Optimized deep learning approach for lung cancer detection using flying fox optimization and bidirectional generative adversarial networks

Research contributions

The main contributions of this research are:

1. To employ effective pre-processing strategies to handle missing values using multiple imputations by chain equations and scale the features using standard and min-max scaler techniques.

2. To incorporate a metaheuristic bio-inspired optimization algorithm such as FFXO to choose the most optimal features that best enhance the classifier model’s reliability.

3. To implement Bi-directional Generative Adversarial Networks to perform lung tumor classification based on the features selected by the FFXO algorithm and demonstrate its performance efficiency by comparing it against the traditional classification algorithms.

This study is motivated by the necessity to solve these problems with an original feature selection and classification deep learning approach. Even though the diagnosis of lung cancer with deep learning models has greatly improved, their efficiency is often limited by high dimensional data and poor feature selection models. At the same time, other models usually have to deal with the datasets’ limitations, making them overfitted. Many researchers have tried to enhance the accuracy of classification with the help of feature selection and optimization methods. Unfortunately, most traditional methods, including PCA and recursive feature elimination (RFE), do not consider non-linear relationships among features. To address these shortcomings, the research utilises a novel hybrid approach that combines Bi-GAN based classification and Flying Fox Optimization (FFXO) feature selection.

However, there are still significant gaps that need to be addressed:

• In adequate feature selection problem in deep learning algorithms—Many deep learning algorithms completely ignore feature selection and accept high-dimensional input data, which leads to overfitting and a lack of interpretability.

• Difficulties with small and highly imbalanced datasets—While deep learning models are proven to perform well with high volumes of data, their biases are evident with the limited medical datasets, making them generalize poorly with high variance.

• Impractical lung cancer classification in real-time—The computational inefficiency of existing hybrid models is that high-performing models, such as LungX-Net and adaptive optimization techniques, require a substantial amount of resources which are not feasible for real-life scenarios.

• Deep learning frameworks’ lack of generalization—Existing models have subpar performance on different datasets because of the limited diversity of the dataset and inefficient feature selection processes.

This study aims to effectively fill the identified gaps by offering an optimised deep learning framework that incorporates:

• Adversarial training that improves feature representation and generalisation on small datasets is applied to lung cancer classification processes through Bi-GAN.

• FFXO enables better feature selection by applying the reduction of dimensionality technique while retaining important attributes, increasing the model’s interpretability and minimising overfitting.

• Class imbalance is countered through synthetic data generation using Bi-GAN, improving generalisation performance and boosting model robustness in these critical areas.

• Combining the elements above will produce a hybrid model with low computational expense that is accurate and deployable in clinical settings.

Paper organization

The remainder of the article is organised as follows. “Related Works” discusses the existing techniques proposed in the literature for classifying lung tumors using AI-based methods. “Proposed Methodology” provides a detailed description of the proposed methodology, which involves data pre-processing, selecting significant features, and categorising lung cancer. “Results and Discussion” presents the results obtained from experiments performed using the lung cancer dataset on the proposed methodology and draws a comparison with other state-of-the-art techniques. “Justification for Model Type Used” concludes the present research.

Motivations of current research

Conventional approaches, such as radiographic procedures, cannot always identify minor irregularities, and human comprehension is sometimes arbitrary and prone to inaccuracies. Deep learning models present an intriguing solution because of their capacity to identify intricate relationships, and automate the diagnostic procedure. Enhanced Flying Fox Optimisation (FFXO) is intended to improve model performance by resolving issues such as overfitting and premature convergence. By producing synthetic data, Bi-GANs can assist in addressing data restrictions and class disparities. By tackling these issues, the present research aims to enhance the reliability of lung cancer detection, improving clinical results and reducing the rate of death.

