Bio-Inspired Spider Wasp Optimization-Based Convolutional Neural Network for Lung Cancer Classification

Lung cancer is one of the deadliest diseases that occur due to the abnormal growth of cells in the lungs, causing damage to healthy lung tissues [1]. The common cause of this cancer is tobacco usage and it accounts for nearly 80% of lung cancer deaths [2]. The study by the World Health Organization (WHO) on cancer reported that lung cancer has led to approximately 1.61 million deaths and it is more prominent among males than females. Early and precise identification of this disease is crucial for improving the patient's health and survival rate [3]. Currently, various imaging techniques including magnetic resonance imaging (MRI), Computed Tomography (CT), and X-ray are employed for lung cancer prediction [4]. Among these techniques, CT was widely used for lung cancer detection because of its effectiveness in capturing minute variations in the lung region [5]. Conventionally, the medical images are manually examined by healthcare professionals or radiotherapists to detect the presence of cancer [6]. However, this manual examination of cancer is time-consuming and it depends on the knowledge of the healthcare experts. Thus, researchers focused on developing an automatic framework leveraging advanced technologies such as artificial intelligence (AI), computer-aided processing, big data analysis, etc. [7]. In this automatic mechanism, the system was trained using the data containing both healthy and cancer features to understand the variations between them [8].

In recent times, AI approaches such as machine learning (ML) and deep learning (DL) have earned greater attention in image processing tasks. The AI-based models analyze the medical images and capture the features and hierarchical relations to detect the presence of lung cancer [9, 10]. Initially, the researchers utilized ML algorithms such as decision trees (DT), random forest classifier (RF), logistic regression, and support vector machine (SVM), for lung cancer detection [11, 12]. They process the medical images and train using either supervised or unsupervised learning strategies for classification tasks. The studies on ML models concluded that they offered automatic classification with increased accuracy than the manual process. Despite its increased performances, it faces challenges such as large computational demands, algorithmic bias, lack of adaptability, reduced accuracy in test cases, data dependency, etc. [13-15]. On the other hand, the DL models including deep neural network (DNN), CNN, and feed-forward neural network (FFNN) have shown promising results in image classification [16].

Among these neural networks, CNN is more suitable for medical image analysis because of its potential to capture features and relations within the images more effectively. Hence many researches are conducted on lung cancer classification using CNN models [17, 18]. The existing works developed techniques such as Fuzzy Particle Swarm Optimization with CNN [19], Deep CNN [20], 2-dimensional CNN [21], 3-dimensional CNN [22], autoencoder-based CNN [23], etc., for lung cancer detection. These strategies mainly focused on differentiating malignant and benign cells in the lung and achieved accuracy of around 86% to 95%. However, these frameworks face challenges such as computational complexity, minimum accuracy, instability, less generalization, and limited to binary classification. To address these issues, a novel optimized CNN was developed in this work for detecting and classifying lung cancer from CT images.

The major contributions of the presented framework are described below:

• This study develops a hybrid classifier by integrating a convolutional neural network with spider wasp optimization for lung cancer identification and classification.
• The proposed framework employs an attention neural network for capturing the informative features from the pre-processed images.
• A meta-heuristic-based feature selector was designed using the Waterwheel Plant Algorithm to optimally select the most relevant attribute from the extracted feature sequences.
• The CNN in the developed algorithm learns the hierarchical and spatial feature representations and recognizes the lung cancer pattern, while the SWO optimizes CNN’s training by refining its parameters to optimal value.
• The presented technique was executed in Python software and its outcomes are validated with conventional classifier models using metrics like accuracy, precision, false positive rate (FPR), recall, f-measure, specificity, FNR, and execution time.

