
==== Front
Sci Rep
Sci Rep
Scientific Reports
2045-2322
Nature Publishing Group UK London

39232180
71551
10.1038/s41598-024-71551-8
Article
Computer-aided diagnosis for lung cancer using waterwheel plant algorithm with deep learning
Alazwari Sana 1
Alsamri Jamal 2
Asiri Mashael M. abusharara@kku.edu.sa

3
Maashi Mashael 4
Asklany Somia A. 5
Mahmud Ahmed 6
1 https://ror.org/014g1a453 grid.412895.3 0000 0004 0419 5255 Department of Information Technology, College of Computers and Information Technology, Taif University, P.O. Box 11099, 21944 Taif, Saudi Arabia
2 https://ror.org/05b0cyh02 grid.449346.8 0000 0004 0501 7602 Department of Biomedical Engineering, College of Engineering, Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia
3 https://ror.org/052kwzs30 grid.412144.6 0000 0004 1790 7100 Department of Computer Science, Applied College at Mahayil, King Khalid University, Abha, Saudi Arabia
4 https://ror.org/02f81g417 grid.56302.32 0000 0004 1773 5396 Department of Software Engineering, College of Computer and Information Sciences, King Saud University, PO Box 103786, 11543 Riyadh, Saudi Arabia
5 https://ror.org/03j9tzj20 grid.449533.c 0000 0004 1757 2152 Department of Computer Science and Information Technology, Faculty of Sciences and Arts, Northern Border University, Turaif, 91431 Arar, Saudi Arabia
6 https://ror.org/03s8c2x09 grid.440865.b 0000 0004 0377 3762 Research Center, Future University in Egypt, New Cairo, 11835 Egypt
4 9 2024
4 9 2024
2024
14 206475 3 2024
28 8 2024
© The Author(s) 2024
2024
https://creativecommons.org/licenses/by-nc-nd/4.0/ Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Lung cancer (LC) is a life-threatening and dangerous disease all over the world. However, earlier diagnoses and treatment can save lives. Earlier diagnoses of malevolent cells in the lungs responsible for oxygenating the human body and expelling carbon dioxide due to significant procedures are critical. Even though a computed tomography (CT) scan is the best imaging approach in the healthcare sector, it is challenging for physicians to identify and interpret the tumour from CT scans. LC diagnosis in CT scan using artificial intelligence (AI) can help radiologists in earlier diagnoses, enhance performance, and decrease false negatives. Deep learning (DL) for detecting lymph node contribution on histopathological slides has become popular due to its great significance in patient diagnoses and treatment. This study introduces a computer-aided diagnosis for LC by utilizing the Waterwheel Plant Algorithm with DL (CADLC-WWPADL) approach. The primary aim of the CADLC-WWPADL approach is to classify and identify the existence of LC on CT scans. The CADLC-WWPADL method uses a lightweight MobileNet model for feature extraction. Besides, the CADLC-WWPADL method employs WWPA for the hyperparameter tuning process. Furthermore, the symmetrical autoencoder (SAE) model is utilized for classification. An investigational evaluation is performed to demonstrate the significant detection outputs of the CADLC-WWPADL technique. An extensive comparative study reported that the CADLC-WWPADL technique effectively performs with other models with a maximum accuracy of 99.05% under the benchmark CT image dataset.

Keywords

Computer-aided diagnosis
Computed tomography
Deep learning
Lung cancer
Medical imaging
Hyperparameter tuning
Subject terms

Computational biology and bioinformatics
Lung cancer
issue-copyright-statement© Springer Nature Limited 2024
==== Body
pmcIntroduction

Recently, LC has been one of the essential reasons for cancer-related deaths worldwide1. According to research, 18% of all cancer-related deaths are common causes of death amongst all cancers. Smoking is one of the main reasons for LC, and it has peaked and risen in many countries2. For a few decades, the LC will be more common. To overcome this issue, early detection and exact diagnosis of LC will help enhance patient outcomes3. The person with LC will get 10 to 20% of living five years by the following experiment. Magnetic resonance imaging (MRI) and CT are the archetypal medical processes for early recognition, which enhance patient endurance4. Generally, the early recognition of a cancer incident through precise diagnosis is by applicable dealing that can improve the probability of a complete cure. Despite the medical utensils, highly authorized experts are needed to clarify medical data to diagnose sickness. This is because the experts have differences due to the high complications of medical images. Computer-aided diagnosis (CAD) is vital in the healthcare industry. In recent years, traditional DL and machine learning (ML) models have been deployed5.

Improving computer-assisted analysis models is very challenging for medical applications, and several studies and finance studies have been conducted on numerous diseases6. Past research has proved that DL-based CAD models can successfully boost the proficiency and accuracy of medical diagnosis7. This is mainly to identify several common cancers like breast and LCs. DL-based CAD models can automatically remove high-level features from original images by utilizing dissimilar model structures compared to conventional CAD models8. Moreover, DL-based CAD networks have few restrictions, such as time consumption, low sensitivity and high FP. A cost-effective, fast, and highly sensitive DL-based CAD network for LC forecast is required immediately. One of the main advantages of employing DL in a CAD network is that it can execute endwise detection by testing significant features in a training method9. It is predictable and can simplify its learning, and malicious nodules can be recognized in novel cases when the network is trained10.