Data preprocessing

Feature selection using FFXO

The FFXO algorithm is a metaheuristic optimization technique inspired by the natural environment that simulates the feeding habits of big fruit bat species known as flying foxes. These bats have organized exploration strategies; they use group movement and interaction to find food reserves across large distances. This optimization algorithm simulates a collection of flying foxes investigating a solution set to replicate their behavior. It is a potential feature selection technique for lung cancer detection as it can effectively explore high-dimensional pattern regions, enhance classification quality, and manage chaotic data to improve evaluation accuracy. The algorithm begins by generating arbitrary solutions, which are considered to be the location of the flying foxes as represented in Eq. (5),

(5) $$s=(s_1,\dots .,s_k).$$

The flying foxes move towards a cooler location to mitigate the hot temperature. This migration behavior can be formulated as shown in Eq. (6),

(6) $$s_{a,b}^{k+1}=s_{a,b}^k+v.rand(coo{l}_b-s_{a,b}^k).$$

In the above equation, $v$ is an arbitrary value that is fixed in nature, and $coolb$ denotes the location of the flying fox, which is considered the most optimal solution. Upon finding the coolest location, it also explores the next best location to prohibit strangling. The updated locations are represented as given in Eqs. (7) and (8),

(7) $$cur{r}_{a,b}^{k+1}=s_{a,b}^k+ran{d}_{1,b}.\left(coo{l}_b-s_{a,b}^k\right)+ran{d}_{2,b}.(s_{R_1,b}^k-s_{R_2,b}^k)$$

(8) $$s_{a,b}^{k+1}=\{\begin{array}{c}cur{r}_{a,b}^{k+1},ifb=corran{d}_j\ge p\\ s_{a,b}^k,otherwise\end{array}.$$

In the Eq. (8), $ran{d}_j$ represents an arbitrary value which lies between zero and one, $p$ denotes a constant value that corresponds to the likelihood factor, $s_{R_1,b}^k$ and $s_{R_2,b}^k$ are members from the group that are generated arbitrarily. $cur{r}_{a,b}^{k+1}$ represents the solutions that are currently generated is ensured that the current solutions are not duplicates of the previously generated solutions. If $cur{r}_{a,b}^{k+1}$ is found to be the optimal solution, then it is set to be the next coolest location. Otherwise, the flying fox is assumed to be in the previous location. The flying foxes die for two reasons, and in such cases, the locations of new flying foxes need to be determined. Firstly, the flying foxes die when the temperature is too high, so they cannot return to their original locations. The dead flying foxes are replaced by new ones and the location of the replaced flying foxes is determined as given in Eq. (9),

(9) $$s_{a,b}^{k+1}=\frac{{\sum }_{c=1}^{N}r{p}_{c,b}^k}{n}.$$

Secondly, the flying foxes will also die when circumscribed by other flying foxes in the group. Therefore, the probability of flying foxes ( $p_{best}$) to be present in the coolest place is computed using the Eq. (10) as,

(10) $$p_{best}=\frac{num-1}{to{t}_{size}}.$$

In the above equation, $num$ denotes the total number of flying foxes that have the same set of optimal solutions and $to{t}_{size}$ denotes the total population of flying foxes in the search space. The process of copulation between two flying foxes is performed by arbitrarily selecting two members ( $G_1,G_2$) as parents to produce new successors. The resultant successors are determined using the representations in Eqs. (11) and (12),

(11) $$successo{r}_1=H.G_1+\left(1-H\right).G_2$$ (12) $$successo{r}_2=H.G_2+\left(1-H\right).G_1.$$

H in the above equation, it denotes an arbitrarily generated value that lies within the range of zero and one. The algorithmic steps involved in the FFXO algorithm are presented in Algorithm 1.