An optimized CNN was developed to classify lung cancer using lung CT images. The presented framework comprises five modules: image acquisition, image preprocessing, feature analysis, feature selection, and optimized model training and classification. In the data collection module, the CT images of healthy and lung cancer-affected individuals are gathered and imported into the system. The accumulated images are preprocessed following steps like noise filtering, image resizing, and background removal. Also, the U-Net algorithm was used for segmenting the region of interest (ROI) from the preprocessed images. In the feature extraction module, ANN was designed to capture the relevant attributes from the segmented images, which reduces the dimensionality of the data and improves training speed. Consequently, the WPA-based feature selector was modeled to select the most informative and highly correlative features from the extracted feature sets, enabling the classifier to concentrate on major features for tumor classification. Finally, an optimized CNN was developed by integrating the efficiency of SWO into it. The SWO continuously refines and optimizes the hyperparameters of CNN, enabling it /to adapt to evolving changes and enhancing its learning capacity. The architecture of the developed lung cancer classifier model is displayed in Figure 1.

1.png

Figure 1. Architecture of SWoCNN for lung cancer classification

3.1 Image acquisition

The first stage of the developed framework is the lung CT image acquisition from both healthy and cancer-affected persons as shown in Figure 2. The gathered images are labeled to train the classifier to recognize the patterns of lung cancer. The presented study used the publicly available Lung CT scan database from Kaggle site [27].

2a.png

(a)

2b.png

(b)

Figure 2. Dataset images of different classes: (a) Lung cancer; (b) Normal

3.2 Image preprocessing

Image preprocessing defines the data preparation phase, where the system converts the raw CT images into an appropriate format for further processing. This stage follows a sequence of processes such as noise filtering, image resizing, and contrast enhancement or background elimination. In the noise filtering step, a Gaussian filter was applied to the images to eliminate the noise or unwanted random variations within each image. The Gaussian filter removes the noise attributes and replaces them with the average value of nearby or surrounding attributes, which is estimated by Gaussian distribution. The mathematical formulation of the Gaussian filter is expressed in Eq. (1):

$G s(m, n)=C_{i m}(m, n) \times \frac{1}{2 \pi \partial^2} e^{-\frac{m^2+n^2}{2 \partial^2}}$          (1)

where, Gs(m, n) represents the Gaussian filter, C_im(m, n) denotes the input CT image, (m, n) indicates the kernel coordinates, and ∂ refers to the standard deviation. Consequently, the filtered images are resized into a common size to boost the dataset consistency. Finally, background removal was done to discard the unwanted regions from the CT images. This process is done using image thresholding. This step helps the system to focus on useful regions in the images. In addition, data augmentation was performed to reduce the risk of overfitting. This step creates new training samples by performing steps like rotation, random translation, scaling, image mixing, etc. Figure 3 displays the preprocessed images.

3.png

Figure 3. Preprocessed images

Further, image segmentation was done using the U-Net algorithm to isolate ROI from the preprocessed images [28]. This technique has the potential to segment the regions with training using less data. It consists of three main elements namely: encoder, bottleneck, and decoder. The encoder component down-samples the preprocessed images through convolution and pooling operations and maps the regions within each image. The bottleneck is placed between encoder and decoder and it maps the most informative regions while retaining spatial information. The encoder performs upsampling and reconstructs the image from the feature representation. The reconstructed image is the segmentation output. Figure 4 displays the segmentation outcomes.

4.png

Figure 4. Segmentation images

3.3 Feature extraction

An attention model is a kind of DL approach, which tends to concentrate on significant parts of the input data. In the proposed work, it was used to capture the relevant features from the segmented results. Feature extraction is one of the crucial steps in cancer classification, which aims to minimize the data dimension and assists the classifier in quickly learning the feature representations [29]. The architecture of ANN is similar to the structure of the encoder-decoder and finds the relevant features through the calculation of attention weights. It includes three main components namely: encoder, attention, and decoder. The encoder accepts the segmented images as input and generates hidden states (features present in the segmented images). The attention module determines the relevance between the encoder’s hidden state and the target hidden state and produces similarity scores as outcomes, which are mathematically expressed in Eq. (2):

