous lung cancer detection methods, which demonstrated effective performance in terms of high accuracy.

3.2 Dataset description

Two different datasets have been considered to prove the generality.

Dataset-1 (Chest CT-Scan images Dataset): The gathered images of CT scans are collected from datasets using the below link https://www.kaggle.com/datasets/mohamedhanyyy/chest-ctscan-images “Access date: 2023-03-02”. The dataset contains 3 types of CT images: Large cell carcinoma, squamous-cell carcinoma, and Adenocarcinoma. It contains a test set, validation set, and training set. The total numbers of CT images are 1000. The CTs are presented in the jpg format. The training and test sets percentages are 70% and 20%, respectively. Images are not in dcm format, the images are in jpg or png to fit the model. Data contain 3 chest cancer types which are Adenocarcinoma, Large cell carcinoma, Squamous cell carcinoma, and 1 folder for the normal cell. Adenocarcinoma of the lung: Lung adenocarcinoma is the most common form of lung cancer accounting for 30 percent of all cases overall and about 40 percent of all non-small cell lung cancer occurrences. Adenocarcinomas are found in several common cancers, including breast, prostate and colorectal. Adenocarcinomas of the lung are found in the outer region of the lung in glands that secrete mucus and help us breathe. Large cell carcinoma lung cancer grows and spreads quickly and can be found anywhere in the lung. This type of lung cancer usually accounts for 10 to 15 percent of all cases of NSCLC. Large-cell undifferentiated carcinoma tends to grow and spread quickly. Squamous cell carcinoma is a type of lung cancer is found centrally in the lung, where the larger bronchi join the trachea to the lung, or in one of the main airway branches. Squamous cell lung cancer is responsible for about 30 percent of all non-small cell lung cancers, and is generally linked to smoking. And the last folder is the normal CT-Scan images.

Figure 2.

Collected sample CT images for the detection of lung cancer.

Dataset-2 (CT images Dataset): The Medical segmentation Decathalon (MSD) has ten different tasks including Brain, Hippocampus, Lung, Heart, Liver, Spleen, Pancreas, Hepatic vessels, Prostrate and colon images. The dataset for Lung images comprised of 8.53 GB data. The lung dataset was comprised of patients with non-small cell lung cancer from Stanford University (Palo Alto, CA, USA) publicly available through TCIA and previously utilized to create a radiogenomic signature [20, 21, 22]. Briefly, 96 preoperative thin-section CT scans were obtained with the following acquisition and reconstruction parameters: section thickness, $<$ 1.5 mm; 120 kVp; automatic tube current modulation range, 100–700 mA; tube rotation speed, 0.5 s; helical pitch, 0.9–1.0; and a sharp reconstruction kernel. The tumour region was denoted by an expert thoracic radiologist on a representative CT cross section. The gathered CT scan inputs are collected from datasets using the below link http://medicaldecathlon.com/ “Access date: 2023-03-02”. This dataset contains three types of classes. The normal and diseased images are presented here. Here, the files are present in jpg format.

Hence, the term $K_{T}^{S}$ is the gathered CT images of lung cancer, where $t=1,2,\ldots,T$ . The sample CT images for the detection of lung cancer are given in Fig. 2.

4. Detection of lung cancer using multi serial hybrid convolution-based residual attention network via CT Images

4.1 Convolutional neural network

CNNs are particularly useful for identifying images to recognize categories, classes, and objects. It is used to be quite efficient for predicting signal data, time series, and audio. The convolution layer serves as the fundamental building element for developing a CNN [18] architecture. A hierarchy of features can be developed by piling many layers on top of one another. The output layer is typically completely connected in the majority of CNN detection models, but the CNN architecture can be changed depending on the objective. A conventional layer is partially connected to the final output layer. They found that the computational time is faster with this CNN design than it with conventional CNN layers are fully connected layers. There are many weights used. The nodules from prediction labels are finally detected by these weights. The term $g_{k}$ is feature map and it is associated with one weight. Typically, the convolution layer feature map has three steps. The sub-weight is indicated by $x_{k,j}$ and the biases are denoted by $c_{k}$ . The input channel is indicated by $y_{j}$ . The convolution $g_{k}$ is measured using Eq. (1).

$\displaystyle g_{k}=\sum_{j}{x_{k,j}}*y_{j}+c_{k}$ (1)

The ReLu layer is validated using Eq. (2).

$\displaystyle g(y)=\max(0,y)$ (2)

The leaky rectifier linear unit is measured using Eq. (3).