Several algorithm parameters in the Bi-GAN + FFXO model affect classification accuracy, training stability, and computational resource efficiency. The learning initial rate (0.0001 with Adam optimizer) provides stabilization by allowing smooth convergence. A batch size of 32 is a good compromise between computation and generalization, preventing overfitting while allowing sufficient stable gradient updates. The number of training epochs is set to 100 to provide reasonable learning, with an early stopping allowance to assist in conserving computation if validation performance has not improved for some time. Feature selection is done using FFXO, where 15 features with most relevance to the classification task are selected. This strategy provides enough dimensionality reduction to maintain classification accuracy. The population size in FFXO (50) limits the number of considered feature subsets and it provides a compromise between computational efficiency and diversity of the search space. In addition, mutation (0.3) and crossover (0.7) rates provide tradeoff selection ensuring diversity in the feature sets selected and a balance between exploration and exploitation. In Bi-GAN with 128 latent vector size, model robustness and data synthetic generation is enhanced, while three hidden layers in the generator and the discriminator networks improve data feature extraction. All these parameters enhance the generalisation capability of the model, along with FFXO, which improves feature selection, and Bi-GAN, which increases feature learning and dataset augmentation. Metaheuristic optimization paired with adversarial training achieves maximised classification accuracy (98.7%) for computational power. Future research may consider hyperparameter tuning strategies like Bayesian optimization or grid search to further optimise these values.

Classification using Bi-GAN

In a GAN approach, two neural networks—the discriminator and the generator, compete against one another in a game-like fashion, usually taking the structure of a game with no losers in which the success of one entity equals the setback of another entity. The discriminative network, also known as the discriminator, assesses the newly created data points that the generative network pulls from a minimal dimensional range. The discriminator separates the newly generated datasets produced by the generator from the original data distribution, while the generator trains how to transform from random variability to a data distribution of the actual information.

Bi-GANs are primarily known for their ability to encode and decode data, learning rich latent representations of input features. While Bi-GANs are often applied in image generation, data augmentation, and anomaly detection, their unique ability to map data to a latent space makes them highly effective for classification tasks, particularly in medical diagnosis. The primary motivation for using Bi-GAN in this study is its ability to enhance feature representation learning. Traditional deep learning models, such as CNNs and RNNs, rely on manually extracted or predefined features, which can limit their effectiveness in handling complex medical datasets. In contrast, Bi-GAN automatically learns intricate patterns in the data by mapping it to a structured latent space, which enables the model to differentiate between cancerous and non-cancerous lung tissue more effectively. The encoder network in Bi-GAN captures discriminative features, improving classification performance. Another key reason for using Bi-GAN is its ability to handle class imbalance and improve generalisation. Medical datasets, particularly those related to lung cancer, often suffer from imbalanced class distributions, where there are significantly fewer malignant cases compared to healthy samples. A traditional classifier may become biased toward the dominant class, leading to poor recall and high false negatives, which can be detrimental in a medical context. Bi-GAN enhances the classifier’s generalisation ability by generating synthetic data representations, ensuring better discrimination between positive and negative cases.

An encoder is added to the basic GAN model to create an enhanced version known as bidirectional GAN, or Bi-GAN. To better assist the generator in creating artificial datasets that are more conceptually affluent, the Bi-GAN can now learn a reverse translation from the actual information to the hidden space through the incorporation of an encoder. Here, the encoder helps the Bi-GAN model understand the hidden representation of the exact information, an important function of the encoder. Similar to a regular GAN, there are two phases in the training process. The discriminator must first be trained to maximize the goal parameter without changing the generator, as shown in Eq. (13). The generator and encoder are trained to minimise the same desired function connected to the discriminator in the subsequent step. The structure of Bi-GAN is depicted in Fig. 2.

