The world’s leading cause of cancer death is lung cancer, and screening by low dose computed tomography (CT) has proven to improve mortality (1). Radiological detection of lung tumors is a difficult task, and the increasing shortage of radiologists contributes to delayed diagnosis and worse prognosis (2, 3). The use of artificial intelligence (AI) can enhance the overall outcome of the disease (4).

Genetic algorithms are general search algorithms based mostly upon the concept of evolution seen in nature. Even with today’s high-performance computers, using an exhaustive search to get the optimal solution for even relatively small problems can be very expensive. Diaz et al. (5) utilized genetic algorithm as a method of feature (genes) selection to support vector machine and artificial neural network to classify lung cancer status of a patient. Genetic algorithm (GA) successfully identified genes that classify patient lung cancer status with notable predictive performance.

Daliri et al. (6) proposed a hybrid method combining a GA for feature selection with extreme learning machines (ELM) for the classification of lung cancer data. The dimension of the feature space is reduced by the GA in this scheme and the effective features are selected appropriately. The data are then fed to a fuzzy inference system (FIS), which is trained by the fuzzy ELMs approach. The method shows accuracy of 97.5% in the diagnosis of lung cancer.

Dehmeshki et al. (7) proposed a shape-based genetic algorithm template matching (GATM) method for the detection of nodules with spherical elements. A spherical-oriented convolution-based filtering scheme is used as a pre-processing step for enhancement. The proposed method resulted in a detection rate of approximately 90%, with the number of false positives at approximately 14.6/scan (0.06/slice). We aimed to develop an optimal classifier based on computational intelligence techniques for the precise diagnosis of lung cancer. To robustly cross-validate the proposed classifiers, two independent lung cancer databases were utilized. Although a previous study by Agrawal et al. (8) investigated the same dataset, the methodology adopted in the present work is fundamentally different. In this study, we explore a novel computational intelligence framework with the objective of achieving improved diagnostic performance and more robust classification outcomes.

The computational framework for this study utilized MATLAB (with all relevant toolboxes), Neurosolutions version 5.07, and XLSTAT 2011 for data processing, model development, and statistical analysis. The flow chart depicted in Figure 1 shows the process followed to carry out the research work using genetic algorithms of artificial intelligence techniques.

Image acquisition.Images of chest CT scans were acquired from multiple sources. These images have been segregated into two databases as follows:

Database I: The Database I contains 80 Lung CT scan images of 80 patients. After processing the images, region of interest (ROI) in relation to tumor were identified and located for lung tumors and used from Database I.

Database II: The Database II comprises 76 Lung CT scan images of 76 patients. After processing the images, ROI in relation to tumor were identified and located for lung tumors and used from Database II and the benign images were used from the set of Database I for the development of the system. Annotations of the lung tumor CT scan images in Database II were performed with the assistance of local medical experts, and all tumor samples in Database II were confirmed to be malignant.

Nature of decision boundary of classifier and feature extraction.After acquisition and preprocessing of CT scan images of lung tumors, feature vectors were determined for both Database I and II. In order to form the feature vector for each tumor image, image statistics parameters, image histogram or transform domain features were used. Feature vector is denoted by feature vector (FV).

FV=[TD1, TD2, ………… TD128, Average, SD, Entropy, Contrast, Correlation, Energy, homogeneity]

Where the Transform Domain (TD) may correspond to the Discrete Cosine Transform (DCT), Fast Fourier Transform (FFT), Walsh–Hadamard Transform (WHT), or histogram coefficients, resulting in a FV of dimension 1×135. In the case of FFT-based features, the FV dimension is reduced to 1×71, as only the first 64 FFT coefficients are retained.

For proper understanding of the input feature space and the nature of the decision boundary separating benign samples from the malignant samples, typical scatter plots were scrupulously examined, where one input feature was plotted against another feature. There were several permutations and combinations of such input features, and this might lead to numerous scatter plots for Database I as well as Database II.