$\displaystyle g(y)=\max(0,y)+\beta\min(0,y)$ (3)

There are many weights used. The nodules from one of the detection labels are finally detected by these weights. Across the labels, they apply the softmax, which normalizes into two distributions. The spatial position vector is denoted by $y$ . The softmax is indicated by $t(y)$ and it is given in Eq. (4). The normalized constant is indicated by $a$ .

$\displaystyle t(y)=\frac{f^{y}}{a}$ (4)

The term $t(y)$ is measured using Eq. (5). Set the term $A$ that denotes the detection label.

$\displaystyle t(y)=\sum_{j}f^{yj}$ (5)

The final convolution hidden layer can be connected to a traditional output layer. The output layer is typically completely connected in the majority of CNN detection models, but the CNN architecture can be changed. The final layer is partially coupled to a conventional layer. The adopted layer structure is provided in Fig. 3.

Figure 3.

Adopted CNN structure.

4.2 Residual attention network

The Residual attention network [19] model decreases the training complexity to resolve the overfitting issues and the residual connection and also stops the gradient from vanishing. Combining the two methods mentioned above results in the Residual attention conventional network. They use convolutions in the Residual attention module in place of the original CNN’s ordinary convolution. On the one hand, the receptive field expands and allows for the detection and segmentation of significant targets. The target may be precisely located to the higher resolution compared to downsampling. The mobility of information can be increased and major information loss can be avoided by combining residual connections. An important reference for enhancing the propagation of feature vectors is the fact that Residual attention may be incorporated into other CNN. They also present residual connectivity. In addition to lower model training complexity, it decreases overfitting, and the residual connection also stops the gradient from vanishing. Combining the two methods mentioned above results in the Residual attention Convolutional network. They use dilated convolutions in the Residual attention module in place of the original CNN’s ordinary convolution. Then, the receptive field expands and allows for the detection and segmentation of significant targets. Then, the target may be precisely located thanks to the higher resolution compared to downsampling. Figure 4 shows the layer description for the RAN network adopted.

Figure 4.

Adopted RAN network.

4.3 Proposed MSHCRAN-based lung cancer detection

The developed MSHCRAN-based deep network is designed to predict lung cancer from images of CT scans with greatly increased prediction accuracy. The complex sequence from the CT images is effectively learned with a developed hybrid deep learning network. Here, two deep learning networks CNN and RAN, are integrated to formulate the hybrid network. The CNN and RAN networks are connected serially and hence, it is named a Multiseriel hybrid network. The CNN output is directly given to the RAN network. So, the performance of the developed model is highly improved. CNN is adopted to detect lung cancer because it is simple to comprehend and quick to train. The algorithm predicts images with the highest accuracy, but it requires a large amount of training data to work well, and it cannot encode the position and orientation of objects. The main disadvantage of the CNN model includes overfitting, bursting gradients, and class imbalance. Several problems could make the model perform worse. The RAN model is used to foster more innovation and competitiveness. It is utilized to enhance the lung cancer detection model’s network performance. That makes things more flexible. It enhanced the integration of the established system. But, decreasing error rates is a challenging task. Moreover, learning could be quite ineffective if the network is too shallow. To overcome these issues, the suggested MSHCRAN-related detection of the lung cancer system is implemented to provide accurate classification outcomes. The implemented MSHCRAN-related detection of lung cancer system increased the effectiveness in terms of accuracy. It effectively maximizes accuracy. The diagrammatic representation of the newly developed MSHCRAN-related detection of the lung cancer system is shown in Fig. 5.

Figure 5.

Diagrammatic representation of newly developed MSHCRAN-based lung cancer detection model.

5. Results and discussions

5.1 Experimental setup

Python was utilized to implement the MSHCRAN-based lung cancer detection system. The developed lung cancer detection system performance was compared among different conventional lung cancer classifiers with several evaluation metrics. A population rate of 10 and an iteration gap of 10 were used for the experimental analysis of the suggested system. The validation employed the classifiers like CNN [14], ResNet [18], LSTM [19], and CRAN [20].

5.2 Evaluation measures

The performance metrics used for the developed detection of lung cancer systems are explained as follows.

(a) FNR: False Negative rate also called as miss rate is the probability that the true positive is mistaken by the test entry. It is calculated using Eq. (6).

$\displaystyle F3=\frac{\textit{CU}_{t}}{\textit{TV}_{t}+\textit{TV}_{n}}$ (6)

(b) NPV: Negative predictive value describes the performance of diagnostic test taken. Precision plays an equal role with NPV in performance measure. It is evaluated using Eq. (7).