This study introduces a computer-aided diagnosis for LC by utilizing the Waterwheel Plant Algorithm with DL (CADLC-WWPADL) approach. The primary aim of the CADLC-WWPADL approach is to classify and identify the existence of LC on CT scans. The CADLC-WWPADL method uses a lightweight MobileNet model for feature extraction. Besides, the CADLC-WWPADL method employs WWPA for the hyperparameter tuning process. Furthermore, the symmetrical autoencoder (SAE) model is utilized for classification. An investigational evaluation is performed to demonstrate the significant detection outputs of the CADLC-WWPADL technique. The major contribution of the CADLC-WWPADL technique is listed below.The CADLC-WWPADL technique employs the lightweight MobileNet for effectual feature extraction, making it ideal for resource-constrained environments. This technique confirms high performance while efficiently capturing significant features. Its contribution is in balancing computational efficiency with robust feature representation.

The CADLC-WWPADL approach implements the WWPA model to fine-tune hyperparameters, improving the performance of the model through effectual exploration of the hyperparameter space. This methodology optimizes parameter settings to attain the optimum possible outcomes. The contribution is in attain greater performance with lesser computational effort.

The SAE method is employed for classification, efficiently capturing complex features and improving the accuracy of the prediction through its state-of-the-art feature representation. This technique enhances the precision of classifications by employing advanced data encoding models. Its usage confirms more accurate and reliable outcomes in the classification procedure.

The approach integrates MobileNet for effectual feature extraction, WWPA for precise hyperparameter tuning, and SAE for advanced classification. This combination creates a robust pipeline that improves model effectiveness and accuracy. The novelty is in impeccably integrating these models to optimize performance across feature extraction, tuning, and classification phases.

The remaining sections of the article are arranged as follows: Section “Related works” offers the literature review, and section “The proposed model” represents the proposed method. Then, section “Results and discussion” elaborates on the results evaluation, and section “Conclusion” completes the work.

Related works

Shah et al.11 employ the DL model of convolutional neural network (CNN) for identifying a Lung Nodule. In this study, an ensemble method was presented to address the problem of lung nodule recognition. The study integrated the performance of two or more CNNs instead of utilizing only one DL technique; this enables them to execute and guess the result with exactness. In12, a “denoising first” 2-path CNN (DFD-Net) method is proposed. The developed technique is controlled by classification and denoising elements in an end method. Initially, an enduring learning denoising method (DR-Net) was mainly utilized for denoising purposes. The two paths primarily pointed to a joint combination of global and local features. Shakeel et al.13 proposed a novel and enhanced image processing (IP) and an ML model to estimate LC. The collective images are generally used by employing the multi-level brightness-preserving method. From the noise-removed lung CT picture, the precious area is divided using an enhanced DNN, in which parts utilize network layers and several features are removed.

In14, a hybrid metaheuristic and CNN technique is mainly proposed, followed by the result vector of the method. Then, the outcome solution vector was distributed to the Ebola Optimizer Search Algorithm (EOSA) to pick out the optimum integration of weights and preferences to learn the CNN method for handling detection issues. Venkatesh and Bojja15 published a novel that viewed the CSO process with the highest threshold for the segmentation purpose, CNN as the classification algorithm and LBP as the removal of feature process on CT images to detect the LC. IoT advanced technology is also mainly executed by executing a Raspberry PI processor. In16, a successive technique for LC diagnosis is used. Thus, two well-organized classification models, such as the CNN and feature-based method, are employed. Using a novel optimization technique, the enhanced Harris hawk optimizer improves the CNN classification model. Next, the classification algorithms such as Haralick and LBP feature are primarily functional to the customary dataset from the CNN classifiers.

Ji et al.17 designed an effectual one-phase technique for automatic LC recognition in CT images called the ELCT-YOLO model. The developed methodology utilized a new Cascaded Refinement Scheme (CRS) collected from two dissimilar kinds of Receptive Field Enhancement Modules (RFEMs) models. Wankhade and Vigneshwari18 designed an effectual model for primary and precise analysis named cancer cell detection utilizing hybrid NN (CCDC-HNN). The factors are removed from CT scan pictures by employing DNN. In the research, an improved 3D-CNN was applied to enhance the accuracy of the diagnosis. Shen et al.19 presented a novel weakly-supervised lung cancer detection and diagnosis network (WS-LungNet). The model employs Semi-CADe using adversarial learning for segmentation and CNA-CADx using cross-nodule attention mechanisms for detection processes. In20, a Deep Fused Features-Based Cat-Optimized Networks (DFF-CON) technique is introduced. This model implements Deep CNN (DCNN) and cat-optimized CNN for segmentation and detection.