(13) $$\begin{array}{rl}\underbrace{min}G,E \underset{D}{\underbrace{max}}H\left(D,E,G\right)=& E_{a\sim dA}\left[E_{c\sim dE\left(\cdot |a\right)}\left[logD\left(a,c\right)\right]\right]\\ & +E_{c\sim dC}[E_{a\sim dG\left(\cdot |c\right)}\left[\mathrm{l}\mathrm{o}\mathrm{g}\left(1-D\left(a,c\right)\right)\right]].\end{array}$$

The discriminator determines a single objective function, even though the generator and the discriminator are trained concurrently. Through the discriminator, both the encoder and the generator get passive modifications. In contrast to the conventional GAN method, the discriminator consists of an additive step. The fusion process combines data with its hidden space E(a). The discriminator then receives both of these combined sets of information. During the optimisation, it can be seen that G and E are reversed representations of each other, which are represented as a = G (E(a)) and c = E (G(c)). The algorithm to train the Bi-GAN technique is presented in Algorithm 2.

After identifying and selecting the essential aspects, classification was performed using Bi-GANs. Classifiers like decision trees and support vector machines have difficulty using medical data due to its high dimensionality, class imbalance, and many complex interactions among the features. Such class imbalance problems are often mitigated in the models using Bi-GANs’ ability to create synthetic features. Moreover, Bi-GANs enhance the model’s robustness. Standard GANs typically consist of a Generator (G) that generates synthetic samples and a Discriminator (D) that attempts to distinguish real samples from the fake ones. Bi-GAN incorporates an Encoder (E) that learns to map real data into a latent vector.

This is done to enhance feature extraction and classification. There are two main steps in the training of Bi-GAN. Initially, the D undergoes training to differentiate between genuine patient records and those produced through the G processes. The Generator learns how to make realistic features of lung cancer, which increases the quality of samples it tries to produce, which closely matches actual patient data. Furthermore, the Encoder (E) model translates actual samples into latent space, which helps in modeling the fundamental structure of lung cancer-associated patterns. This iterative training process continues in a two-way direction for the generator and the discriminator, ensuring that the generator improves the quality of the data generated. In contrast, the discriminator improves in knowing what information is accurate and what is counterfeit. At the end of the training, the classification of lung cancer using the Bi-GAN model is highly accurate as there is an integrated selection of feature processes via FFXO and data generation processes through adversarial learning.

Multiple assumptions were made during the experiment to maintain the accuracy and replicability of the methodology. The model was assumed to perform reasonably well on previously unseen data because the dataset was a representative sample of real-world lung cancer patients. In addition, it was believed that the missing values were missing at random, which justified the use of MICE for imputation. The selected features were chosen using a combination of statistical methods (i.e., correlation matrix) and clinical expertise, ensuring that those chosen features were significant risk factors for lung cancer. In addition, it was assumed that the Bi-GAN model could generalise well to new test data, mainly because the model was put through a rigorous 5-fold cross-validation process.

Although the approach uses deep learning in the scope of lung cancer detection, it incorporates several novel contributions that separate it from other methods. The combination of FFXO with Bi-GAN represents an essential feature selection and classification development. Instead of traditional implementations of deep learning where structure changes or hyperparameter tuning are the only options, this research alters feature selection at the level of metaheuristic optimization, resulting in higher accuracy, lower computational costs, and enhanced generalisation.

Perhaps the most important innovation in this work is using FFXO as a metaheuristic feature selector. Unlike conventional techniques like principal component analysis (PCA) or recursive feature elimination (RFE), FFXO has the capability to selectively and adaptively retrieve relevant attributes by making use of both exploration and exploitation. These types of techniques enable the model to effectively remove irrelevant features, enhance learning, and mitigate overfitting. While feature selection is one of the key components of deep learning models, it tends to be neglected. By using FFXO in this study, the efficiency of classification processes is significantly augmented while still retaining high levels of prediction accuracy.

This study seeks to expand the use of Bi-GAN for detection of lung cancer. This area appears to be pioneering and Bi-GAN approach differs from the classic use of CNNs, SVMs, or decision trees in that it performs classification and also generates synthetic data to address class imbalance problems. Bi-GAN’s adversarial framework allows it to capture data patterns and extract features that are more complex, resulting in improved performance of the classification model. The introduction of an encoder in Bi-GAN enables the model to learn deep latent feature representations. This improvement enhances the model’s ability to accurately classify lung cancer vs. non-cancer cases. This characteristic of the proposed approach is advantageous in situations where there are few labeled medical cases, making it more practical in allowing the model to generalize to novel cases.