Figure 2 depicts a scatter plot representing the relationship between averageand entropy features for database II. Similar plots were obtained for Database I. Meticulous inspection of these scatter plots showed that the decision boundary discriminating between benign and malignant class is not linear. Seemingly, it was highly nonlinear and complex because of partial or complete overlapping of features for both the classes. Such a complex and highly nonlinear decision boundary can’t be estimated by any statistical classifier of linear discriminant analysis. In addition, no mathematical equation can be assigned to two classes namely, malignant and benign was intractable and an ill-posed classification problem. Only the classifier based on the neural networks and support vector machines endowed with innovative computational intelligence techniques have proven remarkable ability to solve such complex pattern recognition problems.

The four different types of knowledge bases developed earlier for Database I as well as database II are not only based on:

(i) Two Dimensions (2D)-DCT coefficients (128) and Image Statistics parameters (7)

The knowledge base created in Excel is of dimension 80×135 (Database I)

The knowledge base created in Excel is of dimension 76×135 (Database II)

(ii) 2D-FFT coefficients (64) and Image Statistics parameters (7)

The knowledge base created in Excel is of dimension 80×71 (Database I)

The knowledge base created in Excel is of dimension 76×71 (Database II)

(iii) 2D-WHT coefficients (128) and Image Statistics parameters (7)

The knowledge base created in Excel is of dimension 80×135 (Database I)

The knowledge base created in Excel is of dimension 76×135 (Database II)

(iv) Image Histogram coefficients (128) and Image Statistics parameters (7)

The knowledge base created in Excel is of dimension 80×135 (Database I)

The knowledge base created in Excel is of dimension 76×135 (Database II) but also carefully selected image statistics parameters.

The above knowledge bases were used to simulate Multi-layer Perceptrons (MLP), Generalised Feed Forward Neural Networks (GFF-NN), Modular Neural Networks (MNN) and Support Vector Machines (SVM) using different learning rules. These simulations were performed on knowledge bases constructed from various transform domains such as DCT, FFT and WHT for both database I and database II. In addition, knowledge bases containing Histogram coefficients were also used for simulation of the neural networks.

Selection of activation functions.In order to determine the optimal activation function for experimentation on the DCT, FFT, WHT and HISTOGRAM knowledge bases, experiments were carried out using MLP neural network to achieve maximum average classification accuracy on the Cross Validation (CV) dataset.

The best results obtained for activation function with respect to knowledge bases are listed as below:

i) DCT Knowledge base: TANH; ii) FFT Knowledge base: LINTANH; iii) WHT Knowledge base: TANH; iv) HIST Knowledge base: LINTANH.In this research work, the entire experimentation was carried out using the above mentioned activation function for the classifier with respect to each knowledge base.

Choice of data-partitions.To decide the appropriate data partitioning percentage for the classifiers, experimentation was carried out on MLP neural network with respect to DCT knowledge base and the obtained results are portrayed in Table I.

The work was carried out for database I and database II following Advanced Genetic Options as portrayed in Table II for genetic optimization of neural networks, with a fixed crossover probability of 0.9 across all evaluated combinations.

The range and values of the other parameters during simulation of the genetically optimized neural network are as follows:

No. of Epochs=10,000; Population Size=50; Maximum Generations=100; Maximum Evolution time=60 min; Cross-validation termination=Terminate after 100 epochs without any improvement; Step size optimization=0 to 1; Momentum optimization=0 to 1; The number of processing elements (PEs) was optimized within a range of 1 to 50. Table III presents only those classifiers that achieved 100% average classification accuracy (ACA) on the cross-validation dataset.

Based on the observations of the above computer simulation experiments and the comparison among different classifiers, it was noticed that the knowledge base formed using 2D-FFT coefficients resulted into the best classifier performance, when Modular Neural Network (Topology II) with Momentum learning rule was employed (Figure 3).