The proposed model

This study designs the CADLC-WWPADL technique for automated LC diagnosis. The primary purpose of the CADLC-WWPADL approach is to detect and classify the existence of LC on CT scans. To accomplish this, the CADLC-WWPADL approach comprises three significant operations: lightweight MobileNet model, WWPA-based hyperparameter tuning, and SAE-based LC classification. Figure 1 portrays the overall process of the CADLC-WWPADL approach.Fig. 1 Entire procedure of CADLC-WWPADL model.

Feature extraction module

The lightweight MobileNet model is employed to derive feature vectors21. MobileNet is an excellent choice for feature extraction due to its lightweight architecture and effectualness, which is optimized for mobile and edge devices. Its usage of depthwise separable convolutions substantially mitigates computational cost and model size while maintaining robust performance. This allows for real-time processing with minimal latency, making it ideal for applications with limited resources. Moreover, MobileNet's pre-trained models are appropriate for transfer learning, giving high-quality feature extraction with less training data. Its balance between accuracy and effectiveness makes it a practical choice for deployment and performance. Figure 2 illustrates the structure of the MobileNetV2 model.Fig. 2 Architecture of MobileNetV2 approach.

The MobileNetV2 has multiple layers: an inverse residual module, depthwise (DWConv) and pointwise convolution (PWConv) modules, which are discussed in this section. The DWConv and PWConv mainly decrease the computation cycles and parameters. The inverse residual module predominantly reduces information loss and avoids the vanishing gradient problems. Meanwhile, they minimize the number of operations and parameters and reduce the training time, which is effective for the small image block. A lightweight MobileNet reclassification model reduces data loss and realizes effective classification. The presented technique comprises five modules: Modules A, B, C, D, and E. A and B modules are DWConv with the addition of 1 × 1 Conv; an inverse residual module is Module C; a squeeze excitation (SE) module is Module D; and the final stage module is E. DWConv is optimum at decreasing the computational outcome. However, it shows lower accuracy. Calculating 1 × 1 PWConv in Modules A and B helps mitigate these problems. The inverse residual module in Module C minimizes the data loss problems led by dimensional conversion. Module D is a lightweight attention model for decreasing computation effort. Module E exploits a 1 × 1 Conv for expansion to mitigate the issues of higher resource consumption from the network output. After the pooling layer, the Relu6 function optimizes the network speed. The Relu6 function preserves the network accuracy.1 Relu6=minmax0,x,6

Due to dimension transformation, the network exploits 1 × 1 Conv for a linear outcome to avoid data loss. Further, the drop layer reduces the computation, accelerates the convergence, and alleviates the over-fitting. Processing this effective last-stage model raises the computation rate while preserving accuracy.

Hyperparameter tuning module

In this work, the WWPA is applied to alter the hyperparameter values of the lightweight MobielNet model22. The WWPA approach is selected for tuning the hyperparameters of the lightweight MobileNet method due to its capability to navigate convolutional hyperparameter spaces and optimize performance effectually. WWPA outperforms in balancing exploration and exploitation, which assists in finding optimal hyperparameter values with fewer evaluations. Its adaptive behaviour confirms robust optimization even in high-dimensional spaces, making it appropriate for fine-tuning MobileNet. Related to other methodologies, the effectualness of WWPA in converging to high-quality solutions with minimal computational overhead presents crucial merits for optimizing lightweight techniques. Figure 3 specifies the overall structure of the WWPA model.Fig. 3 Overall structure of WWPA technique.

The WWPA model is based on the real behaviour of waterwheels, which uses a group of individuals to search for a better solution to the problem in the search range. The population of WWPA has dissimilar values for the problem variable due to the various positions of the waterwheel within the search range. The vector is a graphical representation of different solutions to the problems, with every waterwheel signifying the other vectors. Consider that the matrix is applied to portray the whole WWPA population, including each waterwheel variation. In the initial phase of WWPA, the initial position of the waterwheel in the search range is randomly defined.2 P=P1⋮Pi⋮PN=p1,1…p1,j…p1,m⋮⋱⋮...⋮pi,1…pi,j…pi,m⋮...⋮⋱⋮pN,1…pN,j…pN,M

3 pi,j=lbj+ri,j·ubj-lbj,i=1,2,…,N,j=1,2,…,m

Here, the count of waterwheels is N, and the amount of parameters is m, correspondingly; an arbitrary numeral within 0,1 is ri,j; the low and up borders of the jth problem variables are lbj and ubj; the population matrix of waterwheel positions is P; the ith waterwheels (solution candidate) is Pi, and pi,j (problem parameter). Meanwhile, every waterwheel refers to the different approaches to the problems; its objective function is independently defined. The vector effectively represents the value that makes up the objective function of the problem.4 F=F1⋮Fi⋮FN=FX1⋮FXi⋮FXN

In Eq. (4), the vector having the objective function value is F, and the predicted value for the ith waterwheels is Fi. The assessment of objective function is used as a primary yardstick to select the optimum solution. This implies that the maximum values of the objective function correspond to the best member (i.e., the best solution candidate). On the other hand, the maximum value corresponds to the worst member (viz., worst solution candidate). Due to the random movement of waterwheels, the present optima changes over time in the search space.