Experimental setup

The experiments were conducted on a Windows-based machine with an Intel® Core^TM i5-7200U CPU (2.50 GHz) and 4 GB RAM. The implementation was performed in Python, using deep learning frameworks such as TensorFlow, Keras, and PyTorch, along with essential libraries including Matplotlib, Scikit-learn, and Pandas for data handling and visualization. The dataset used for evaluation was obtained from Kaggle’s publicly available lung cancer dataset, which consists of 284 instances and 16 attributes, comprising both categorical and numerical variables. The methodology ensures that proper feature engineering, optimization techniques, and deep learning architectures are employed to maximize classification performance.

Dataset description

The lung cancer dataset used in this research is composed of 16 attributes, which is a combination of numerical and categorical variables (https://www.kaggle.com/datasets/akashnath29/lung-cancer-dataset). There are a total of 284 instances in the dataset, which aids in making a classification based on the cancer risk status. The dataset employed in the experimentation can be downloaded from the link given. A detailed description of the variables available in the dataset is presented in Table 2.

Experimental evaluation

One technique that can be used to comprehend the fundamental connections between different data characteristics in a collection of data is feature correlation. Determining the relationships that exist between the data attributes and the manner in which every attribute affects the resulting feature are just two applications where feature correlation can be helpful. In this work, the correlation between the features is computed using the correlation coefficients of a attribute vector W which is of the dimension $m\times n$ represented as $W=[w_1,w_2,\dots ,w_n]$ where $w_1,w_2,\dots ,w_n$ denotes the set of attributes, m denotes the length of each of the attribute and n denotes the total number of attributes. Figure 3 represents the resultant correlation matrix obtained for the given dataset. Based on the correlation analysis, it can be observed that there is a moderate positive correlation (0.29) between alcohol consumption and lung cancer which indicates that an individual with the habit of consuming alcohol is likelier to have lung cancer.

The correlation matrix as shown in Fig. 3 visually represents the relationships between different attributes in the dataset. The analysis reveals that alcohol consumption (0.29 correlation), wheezing and coughing (0.25 correlation), and fatigue (strong correlation) are among the most influential features in lung cancer classification. On the other hand, attributes such as age (0.11 correlation) and gender (0.054 correlation) exhibit weak relationships with lung cancer occurrence. These findings suggest that lifestyle factors and respiratory symptoms are stronger indicators of lung cancer risk compared to demographic characteristics. The correlation analysis helps in understanding the underlying patterns in the data, further reinforcing the significance of the FFXO-based feature selection approach in this study.

Features such as wheezing and coughing are moderately correlated (0.25) indicating that these kind of respiratory issues increases the risk of lung cancer. Another essential predictor identified from the correlation matrix is fatigue, which shows a strong correlation with lung cancer. The features having low correlation with lung cancer include age (0.11), gender (0.054) and chronic disease (0.14). Further, strong inter-feature correlations are found between anxiety and fatigue, which is 0.56 and anxiety and wheezing, which is 0.48. Features including fatigue, alcohol consumption, wheezing, and coughing that have stronger associations with lung cancer are probably crucial for lung cancer detection models. However, in this particular dataset, characteristics such as age and gender exhibit weak correlations and may not have as much of an influence on lung cancer prediction. The features selected by the FFXO algorithm (Table 3) include smoking, anxiety, fatigue, allergy, wheezing, coughing, alcohol consumption, shortness of breath, and chest pain, as these attributes exhibited strong correlations with lung cancer risk. The selection of these features ensures that the model focuses on the most relevant factors, thereby enhancing classification accuracy while eliminating irrelevant attributes that may introduce noise into the learning process. This feature selection process significantly contributes to the robustness of the proposed approach, as it ensures that only meaningful variables influence the classification outcomes.