$S w_{x t}^y=\omega\left(\tanh \left(w_e h_{x t}^y+b_{w e}\right)\right)+\beta$          (2)

where, $S w_{x t}^y$ represents the similarity score, ω denotes associated with ANN, tanh defines the activation function, $h_{x t}^y$ indicates the hidden state generated by the encoder, b_we denotes the bias vector associated with the attention component, w_e represent the weight vector of the attention module, and β indicates the bias vector of ANN. The resultant of the attention component is converted into attention weights using a Softmax function, as defined in Eq. (3):

$A w_t={soft} \max \left(S w_{x t}^y\right)$          (3)

where, Aw_t denotes the attention weights, which represents the relevance of the extracted feature for lung cancer categorization. These weights are applied to the encoder's hidden state and all the relevant information is extracted based on their importance through a weighted sum function, as represented in Eq. (4):

$W_{s t}=\sum_{t=1}^{T_m} A w_t h_{x t}^y$          (4)

where, W_st denotes the weighted sum and T_s indicates the maximum iteration. The decoder component receives the weighted sum and transforms it into a feature vector. This extracted feature vector contains all important information related to lung cancer classification.

3.4 Feature selection

Feature selection represents the process of selecting the most relevant attributes while eliminating the irrelevant ones. This step aims to enhance interpretability, reduce complexity, and increase accuracy. In the developed work, a novel feature selector was modeled using the meta-heuristic optimization algorithm named “Waterwheel plant algorithm [30].” The WPA is a bio-inspired approach modeled based on the natural characteristics of waterwheel plants to solve complex optimization problems. This algorithm is mathematically designed following the hunting expedition behavior of waterwheel plants. They use search agents to find the prey. Here, the concept of this algorithm is applied to select the most relevant and highly informative features for classifier training. The feature selection process begins with the initialization of extracted feature sets in the search space. In WPA, the extracted feature sets by ANN are expressed as a matrix using Eq. (5) and each feature is initialized in the search space using Eq. (6):

$F_e=\left[\begin{array}{l}F_{e l} \\ \vdots \\ F_{e i} \\ \vdots \\ F_{e Q}\end{array}\right]_{Q \times R}=\left[\begin{array}{cccc}f_{e(1,1)} & \ldots & f_{e(1, j)} & \dots & f_{e(1, R)} \\ \vdots & \ddots& \vdots & \ddots & \vdots  \\ f_{e(i, 1)} & \ldots & f_{e(i, j)} & \ldots &  f_{e(i, R)} \\ \vdots & \ddots & \vdots & \ddots & \vdots  \\ f_{e(Q, 1)} & \ldots & f_{e(Q, j)} & \ldots & f_{e(Q, R)}\end{array} \right]$          (5)

$\begin{gathered}f_{e(i, j)}=L_{b j}+\kappa_{i, j} \cdot\left(U_{b j}-L_{b j}\right) \\ i=1,2,3, \ldots \ldots, Q, j=1,2,3, \ldots ., R\end{gathered}$          (6)

where, F_e indicates the extracted feature sets, F_ei denotes the i^th feature (candidate solution), f_e(i,j) represents the j^th problem variable, κ_i,j represents a random number (0, 1), U_bj and L_bj refers to the upper bound and lower bound. Further, the fitness value was determined for each feature based on their relevance to the tumor classification task, which is defined in Eq. (7):

$\gamma_v=\gamma_v\left[\begin{array}{l}F_{e 1} \\ \vdots \\ F_{e i} \\ \vdots \\ F_{e Q}\end{array}\right]_{Q \times R}$          (7)

The candidate solution with a high fitness value indicates the best or optimal solution and the solution with the lowest value denotes the worst. After fitness evaluation, the next step is exploration, where the system explores the search space and updates its initial values. This exploration helps to find the optimal solution and mitigates the possibility of local optima. The exploration phase of WPA is modeled using Eqs. (8) and (9):

$\vec{A}=\vec{d} \cdot\left(f_e(t)+2 \rho\right)$          (8)