Phase 1: exploration

Due to the keen sense of smell, Waterwheel is a powerful predator that allows one to determine pests’ origin. A waterwheel begins attacking the insects that get within the range. It initiated an attack and continued its pursuit after finding the prey. The WWPA’s exploration ability might be enhanced to locate the optimum search space and escape from the local optima by mimicking the waterwheel attacks on the insect, leading to significant variation in the waterwheel's location within the search range. Using the simulation technique of the waterwheel towards the bug, Eq. (5) obtains a new position of the waterwheel. The prior location will be abandoned if the objective function values are enhanced by fluctuating the waterwheels.5 W=r1·Pt+2K

6 Pt+1=Pt+W·2K+r2

where r1 and r2 are random variables within [0, 2] and [0,1]. Furthermore, a vector W displays the circle’s diameter where the waterwheel plant searches for the possible location, and K is an exponential parameter within [0, 1]. Equation (7) adjusts the waterwheel's position if the solution remains unchanged after three iterations.7 Pt+1=GaussianμP,σ+r1Pt+2KW

Phase 2: exploitation

An insect is swallowed by the waterwheel and transmitted into a feeding tube. The model of transferring the insect into the suitable tube resulting in minor modifications in the waterwheel location in the search range raises the capability of exploitation in the local search, and the best solution is converged near the one that has been discovered already. It emulates the natural behaviours of waterwheels by defining a new random position as the best location for consuming insects' waterwheels. The subsequent equation demonstrates that the waterwheel is dislocated to the latest location if the target function value exceeds the initial position.8 W=r3·KPbestt+r3Pt

9 Pt+1=Pt+KW

where the random integer in the range of [0, 2] is r3, the existing solution at thetth iteration is Pt, and the better solution is Pbest.

Like the exploration stage, if the outcome does not enhance for three iterations, the subsequent mutation can be used to escape from the local minima.10 Pt+1=r1+KsinFCθ

In Eq. (10), F and C are random variables within [5,5]. Furthermore, the values of K exponentially decrease by the following expression:11 K=1+2∗t2Tmax+F

The WWPA method derives an FF to attain a highly effectual classifier. It determines a positive numeral to characterize the good of the candidate solutions. The lessening of the classifying rate of error is assumed as the FF.12 fitnessxi=ClassifierErrorRatexi=No.ofmisclassifiedinstancesTotalNo.ofinstances×100

Image classification module

The SAE model is exploited to classify the LC. Initially, the structure of SAE is established23. The SAE method is advantageous for classification tasks as it outperforms in capturing complex, high-dimensional data structures and mitigating dimensionality through unsupervised learning. Its symmetric architecture confirms that the encoded factors are meaningful and efficient, conserving significant data while discarding noise. This can pave the way to an enhanced feature representation, improving classification methodologies' performance. Furthermore, the SAE can uncover complex patterns in the data that might be missed by conventional approaches, making it a robust option for handling various and complex datasets. Figure 4 depicts the architecture of SAE.Fig. 4 SAE architecture.

The NN model has an input image, which is decreased to a low-dimensional form before recreating it. This allows the network to learn relevant features of an image for the subsequent segmentation. The SAE comprises multiple layers of encoding used to convert the input images into lower-size feature vectors, and decoded convert the feature vectors into output images. The SAE is made up of 3 encoded and decoded blocks. The mathematical modelling of the SAE structure is given below.

In this setup, each encoder block is assigned to maximize the number of feature mappings while reducing the spatial dimension of the input dataset. Each encoder block exploits a max-pooling layer with the ReLU function, two convolutional layers with different filtering dimensions, and a batch-normalized layer to decrease the input's data dimensionality while preserving significant features.13 hk=fkWk∗hk-1+bk,k=1,2,3

In Eq. (13), the activation function of the encoded block indexed k is fk. hk-1 refers to the prior encoded block or the input x if k=1. The weight and bias of the encoder block indexed k are Wk and bk.

At the same time, each decoder block performs the reverse process of the encoded block. This can be accomplished by using all the decoded blocks with an upsampling layer to extend the spatial dimension of the feature map. Then, the two convolutions with filter counts similar to those in the respective encoded block are used. Subsequently, the final decoded block is 11 convolutions with a softmax function that generates the segmentation mask with the number of class channels.14 gk=fkWk∗gk-1+bk,k=1,2,3

whereas gk-1 denotes the resultant of a prior decoded layer or the outcome of kth encoded layers if k=3, Wk and bk show the weight and bias of indexed k decoded layers, fk represents the activation function of kth decoder layers.

The output layer is shown below:15 y=softmaxWoιιt∗g1+boιιt

In Eq. (15), the weights and biases of output layers are Wout and bout, and y is the forecasted output of SAE.

Results and discussion

In this section, the simulation value of the CADLC-WWPADL method can be investigated by implementing the benchmark CT image dataset24, containing 100 instances and three classes, as portrayed below in Table 1.Table 1 Specifications on database.

Class	Instance numbers	
Normal	35	
Benign	32	
Malignant	33	
Total instances	100	

Figure 5 shows the confusion matrices generated by the CADLC-WWPADL methodology with 80:20 and 70:30 of the TR/TS set. The simulated values exhibit efficient recognition with all three classes.Fig. 5 Confusion matrices of (a–c) TR phase of 80:70 and (b–d) TS phase of 20:30.