Based on the correlation analysis and application of FFXO algorithm, the features selected for classification purposes are presented in Table 3. The features such as smoking, anxiety, fatigue, allergy, wheezing, coughing, alcohol consumption, shortness of breath and chest pain are considered to be the significant attributes for effective lung cancer categorization.

Performance assessment

The performance of the proposed system is assessed based on the metrics such as accuracy, F1-score, precision and recall. In a balanced dataset with about identical amounts of lung cancer and non-lung cancer samples, accuracy offers a clear assessment of how effectively the Bi-GAN model distinguishes between the two classes. The information set may be unbalanced in healthcare applications such as lung cancer detection, with a higher number of negative (healthy) instances than positive (cancer) instances. When this occurs, accuracy on its own may be deceptive since a model that classifies the majority of specimens as negative may yet have high accuracy even when it is unable to identify real instances of cancer. For this reason, other measures like F1-score, recall, and precision are required. A high precision score in the classification of lung cancer indicates that the Bi-GAN model minimises errors in diagnosis, an important feature that helps prevent inaccurate identification of cancer in healthy persons. By ensuring that the majority of lung cancer cases are accurately identified by the model, recall reduces the possibility of false negatives. Since the F1-score strikes a compromise between recall and precision, it is perfect for assessing the Bi-GAN in situations where false positives and false negatives might have serious repercussions.

The performance of the Bi-GAN model used in this research for lung cancer detection is compared against traditional ML and DL models along with optimization techniques as well as with the existing works in the literature. Initially, the performance of the Bi-GAN model is compared with the ML models without employing feature selection techniques and the outcomes are presented in Table 4 and Fig. 4. Decision tree model produced an accuracy of 87.6% with 86.9% F1 score. Random forest model exhibited an accuracy, precision, recall values of 88.9%, 87.4% and 87.6% respectively. Naïve Bayes classifier offered predictions with 90.8% accuracy and 89.6% precision levels. Logistic regression model was capable enough to produce predictions with 91.5% accuracy, 90.3% precision, 89.6% recall and 91.2% F1 score. AdaBoost classifier produced higher accuracy among other models with 92.7% accuracy and 91.3% precision. It was observed that the Bi-GAN model produced maximum accuracy of 94.3%, precision of 93.5%, recall of 93.7% and F1 score of 94.1%.

The comparison of machine learning models without feature selection (Table 4 and Fig. 4) demonstrates that traditional classifiers such as decision tree (87.6%) and random forest (88.9%) produce relatively lower accuracy levels. In contrast, the proposed Bi-GAN model achieves 94.3% accuracy, indicating its superior capability in recognizing lung cancer patterns. However, machine learning models often struggle with high-dimensional data, leading to suboptimal performance. This challenge is addressed in the subsequent experiment, where feature selection is applied before classification.

Further, the ML models were also assessed after applying selected features for performing classification and the results are presented in Table 5 and Fig. 5. It can be seen that the accuracy and other metric values for all the models have increased after using feature selection. The accuracy of decision tree was 89.6% with 88.2% precision value. Random forest and naïve Bayes classifier showed accuracy improvement as 90.6% and 92.3% respectively. Logistic regression model produced an accuracy, precision, recall and F1 score values of 93.6%, 92.1%, 92.4% and 93.3% respectively. AdaBoost classifier showed better accuracy in making predictions with 94.9%. However, the accuracy of Bi-GAN model also improved to 96.7% with precision and recall of 95.3% and 95.7% respectively.