$f_e(t+1)=f_e(t)+\vec{A} \cdot\left(2 \rho+\vec{d}_1\right)$          (9)

where, $\vec{A}$ denotes the direction of moving in the search space, f_e(t+1) indicates the updated solution, f_e(t) defines the current solution, $\overrightarrow{d_1} \vec{d}$ refers to the random vector, and $\rho$ represents the optimization factor. Consequently, the system exploits the current solutions to estimate the optimal features. This exploitation phase follows the hunting and feeding characteristics of WPA and it is formulated in Eqs. (10) and (11):

$\overrightarrow{A^{\prime}}=\vec{d}_2 \cdot\left(\rho \cdot f_e^{\prime}(t)+\vec{d}_3 f_e(t)\right)$          (10)

$f_e(t+1)=f_e(t)+\rho \overrightarrow{A^{\prime}.}$         (11)  

where, $\overrightarrow{A^{\prime}}$ denotes the direction of exploitation, $\vec{d}_2$ and $\vec{d}_3$ indicates the random vector. Further, fitness value was determined for the updated solutions. Then, the feature with a fitness value greater than 0.5 is selected and this process continuous until reaching maximum iteration or maximum convergence rate. The selected features are fed into the classifier for training.

3.5 Optimized CNN for lung cancer classification

A meta-heuristic optimizer-based CNN was developed to accurately identify and categorize lung cancer by processing the CT images [31]. The developed algorithm incorporates the optimization and exploration characteristics of SWO into CNN, enabling the system to adapt to the evolving variations in the image features and to train the model with optimal values. The CNN is a popular DL model used for performing tasks such as object detection, speech recognition, video labeling, image classification, facial recognition, etc. It consists of five important layers namely: input, convolutional, pooling, and fully-connected (FC) layers. The input layer accepts the selected feature sets from the WPA module as classifier input and transforms them into a suitable format for feature learning. The convolutional layer is the building block of CNN, which is responsible for understanding the patterns within each image. It applies a sequence of kernel filters on the input data and creates a feature map, which represents the patterns influencing healthy and lung cancer-affected images. The feature map is mathematically expressed in Eq. (12):

$F_{m a p}(p, q)=\sum_{i=0}^{x-1} \sum_{j=0}^{y-1} S_f(x, y) * K(p-x, q-y)$         (12)

where, F_map(p, q) indicates the feature map at the position (p, q), and K(p-x, q-y) represents the kernel filter applied at the location (p-x, q-y). Each convolutional layer is followed by a Rectified Linear Unit (ReLU) activation function to introduce non-linearity in the system. After each convolution+ReLU combination, a pooling layer was placed to reduce the spatial dimension, while retaining the significant information. This layer aims to minimize the memory and helps prevent overfitting by performing pooling operations. The pooling operation defines the process of sliding a 2-D filter over each channel of the feature map. In the designed CNN, a max pooling layer was used, which chooses the maximum element from the feature map covered by the 2-D filter. After multiple convolutional and pooling layers, the resultant feature map was flattened into a one-dimensional vector. This learned feature vector contains the patterns and relations between them, enabling them to categorize the lung cancer classes. The FC layer accepts this feature and computes the lung cancer classification through a Softmax function. The Softmax function returns the possibility of the input image belonging to the class, as represented in Eqs. (13) and (14):

$P_{v l}\left(C_{i m}\right)=\operatorname{Soft} \max \left(W_{f c} * f_{l e}\right)+b^{\prime}$         (13)

$C_{c l}= \begin{cases} { if }\left(0>P_{v l} \leq 0.2\right), & { normal } \\  { if }\left(P_{v l} \geq 0.3\right), & S C C\end{cases}$         (14)

where, P_vl indicates the probability value, C_cl denotes the classification function, W_fc defines the weight matrix of the FC layer, f_le represents the flattened feature vector, and b' refers to the bias vector of the FC layer. Consequently, the loss or error was determined to estimate the efficiency of the classifier and it is formulated in Eq. (15):