Table 2 and Fig. 6 report the LC recognition outputs of the CADLC-WWPADL technique with 80:20 of the TR/TS set. The investigational outputs showed that the CADLC-WWPADL model correctly categorized the images. On 80% TR, the CADLC-WWPADL method attains an average accuy of 92.50%, precn of 89.48%, sensy of 88.04%, specy of 94.19%, and Fscore of 88.50%. Besides, with 20% of TR Phase, the CADLC-WWPADL model achieves an average accuy of 96.67%, precn of 96.67%, sensy of 95.24%, specy of 96.97%, and Fscore of 95.68%, respectively.Table 2 LC detection output of CADLC-WWPADL technique with 80:20 of TR/TS set.

Class	Accuy	Precn	Sensy	Specy	Fscore	
TR (80%)	
 Normal	92.50	85.71	96.77	89.80	90.91	
 Benign	91.25	87.50	84.00	94.55	85.71	
 Malignant	93.75	95.24	83.33	98.21	88.89	
 Average	92.50	89.48	88.04	94.19	88.50	
TS (20%)	
 Normal	100.00	100.00	100.00	100.00	100.00	
 Benign	95.00	100.00	85.71	100.00	92.31	
 Malignant	95.00	90.00	100.00	90.91	94.74	
 Average	96.67	96.67	95.24	96.97	95.68	

Fig. 6 Average of CADLC-WWPADL technique with 80:20 of TR/TS set.

The LC recognition outputs of the CADLC-WWPADL method with 70:30 of TR/TS set are described in Table 3 and Fig. 7. The investigational outputs show that the CADLC-WWPADL method appropriately categorized the images. According to 70% TR, the CADLC-WWPADL method gets an average accuy of 99.05%, precn of 98.33%, sensy of 98.55%, specy of 99.35%, and Fscore of 98.40%. Also, on 30% TR, the CADLC-WWPADL model achieves an average accuy of 97.78%, precn of 97.62%, sensy of 96.67%, specy of 98.04%, and Fscore of 97.01%, correspondingly.Table 3 LC detection output of CADLC-WWPADL technique with 70:30 of TR/TS set.

Class	Accuy	Precn	Sensy	Specy	Fscore	
TR (70%)	
 Normal	100.00	100.00	100.00	100.00	100.00	
 Benign	98.57	95.00	100.00	98.04	97.44	
 Malignant	98.57	100.00	95.65	100.00	97.78	
 Average	99.05	98.33	98.55	99.35	98.40	
TS (30%)	
 Normal	100.00	100.00	100.00	100.00	100.00	
 Benign	96.67	92.86	100.00	94.12	96.30	
 Malignant	96.67	100.00	90.00	100.00	94.74	
 Average	97.78	97.62	96.67	98.04	97.01	

Fig. 7 Average of CADLC-WWPADL technique with 70:30 of TR/TS set.

Figure 8 shows the classifier analysis of the CADLC-WWPADL methodology with 80:20 and 70:30. Figure 8a–c exhibits the accuracy outcome of the CADLC-WWPADL methodology with 80:20 and 70:30. This outcome specifies that the CADLC-WWPADL model accomplishes increasing values over enhancing epochs. The enriching validation through training exhibits the CADLC-WWPADL approach that learns successfully on the test dataset. Then, Fig. 8b–d shows the loss analysis of the CADLC-WWPADL approach at 80:20 and 70:30. The investigational analysis exhibited that the CADLC-WWPADL technique gets nearer to TR/TS loss. This shows that the CADLC-WWPADL methodology gains proficiency with the test database.Fig. 8 Accuy curve of (a–c) 80:20 and 70:30 (b–d) Loss curve of 80:20 and 70:30.

Figure 9 demonstrates the performance of the CADLC-WWPADL methodology with 80:20 and 70:30. Figure 9a–c represents the PR values of the CADLC-WWPADL methodology with 80:20 and 70:30. The simulated outcomes stated that the CADLC-WWPADL approach lead to raise values of PR. Also, the CADLC-WWPADL approach can attain superior PR values for the two classes. Then, Fig. 9b–d signifies the ROC analysis of the CADLC-WWPADL approach with 80:20 and 70:30. The figure reported that the CADLC-WWPADL methodology led to enriched ROC values. Similarly, it is observed that the CADLC-WWPADL methodology can achieve maximum ROC values with each of the two classes.Fig. 9 PR curve of (a–c) 80:20 and 70:30 (b–d) ROC curve of 80:20 and 70:30.

Table 425 provides a wide-ranging simulation analysis to highlight the improvement of the CADLC-WWPADL technique. The comparison study shows that the CADLC-WWPADL technique outperforms other methods across key performance metrics. It attains an Accuy of 99.05%, Precn of 98.33%, Sensy of 98.55%, and Specy of 99.35%. This performance surpasses ODNN, KNN, DNN, YOLO-DLN, DBN-LND, and DBOMDFF-LCC, prominent methods in the field. The superior outcomes of the CADLC-WWPADL method illustrate its efficiency in precisely classifying and detecting features related to existing methodologies, depicting its advanced capacities in the given tasks.Table 4 Comparison analysis of CADLC-WWPADL technique with other models25.