When feature selection is incorporated (Table 5 and Fig. 5), the performance of all classifiers improves, confirming that selecting the most relevant attributes significantly enhances classification accuracy. Bi-GAN achieves 96.7% accuracy, outperforming other models such as AdaBoost (94.9%) and logistic regression (93.6%). This improvement highlights the effectiveness of FFXO in reducing dimensionality, ensuring that the classifier focuses only on the most important predictors. The results emphasize that a well-optimized feature selection process is crucial in improving lung cancer detection accuracy, making the classification process more efficient.

In addition to this, conventional DL models were also assessed and compared for lung cancer classification. Multi-layer perceptrons (MLP) produced accuracy of 90.6%, precision of 88.6%, recall of 89.2% and F1 score of 89.8%. Deep neural networks (DNN) produced higher accuracy of 92.7% with 90.3% precision value. Gated recurrent units (GRU) produced 93.6%, 92.3% and 92.7% of accuracy, precision and recall values correspondingly. Bayesian neural networks (BNN) showed lower accuracy for lung cancer detection with outcomes accurate with 91.8%. Deep belief networks (DBN) performed better than the other models and produced classifications with 95.5% accuracy. The detailed description of outcomes along with graphical representations are presented in Table 6 and Fig. 6.

The comparison with deep learning models (Table 6 and Fig. 6) further demonstrates the strength of Bi-GAN. Traditional deep learning models such as multi-layer perceptrons (90.6%), deep neural networks (92.7%), and gated recurrent units (93.6%) show competitive performance but still fall short of Bi-GAN’s 96.7% accuracy. The superior performance of Bi-GAN is attributed to its ability to generate synthetic data, thereby improving classification robustness and reducing overfitting issues commonly encountered in medical datasets.

The performance of the DL models in combination with different optimization algorithms such as Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Artificial Bee Colony Optimization (ABC), Ant Colony Optimization (ACO) and Cuckoo Search Optimization (CSO) were also analysed. The results of the analysis are shown in Table 7 and Fig. 7. MLP with GA was found to produce an accuracy of 91.9%, precision of 89.8%, recall of 89.3% and F1 score of 90.5%. DNN with PSO algorithm produced improved accuracy of 93.8% with 92.3% precision. GRU with ABC algorithm showed an accuracy enhancement of 94.6% with F1 score value of 94.1%. BNN with ACO exhibited accuracy of 92.3%, precision of 90.6%, recall of 91.4% and F1 score of 91.8%. DBN model showed much higher accuracy of 97.5% when optimized by CSO algorithm. Bi-GAN model with FFXO algorithm produced superior accuracy of 98.7% compared to conventional ML and DL algorithms with and without feature selection methods.

Finally, few research on lung cancer detection using deep learning approaches that are proposed by other researchers are compared with the outcomes of present research and the results are tabulated and depicted in Table 8 and Fig. 8. Wang et al. (2022) suggested the dual-modality deep learning network for classifying lung tumors and produced an accuracy of 95.8%. The research in Mukherjee et al. (2020) used shallow convolutional neural network to forecast the presence of lung carcinoma and the outcomes were 94.6% accurate with 93.5% precision and 92.7% recall. The research work recommended in Mkindu, Wu & Zhao (2023) 3D multi-scale vision transformer was capable enough to detect tumours in lung nodules with 96.8% efficiency. An accuracy level of 97.5% was achieved by applying Improved Swin Transformer for classifying lung cancer in Sun, Pang & Li (2023). Multiple attention 3D U-Net was utilised for segmenting lung tumors in Chen et al. (2021) and obtained an accuracy of 98.2% with 97.6% F1 score. The maximum accuracy of 98.7% is produced by the Bi-GAN model used in this research for lung cancer classification.

Limitations of current research

A highly intricate model may result from combining the FFXO algorithm with deep learning approach like Bi-GAN. High model complexity might make it more challenging to operate in real-time or resource-constrained situations, such as healthcare facilities or environments with limited resources, by increasing computing costs, storage utilisation, and training times. Insufficient or skewed data may restrict the model’s usefulness and cause poor generalisation to unobserved cases.