This MNN was further re-configured using advanced genetic options Roulette-Rank-Two Point-uniform specifications and the resultant performance measures observed were:

The Average Classification Accuracy on cross-validation: 100%; The number of connection weights and biases, “N”: 2,002; The Average Minimum Square Error (MSE) on cross-validation: 0.092; The time elapsed per epoch per exemplar, ‘τ’: 0.5729 ms; The fitness of chromosomes: 0.0841. The snapshot for the settings of different training parameters for the above-mentioned best classifier is shown in Figure 4.

For database I, the best Neural Network (NN) based classifier was MNN (Topology II). The knowledge base used to train and cross-validate this classifier was using 2D-FFT coefficients. The size of the knowledge base was 80×71, which implies that there are 71 different input features.

Similarly, the experimentation was carried out on database II and the results of various classifiers with the use of different learning rules and different knowledge bases of database II were investigated and only the results showing 100% classification accuracy are tabulated as in Table IV.

As MNN (Topology II) was identified as an optimal classifier on FFT knowledge base for database I, the same MNN (Topology II) classifier was applied on database II to obtain optimum results. The best results obtained for database II were:

The Average Classification Accuracy on cross-validation: 100%; The number of connection weights and biases, ‘N’: 3,556; The average MSE on cross-validation: 0.0028; The time elapsed per epoch per exemplar, ‘τ’: 0.1268 ms; The area under ROC curve for test on CV Dataset, “AZ”: 1.0; The fitness of chromosomes: 0.004548.

From the meticulous observation of Table V, it is obvious that the use of single hidden layer MNN (Topology II) with momentum learning rule on FFT knowledge base for database I and database II resulted in the reasonable and optimal classifier based on Computational Intelligence (C.I.) techniques for the diagnosis of lung cancer as compared to the single hidden layer MLP based classifier.

Thus, MNN (Topology II) based classifier, which is genetically optimized and trained, is recommended as the optimal classifier for the diagnosis of lung cancer from lung tumor CT scan images (Figure 5).

The observation of the above results for database I and database II indicated that two different sets of optimal classifiers with different configurations are required to be designed for best results. To provide a common optimal classifier for different databases, further experimentation was carried out on FFT knowledge base for database I and database II in two stages.

Stage I: As MNN (Topology II) was identified as an optimal classifier on FFT knowledge base for database I, therefore, the same MNN (Topology II) classifier was applied on database II to get optimum results. The best results obtained for database II were:

The Average Classification Accuracy on cross-validation: 100%; The number of connection weights and biases, ‘N’: 3,556; The average MSE on cross-validation: 0.0028; The time elapsed per epoch per exemplar, ‘τ’: 0.1268 ms; The area under ROC curve for test on CV Dataset, “AZ”: 1.0; The fitness of chromosomes: 0.004548.

Figure 6 shows the plot of area under curve of best Classifier for database II. Area under ROC=1.0; Area under convex hull of an ROC curve=1.0.

Stage II: The MLP was previously identified as an optimal classifier on FFT knowledge base for database II, therefore, the same MLP classifier was applied to database I to obtain optimum results. The best results obtained for database I were:

The Average Classification Accuracy on cross-validation: 100%; The number of connection weights and biases, ‘N’: 3,406; The Average MSE on cross-validation: 0.03993; The time elapsed per epoch per exemplar, ‘τ’: 1.12 ms; The area under ROC curve for Test on CV Dataset, “AZ”: 0.98; The fitness of chromosomes: 0.064686. Figure 7 depicts the plot of area under curve of best Classifier for database I (Test on CV Dataset).

The area under the ROC curve was 0.98, while the area under the convex hull of the ROC curve was 0.99.

In this study, we propose a novel genetically optimized classifier for lung cancer detection using chest CT images. Based on the meticulous observation of the obtained result, the MNN (Topology II) based classifier, which is genetically optimized and trained, is recommended as the optimal classifier for the diagnosis of lung cancer from the lung tumor CT scan images.