Methods	Accuy	Precn	Sensy	Specy	
ODNN	92.27	91.45	88.74	88.70	
KNN	96.69	97.20	86.62	92.27	
DNN	95.62	97.11	93.01	89.58	
YOLO-DLN	94.92	96.65	94.86	95.26	
DBN-LND	95.17	98.09	93.67	90.37	
AGFLCC-DGM	98.89	97.05	98.12	99.05	
DBOMDFF-LCC	99.01	98.01	98.02	99.07	
CADLC-WWPADL	99.05	98.33	98.55	99.35	

In Fig. 10, relative accuy and precn results of the CADLC-WWPADL are reported. The results highlighted that the ODNN model reaches worse results, whereas the DNN, YOLO-DLN, and DBN-LND models obtain slightly boosted outcomes. Meanwhile, the KNN, DBN-LND, AGFLCC-DGM, and DBOMDFF-LCC models offer considerably improved performance. Furthermore, the CADLC-WWPADL technique surpassed the other models with increased accuy and precn values of 99.05% and 98.33%.Fig. 10 Accuy and precn outcome of CADLC-WWPADL technique with other models.

Figure 11 compares with accuy and precn analysis of the CADLC-WWPADL technique. The simulated values exhibited that the ODNN method gets poorer outcomes, while the DNN, YOLO-DLN, and DBN-LND methodologies acquire moderately increased outcomes. Then, the KNN, DBN-LND, AGFLCC-DGM, and DBOMDFF-LCC methods provide significantly enhanced performance. Moreover, the CADLC-WWPADL method exceeded the other techniques with raised sensy and specy values of 98.55% and 99.35%, respectively. Thus, the CADLC-WWPADL method can be utilized for automated and accurate cancer detection.Fig. 11 Recal and Fscore outcome of CADLC-WWPADL technique with other models.

Conclusion

In this study, the CADLC-WWPADL technique for automated LC diagnosis is designed. The primary purpose of the CADLC-WWPADL technique is to distinguish and classify the presence of LC on CT scans. The CADLC-WWPADL technique comprises three primary operations: lightweight MobileNet model, WWPA-based hyperparameter tuning, and SAE-based LC classification. The design of WWPA for the hyperparameter tuning process helps accomplish enhanced detection results. Lastly, the SAE model can be utilized to recognize LC effectively. An investigational evaluation is performed to demonstrate the significant detection outputs of the CADLC-WWPADL method. An extensive comparative study reported that the CADLC-WWPADL technique effectively performs with other models with a maximum accuracy of 99.05% under the benchmark CT image dataset. The limitations of the CADLC-WWPADL technique encompass potential threats in scaling to vast datasets due to the lightweight capability restrictions of the MobileNet model for convolutional feature extraction. Furthermore, while the WWPA method assists hyperparameter tuning, its efficiency may differ with diverse dataset behaviours. Future studies may improve the technique's robustness across various datasets, optimize computational effectualness, and explore alternative feature extraction techniques to enhance diagnostic accuracy for lung cancer recognition.

Acknowledgements

The authors extend their appreciation to the Deanship of Research and Graduate Studies at King Khalid University for funding this work through Large Research Project under grant number RGP2/43/45. Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2024R729), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. Researchers Supporting Project number (RSPD2024R787), King Saud University, Riyadh, Saudi Arabia. The authors extend their appreciation to the Deanship of Scientific Research at Northern Border University, Arar, KSA for funding this research work through the project number “NBU-FFR-2024- 2932-06”. This study is partially funded by the Future University in Egypt (FUE).

Author contributions

Conceptualization: Sana Alazwari Data curation and Formal analysis: Jamal Alsamri, Ahmed Mahmud Investigation and Methodology: Sana Alazwari, Mashael M Asiri Project administration and Resources: Supervision; Mashael M Asiri, Mashael Maashi Validation and Visualization: Jamal Alsamri, Somia A. Asklany Writing—original draft, Sana Alazwari Writing—review and editing, Mashael M Asiri All authors have read and agreed to the published version of the manuscript.

Data availability

The data supporting this study's findings are openly available in the SIMBA Public Database repository at https://www.via.cornell.edu/lungdb.html, reference number24.

Competing interests

The authors declare no competing interests.