$\displaystyle N3=\frac{\textit{TV}_{t}}{\textit{TV}_{t}+\textit{CU}_{t}}$ (7)

(c) Sensitivity: The sensitivity evaluates a model’s ability to predict true positives of each vailable category and is given in Eq. (8).

$\displaystyle\textit{SEN}=\frac{\textit{TV}_{n}}{\textit{TV}_{n}+\textit{CU}_{% n}}$ (8)

(d) FPR: The False positive rate is a probability of falsely rejecting the null hypothesis and is calculated using Eq. (9).

$\displaystyle\textit{FPR}=\frac{\textit{TV}_{t}}{\textit{TV}_{t}+\textit{CU}_{% n}}$ (9)

(e) F1-score: It is a measure for model’s accuracy that combines precision and recall scores. It is measured using Eq. (10).

$\displaystyle F1=\frac{2\times\textit{TV}_{n}}{2\textit{TV}_{n}+\textit{CU}_{n% }+\textit{CU}_{t}}$ (10)

(f) MCC: The Mathew’s correlation coofficient is the best single value summarization of confusion matrix and is given in Eq. (11).

$\displaystyle\textit{MCC}=\frac{\textit{CU}_{t}\times\textit{CU}_{n}-\textit{% TV}_{n}\times\textit{TV}_{t}}{\sqrt{(\textit{CU}_{t}+\textit{TV}_{n})(\textit{% CU}_{t}+\textit{TV}_{t})(\textit{CU}_{n}+\textit{TV}_{n})}(\textit{CU}_{n}+% \textit{TV}_{t})}$ (11)

(g) Precision: The precision parameter quatifies the correct positive predictions and is measured using Eq. (12).

$\displaystyle\textit{Pre}=\frac{\textit{CU}_{n}}{\textit{CU}_{t}+\textit{TV}_{% n}}$ (12)

(h) FDR: The False Discovery rate is the expected ratio of the number of false positive classifications to the total number of positive classifications. The value is measured using Eq. (13).

$\displaystyle\textit{FDR}=\frac{\textit{TV}_{t}}{\textit{TV}_{t}+\textit{CU}_{% t}}$ (13)

(i) Specificity: It is the ratio of true negatives to all negative outcomes and is calculated by Eq. (14).

$\displaystyle\textit{Sp}=\frac{\textit{CU}_{n}}{\textit{CU}_{n}+\textit{TV}_{n}}$ (14)

5.3 Performance analysis on dataset 1

Figure 6.

Performance evaluation on implemented deep learning-related lung cancer detection over different detection techniques using dataset-1.

The newly developed MSHCRAN-related lung cancer detection model’s performance is compared to that of previous detection methods using dataset-1, as shown in Fig. 6. The recommended MSHCRAN-based lung cancer detection model outperformed and improved with high precision of 3.15% than CNN, 4.25% than ResNet, 6.52% than LSTM, and 8.88% than CRAN in the tanh activation function analysis. As a result, the suggested MSHCRAN-based detection of lung cancer systems has shown high prediction rates when compared to traditional methods.

5.4 Performance analysis on dataset 2

Figure 7.

Performance evaluation on implemented deep learning-based lung cancer detection over different detection techniques using dataset-2.

The effectiveness of the developed MSHCRAN-based lung cancer detection model compared to several other lung cancer detection techniques using dataset-2 with activation function analysis is shown in Fig. 7. The developed MSHCRAN-based lung cancer detection model shows outstanding performance in terms of enhanced precision of 11.2% compared to CNN, 13.7% compared to ResNet, 15.1% compared to LSTM, and 16.7% compared to CRAN using dataset-2 in the tanh activation function analysis. Compared to other methods, the created MSHCRAN-based lung cancer detection system exhibited great efficacy in terms of accuracy.

5.5 The overall analysis of the developed lung cancer detection system

Table 2
Performance analysis of developed Lung cancer detection model with different techniques