AI when integrated with life sciences has ample potential for cancer research and might help in providing better patient outcomes and interventions (9). Diagnosis is a crucial step in providing better results. AI can be a way to help with the increasing shortage of radiologists and AI is doing that without replacing the traditional diagnostic decision steps, and without influencing radiologist’s decisions (10). However, successful AI integration is obstructed and needs a lot of effort from the staff as there are major disagreements regarding autonomous AI usage across surveys, pre-, during and post-implementation (11).

Initially developed CNN by Ciompi et al., showed a diagnostic accuracy that was comparable to that of an experienced radiologist (12). Nasrullah et al., automated lung nodule detection and classification, and showed reduced misdiagnosis and false positives in early-stage (13). Shen et al., produced more interpretable lung cancer predictions and achieved significantly better results than a 3D CNN alone (14). Additional contributions that have advanced knowledge in this field include the work of Cui et al. (15), whose model demonstrated results highly consistent with expert radiologists in lung nodule classification, and Siddiqui et al., who identified the optimal network configuration as (128–128–20) (16). The majority of these models used the LUNA 16 database, and/or LIDC-IDRI (17, 18). The TCIA lung cancer database has also been used by some of the models (19).

Our novel approach uses genetically optimized modular neural network to achieve better results while trying to be computationally efficient and maintaining the quality of outputs. The results we present are not only aligned with the findings of CNN based models but also add to them (20). Our 100% results though remarkable should be interpreted with caution; this study was conducted with a small sample size compared to its predecessors. Further, the future studies on the model should be conducted as a multi-centered, large sample size, prospective study, similar to the work of Fockens, Kiki et al. (21).

While most of these studies utilized the computational strength of the CNN, our model attempted to focus on the GA-MNN. The work of Elayaraja et al., has shown excellent results in case of cervical cancer using genetic algorithm, reaching a sensitivity of 99.09%, specificity of 99.39%, and accuracy of 99.36% (22). Similar results were seen in the case of pancreatic cancer by Li et al., where genetically optimized back propagated AI was utilized for diagnosis and prognosis (23). In case of lung cancer diagnosis, a recent study showed that AI driven GA-optimized lung cancer segmentation has the potential to provide precise lung cancer diagnosis and at the same time, reduce healthcare disparities in resource-limited settings (24).

Our study aimed to further motivate scientists to apply genetically optimized MNN to their future studies on lung cancer and produce more research that can lead to integration of this technology to a multi-modal approach that helps create patient specific management (25). Efforts to develop multi-modal AI for medical application have been done in case of thyroid carcinoma by Yu et al. (26). Another application of multi modal AI was seen in the work of Gao et al., where it was utilized for assessing treatment response (27). Both studies show the promise that AI brings in multi modal management, providing better patient care in the future. While developing future AI models for healthcare, it is necessary to keep in mind the regulations on AI use and that many radiologists are still unconvinced about the return on investment that AI in radiology brings (28, 29). We also need to keep in mind as stated by Tjoa et al., ‘how AI takes decisions is still a black box and we poorly know many machine made decisions’ (30).

Our genetically optimized MNN model offers a computationally efficient, highly accurate, and a novel approach for lung cancer diagnosis. This shows a future potential of integration into clinical workflows, which may help address diagnostic delays, serve as a triage in case of patient overloads and radiologist shortage, optimize resource utilization, and ultimately improve patient outcomes. This model might serve as a pathway for further development of future genetically optimized neural network models in medical imaging.

The Authors have no conflicts of interest to declare in relation to this study.

Dr. Vijay Agrawal: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing-review & editing; Trushdeep Agrawal: Supervision, Visualization, Writing – original draft, Writing – review & editing, Project administration; Dr. Aruni Ghose: Supervision, Writing – review & editing, Project administration, Resources; Dr. Sola Adeleke: Supervision, Writing – review & editing, Project administration, Resources; Dr. Stergios Boussios: Supervision, Writing- review & editing, Project administration, Resources; Dr. Rajender Singh Arora: Conceptualization, Data curation, Validation, Resources, Supervision.

None.

No artificial intelligence (AI) tools, including large language models or machine learning software, were used in the preparation, analysis, or presentation of this manuscript.