Publisher's note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
==== Refs
References

1. Said Y Alsheikhy AA Shawly T Lahza H Medical image segmentation for lung cancer diagnosis based on deep learning architectures Diagnostics 2023 13 3 546 10.3390/diagnostics13030546 36766655
Said, Y., Alsheikhy, A. A., Shawly, T. & Lahza, H. Medical image segmentation for lung cancer diagnosis based on deep learning architectures. Diagnostics 13(3), 546 (2023).36766655 10.3390/diagnostics13030546
2. Shafi I Din S Khan A Díez IDLT Casanova RDJP Pifarre KT Ashraf I An effective method for lung cancer diagnosis from CT scan using deep learning-based support vector network Cancers 2022 14 21 5457 10.3390/cancers14215457 36358875
Shafi, I. et al. An effective method for lung cancer diagnosis from CT scan using deep learning-based support vector network. Cancers 14(21), 5457 (2022).36358875 10.3390/cancers14215457
3. Mamun, M., Mahmud, M.I., Meherin, M. & Abdelgawad, A. LCDCTCNN: Lung cancer diagnosis of CT scan images using CNN based model. In 2023 10th International Conference on Signal Processing and Integrated Networks (SPIN) 205–212 (IEEE, 2023).
4. Feng, J. & Jiang, J. Deep learning-based chest CT image features in diagnosis of lung cancer. In Computational and Mathematical Methods in Medicine (2022).
5. Tyagi S Talbar SN Predicting lung cancer treatment response from CT images using deep learning Int. J. Imaging Syst. Technol. 2023 10.1002/ima.22883
Tyagi, S. & Talbar, S. N. Predicting lung cancer treatment response from CT images using deep learning. Int. J. Imaging Syst. Technol.10.1002/ima.22883 (2023).10.1002/ima.22883
6. Bushara R A deep learning-based lung cancer classification of CT images using augmented convolutional neural networks ELCVIA Electron. Lett. Comput. Vis. Image Anal. 2022 21 1 10.5565/rev/elcvia.1490
Bushara, R. A deep learning-based lung cancer classification of CT images using augmented convolutional neural networks. ELCVIA Electron. Lett. Comput. Vis. Image Anal. 21, 1. 10.5565/rev/elcvia.1490 (2022).10.5565/rev/elcvia.1490
7. Bhattacharjee A Rabea S Bhattacharjee A Elkaeed EB Murugan R Selim HMRM Sahu RK Shazly GA Salem Bekhit MM A multi-class deep learning model for early lung cancer and chronic kidney disease detection using computed tomography images Front. Oncol. 2023 13 1193746 10.3389/fonc.2023.1193746 37333825
Bhattacharjee, A. et al. A multi-class deep learning model for early lung cancer and chronic kidney disease detection using computed tomography images. Front. Oncol. 13, 1193746 (2023).37333825 10.3389/fonc.2023.1193746
8. Ruan J Meng Y Zhao F Gu H He L Gong X Development of deep learning-based automatic scan range setting model for lung cancer screening low-dose CT imaging Acad. Radiol. 2022 29 10 1541 1551 10.1016/j.acra.2021.12.001 35131147
Ruan, J. et al. Development of deep learning-based automatic scan range setting model for lung cancer screening low-dose CT imaging. Acad. Radiol. 29(10), 1541–1551 (2022).35131147 10.1016/j.acra.2021.12.001
9. Kaur, G. & Sandhu, J. K. Deep learning model for lung cancer detection on CT scan image. In 2023 International Conference on Circuit Power and Computing Technologies (ICCPCT) 1240–1245 (IEEE, 2023).
10. Praveena, M., Ravi, A., Srikanth, T., Praveen, B. H., Krishna, B. S. & Mallik, A. S. Lung cancer detection using deep learning approach CNN. In 2022 7th International Conference on Communication and Electronics Systems (ICCES) 1418–1423 (IEEE, 2022).
11. Shah AA Malik HAM Muhammad A Alourani A Butt ZA Deep learning ensemble 2D CNN approach towards the detection of lung cancer Sci. Rep. 2023 13 1 2987 10.1038/s41598-023-29656-z 36807576
Shah, A. A., Malik, H. A. M., Muhammad, A., Alourani, A. & Butt, Z. A. Deep learning ensemble 2D CNN approach towards the detection of lung cancer. Sci. Rep. 13(1), 2987 (2023).36807576 10.1038/s41598-023-29656-z
12. Sori WJ Feng J Godana AW Liu S Gelmecha DJ DFD-Net: Lung cancer detection from denoised CT scan image using deep learning Front. Comput. Sci. 2021 15 1 13 10.1007/s11704-020-9050-z
Sori, W. J., Feng, J., Godana, A. W., Liu, S. & Gelmecha, D. J. DFD-Net: Lung cancer detection from denoised CT scan image using deep learning. Front. Comput. Sci. 15, 1–13 (2021).10.1007/s11704-020-9050-z
13. Shakeel PM Burhanuddin MA Desa MI Automatic lung cancer detection from CT image using improved deep neural network and ensemble classifier Neural Comput. Appl. 2022 10.1007/s00521-020-04842-6
Shakeel, P. M., Burhanuddin, M. A. & Desa, M. I. Automatic lung cancer detection from CT image using improved deep neural network and ensemble classifier. Neural Comput. Appl.10.1007/s00521-020-04842-6 (2022).10.1007/s00521-020-04842-6