Measures	CNN [14]		ResNet [18]		LSTM [19]		CRAN [20]		MSHCRAN
Dataset-1
FPR	9	.842898	10	.2597	8	.01539	6	.540558	4	.200064
Sensitivity	90	.10062	89	.64622	92	.04804	93	.57352	95	.68322
FNR	9	.899383	10	.35378	7	.951964	6	.426485	4	.31678
FDR	9	.957833	10	.38287	8	.101102	6	.608358	4	.254628
Accuracy	90	.12903	89	.69355	92	.01613	93	.51613	95	.74194
Specificity	90	.1571	89	.7403	91	.98461	93	.45944	95	.79994
Precision	90	.04217	89	.61713	91	.8989	93	.39164	95	.74537
NPV	90	.21495	89	.76908	92	.13231	93	.63958	95	.73855
MCC	80	.25742	79	.38637	84	.03193	87	.03209	91	.48354
F1-score	90	.07138	89	.63167	91	.97341	93	.48249	95	.71429
Dataset-2
FPR	9	.800719	10	.09474	8	.00392	6	.370467	4	.279647
NPV	90	.08157	89	.58333	91	.84605	93	.44636	95	.68909
MCC	80	.51343	79	.70787	84	.03016	87	.22426	91	.51483
Accuracy	90	.25806	89	.85484	92	.01613	93	.6129	95	.75806
Sensitivity	90	.31539	89	.80567	92	.03568	93	.59669	95	.79484
FNR	9	.684613	10	.19433	7	.96432	6	.403313	4	.205161
Precision	90	.43062	90	.12148	92	.18251	93	.77593	95	.82537
FDR	9	.569378	9	.878517	7	.817486	6	.224066	4	.174634
F1-score	90	.37297	89	.9633	92	.10904	93	.68622	95	.8101
Specificity	90	.19928	89	.90526	91	.99608	93	.62953	95	.72035

Figure 8.

Sample classification results from dataset 2.

The effectiveness of the newly suggested MSHCRAN-based detection of lung cancer model is compared with conventional detection methods and it is shown in Table 2. The investigated detection of the lung cancer model’s performance analysis compared various lung cancer detection techniques using dataset 2, which is given in Table 2. According to the analysis, the implemented deep learning-based detection of lung cancer system had a precision higher than 95.75% compared to CNN with 90.04%, ResNet with 89.62%, LSTM with 92%, and CRAN with 93.4% using dataset-1. The analysis with Dataset-2 shows a precision of 90.43% with CNN, ResNet with 90.12%, LSTM with 92%, and CRAN with 93.7%, with the proposed method precision of 95.8%. The implemented lung cancer detection model performed more effectively and accurately than other traditional lung cancer detection methods. Figure 8 shows the sample classification results.

6. Conclusion

A new deep learning-based lung cancer detection system was developed to predict lung cancer in the early stage and reduces lung cancer mortality rates. The collected CT images were gathered from online resources. The images of CT scans were given to the detection section. Here, lung cancer detection could be done using the MSHCRAN technique. The CT images of the lungs could be used to detect lung cancer models. That technique effectively predicted the abnormal spots in the CT images. It effectively detected lung cancer. The developed MSHCRAN-based lung cancer detection model showed high performance in increased precision of 2.56% compared to CNN, 4.35% compared to ResNet, 6.91% compared to LSTM, and 6.44% compared to CRAN using dataset-2. The performance of the newly developed MSHCRAN-based lung cancer detection model was compared to existing detection methods, which showed effective performance in terms of high accuracy.

Footnotes

Acknowledgments

The authors would like to express their gratitude to Medical Imaging Research Lab(MIR Lab), SCRA, India for their technical support. Also, the authors would like to thank the referees for their comments and suggestions on the manuscript.

References

Xie

Xia

Zhang

Song

Feng

Fulham

, et al. Knowledge-based Collaborative Deep Learning for Benign-Malignant Lung Nodule Classification on Chest CT. IEEE Transactions on Medical Imaging. 2019; 38(4): 991-1004. doi: 10.1109/TMI.2018.2876510.

Liu

Setio

AAA

Ghesu

Gibson

Grbic

Georgescu

. No Surprises: Training Robust Lung Nodule Detection for Low-Dose CT Scans by Augmenting With Adversarial Attacks. IEEE Transactions on Medical Imaging. 2021; 40(1): 335-345. doi: 10.1109/TMI.2020.3026261.

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. doi: 10.1109/ACCESS.2020.2992645.

Gerard

Patton

Christensen

Bayouth

Reinhardt

. FissureNet: A Deep Learning Approach For Pulmonary Fissure Detection in CT Images. IEEE Transactions on Medical Imaging. 2019 Jan; 38(1): 156-166. doi: 10.1109/TMI.2018..

Kumar

Fulham

Feng

Kim

. Co-Learning Feature Fusion Maps From PET-CT Images of Lung Cancer. IEEE Transactions on Medical Imaging. 2020; 39(1): 204-217. doi: 10.1109/TMI.2019.2923601.

Guo

Kruger

Wang

Kalra

Yan

. Knowledge-Based Analysis for Mortality Prediction From CT Images. IEEE J Biomed Health Inform. 2020 Feb; 24(2): 457-464. doi: 10.1109/JBHI.2019.2946066.

Liu

Zhao

Zhu

. Integrating Lung Parenchyma Segmentation and Nodule Detection With Deep Multi-Task Learning. IEEE J Biomed Health Inform. 2021 Aug; 25(8): 3073-3081. doi: 10.1109/JBHI.2021.3053023.

Aslani

Emberton

Alexander

Jacob

. Deep Learning-Based Long Term Mortality Prediction in the National Lung Screening Trial. IEEE Access. 2022; 10: 34369-34378. doi: 10.1109/ACCESS.2022.3161954.

Wang

Deng

Zhou

Feng

, et al. A Weakly-Supervised Framework for COVID-19 Classification and Lesion Localization From Chest CT. IEEE Transactions on Medical Imaging. 2020; 39(8): 2615-2625. doi: 10.1109/TMI.2020.2995965.

10.

Masood

Yang

Sheng

Qin

Lanfranchi

Kim

Feng

. Cloud-Based Automated Clinical Decision Support System for Detection and Diagnosis of Lung Cancer in Chest CT. IEEE J Transl Eng Health Med. 2019 Dec 4; 8: 4300113. doi: 10.1109/JTEHM.2019.2955458.

11.

Lessmann

van Ginneken

Zreik

de Jong

de Vos

Viergever

Isgum

. Automatic Calcium Scoring in Low-Dose Chest CT Using Deep Neural Networks With Dilated Convolutions. IEEE Trans Med Imaging. 2018 Feb; 37(2): 615-625. doi: 10.1109/TMI.2017.2769839.

12.

Setio

Ciompi

Litjens

Gerke

Jacobs

van Riel

Wille

Naqibullah

Sanchez

van Ginneken

. Pulmonary Nodule Detection in CT Images: False Positive Reduction Using Multi-View Convolutional Networks. IEEE Trans Med Imaging. 2016 May; 35(5): 1160-1169. doi: 10.1109/TMI.2016.2536809.

13.

Ozdemir

Russell

Berlin

. A 3D Probabilistic Deep Learning System for Detection and Diagnosis of Lung Cancer Using Low-Dose CT Scans. IEEE Trans Med Imaging. 2020 May; 39(5): 1419-1429. doi: 10.1109/TMI.2019.2947595.

14.

Togacar

Ergen

Cömert

. 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: 23-39. doi: 10.1016/j.bbe.2019.11.004.

15.

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. doi: 10.1016/j.measurement.2019.05.027.

16.

Ruan

Meng

Zhao

Gong

. Development of Deep Learning-based Automatic Scan Range Setting Model for Lung Cancer Screening Low-dose CT Imaging. Acad Radiol. 2022 Oct; 29(10): 1541-1551. doi: 10.1016/j.acra.2021.12.001.

17.

Sori

Feng

Godana

Liu

Gelmecha

. DFD-Net: lung cancer detection from denoised CT scan image using deep learning. Front Comput Sci. 2021; 15(152701). doi: 10.1007/s11704-020-9050-z.

18.

Zheng

Han

Soomro

. An inception module CNN classifiers fusion method on pulmonary nodule diagnosis by signs. Tsinghua Science and Technology. 2020; 25(3): 368-383. doi: 10.26599/TST.2019.9010010.

19.

Yuan

Cheng

Fan

Zeng

. Pulmonary Nodule Detection Using 3-D Residual U-Net Oriented Context-Guided Attention and Multi-Branch Classification Network. IEEE Access. 2022; 10: 82-98. doi: 10.1109/ACCESS.2021.3137317.

20.

Masroor

Niri

Sharifinasab

. A multiple-input deep residual convolutional neural network for reservoir permeability prediction. Geoenergy Science and Engineering. 2023; 222(211420). doi: 10.1016/j.geoen.2023.211420.

Residual attention network based hybrid convolution network model for lung cancer detection

Abstract

Keywords

1. Introduction

2.1 Related works

2.2 Problem statement

Table 1 Advantages and disadvantages of lung cancer detection using deep learning with ct images

3.1 Proposed lung cancer detection model

4.1 Convolutional neural network

5.1 Experimental setup

5.2 Evaluation measures

Table 2 Performance analysis of developed Lung cancer detection model with different techniques

Footnotes

Acknowledgments

References

Table 1
Advantages and disadvantages of lung cancer detection using deep learning with ct images

Table 2
Performance analysis of developed Lung cancer detection model with different techniques