14. Mohamed TI Oyelade ON Ezugwu AE Automatic detection and classification of lung cancer CT scans based on deep learning and Ebola optimization search algorithm PLoS One 2023 18 8 e0285796 10.1371/journal.pone.0285796 37590282
Mohamed, T. I., Oyelade, O. N. & Ezugwu, A. E. Automatic detection and classification of lung cancer CT scans based on deep learning and Ebola optimization search algorithm. PLoS One 18(8), e0285796 (2023).37590282 10.1371/journal.pone.0285796
15. Venkatesh C Bojja P A dynamic optimization and deep learning technique for detection of lung cancer in CT images and data access through Internet of Things Wirel. Person. Commun. 2022 125 3 2621 2646 10.1007/s11277-022-09676-0
Venkatesh, C. & Bojja, P. A dynamic optimization and deep learning technique for detection of lung cancer in CT images and data access through Internet of Things. Wirel. Person. Commun. 125(3), 2621–2646 (2022).10.1007/s11277-022-09676-0
16. Guo Z Xu L Si Y Razmjooy N Novel computer-aided lung cancer detection based on convolutional neural network-based and feature-based classifiers using metaheuristics Int. J. Imaging Syst. Technol. 2021 31 4 1954 1969 10.1002/ima.22608
Guo, Z., Xu, L., Si, Y. & Razmjooy, N. Novel computer-aided lung cancer detection based on convolutional neural network-based and feature-based classifiers using metaheuristics. Int. J. Imaging Syst. Technol. 31(4), 1954–1969 (2021).10.1002/ima.22608
17. Ji Z Zhao J Liu J Zeng X Zhang H Zhang X Ganchev I ELCT-YOLO: An efficient one-stage model for automatic lung tumor detection based on CT images Mathematics 2023 11 10 2344 10.3390/math11102344
Ji, Z. et al. ELCT-YOLO: An efficient one-stage model for automatic lung tumor detection based on CT images. Mathematics 11(10), 2344 (2023).10.3390/math11102344
18. Wankhade S Vigneshwari S A novel hybrid deep learning method for early detection of lung cancer using neural networks Healthc. Anal. 2023 3 100195 10.1016/j.health.2023.100195
Wankhade, S. & Vigneshwari, S. A novel hybrid deep learning method for early detection of lung cancer using neural networks. Healthc. Anal. 3, 100195 (2023).10.1016/j.health.2023.100195
19. Shen Z Cao P Yang J Zaiane OR WS-LungNet: A two-stage weakly-supervised lung cancer detection and diagnosis network Comput. Biol. Med. 2023 154 106587 10.1016/j.compbiomed.2023.106587 36709519
Shen, Z., Cao, P., Yang, J. & Zaiane, O. R. WS-LungNet: A two-stage weakly-supervised lung cancer detection and diagnosis network. Comput. Biol. Med. 154, 106587 (2023).36709519 10.1016/j.compbiomed.2023.106587
20. Gopinath A Gowthaman P Venkatachalam M Saroja M Computer aided model for lung cancer classification using cat optimized convolutional neural networks Meas. Sens. 2023 30 100932 10.1016/j.measen.2023.100932
Gopinath, A., Gowthaman, P., Venkatachalam, M. & Saroja, M. Computer aided model for lung cancer classification using cat optimized convolutional neural networks. Meas. Sens. 30, 100932 (2023).10.1016/j.measen.2023.100932
21. Zhou Y Wen S Wang D Meng J Mu J Irampaye R MobileYOLO: Real-time object detection algorithm in autonomous driving scenarios Sensors 2022 22 9 3349 10.3390/s22093349 35591039
Zhou, Y. et al. MobileYOLO: Real-time object detection algorithm in autonomous driving scenarios. Sensors 22(9), 3349 (2022).35591039 10.3390/s22093349
22. Alhussan AA Abdelhamid AA El-kenawy ESM Ibrahim A Eid MM Khafaga DS Ahmed AE A binary waterwheel plant optimization algorithm for feature selection IEEE Access 2023 10.1109/ACCESS.2023.3312022
Alhussan, A. A. et al. A binary waterwheel plant optimization algorithm for feature selection. IEEE Access10.1109/ACCESS.2023.3312022 (2023).10.1109/ACCESS.2023.3312022
23. Abinaya S Kumar KU Alphonse AS Cascading autoencoder with attention residual U-Net for multi-class plant leaf disease segmentation and classification IEEE Access 2023 11 98153 98170 10.1109/ACCESS.2023.3312718
Abinaya, S., Kumar, K. U. & Alphonse, A. S. Cascading autoencoder with attention residual U-Net for multi-class plant leaf disease segmentation and classification. IEEE Access 11, 98153–98170 (2023).10.1109/ACCESS.2023.3312718
24. http://www.via.cornell.edu/lungdb.html
25. Alamgeer M Alruwais N Alshahrani HM Mohamed A Assiri M dung beetle optimization with deep feature fusion model for lung cancer detection and classification Cancers 2023 15 15 3982 10.3390/cancers15153982 37568800
Alamgeer, M., Alruwais, N., Alshahrani, H. M., Mohamed, A. & Assiri, M. dung beetle optimization with deep feature fusion model for lung cancer detection and classification. Cancers 15(15), 3982 (2023).37568800 10.3390/cancers15153982