Loss $=\frac{1}{S} \sum_{i=1}^S \sum_{j=1}^C Y_i^j \cdot \log \left(Y_i^{\prime j}\right)$         (15)

where, $Y_i^j$ denotes the ground truth outcome, S refers to the number of training samples, C indicates the number of classes, and ${Y^{\prime}}_i^j$ represents the predicted class. Reducing this loss is significant for improving classification accuracy. Also, the CNN's performance depends on the value of its hyperparameters such as several filters, kernel size, hidden layer, learning rate, batch size, weights, and bias vectors. To optimally select the values of these parameters, the SWO was integrated into CNN. The SWO is a meta-heuristic optimization algorithm developed based on the searching, hunting, mating, and nesting characteristics of female wasps [32]. The female wasps usually do these processes by depositing solitary egg within each spider’s abdomen. In the beginning stage of optimization, the female wasps explore their surroundings for finding suitable spiders. Then, they immobilize and pull them to make nests. After identifying appropriate nests and finding prey, they intend to drag the prey into nests. Unlike conventional optimization models such as genetic algorithm (GA), particle swarm optimization (PSO), grey wolf optimization (GWO), etc., the SWO dynamically switches between exploration and exploitation phases, leading to faster convergence and prevents premature convergence. In addition, the SWO model is more robust and efficient in optimizing the complex and high-dimensional spaces. Moreover, it offers diversified search options, which avoids the local optima. Since the SWO was used for selecting the optimal values in the search range, only the hunting phase is used in the designed model. The mating and nesting phases are not considered as they don’t directly relate with parameter update task, they are intended to reproduction and settlement of wasps in the biological systems The proposed optimization involves four phases: (i) initialization, (ii) fitness evaluation, (iii) searching phase, and (iv) parameter selection and termination.

i. Initialization

The SWO algorithm commences with the initialization of the female wasps. Each female wasp represents the CNN parameter set and it is initialized with a random value in the search space. In the context of CNN parameter tuning, female wasp defines the hyperparameter population, and prey denotes their optimal value The wasp population and the random initialization process are defined in Eqs. (16) and (17):

$\operatorname{Pr}=\left[\begin{array}{cccc}\operatorname{Pr}_{(1,1)} & \operatorname{Pr}_{(1,2)} & \ldots & \operatorname{Pr}_{(1, d)} \\ \operatorname{Pr}_{(2,1)} & \operatorname{Pr}_{(2,2)} & \ldots & \operatorname{Pr}_{(2, d)} \\ \vdots & \vdots & \vdots & \vdots \\ \operatorname{Pr}_{(c, 1)} & \operatorname{Pr}_{(c, 2)} & \ldots & \operatorname{Pr}_{(c, d)}\end{array}\right]$        (16)

$\overrightarrow{\operatorname{Pr}}_t=\vec{L}_u+\vec{r}\left(\vec{K}-\vec{L}_u\right)$          (17)

where, Pr indicates the parameter population, $\vec{L}_u$ $\vec{K}$ denotes the lower and upper bound of the search space and $\vec{r}$ represents the random initialization vector.

ii. Fitness evaluation

For each parameter set, fitness was calculated based on the predefined objective function. The objective function of the SWO is to reduce the loss incurred by CNN, and it is formulated in Eq. (18):

$O b j_{f u n}=\min \left(\frac{1}{S} \sum_{i=1}^S \sum_{j=1}^C Y_i^j \cdot \log \left(Y_i^{\prime j}\right)\right)$          (18)

The parameter set with high fitness indicates the optimal solution and vice versa. Then, the best candidate solution was estimated by sorting them as per their fitness value.

iii. Searching phase

This phase follows how female wasps search for the most suitable spiders for feeding their offspring. They use a constant step for searching for prey by randomly looking into the search space. Similarly, in the developed work, they look for the optimal value for the parameter set in the search range. In this searching phase, the values of the parameter set will be updated in two ways. The random exploration of search space and parameter update is formulated in Eq. (19):

${Pr}(t+1)={Pr}(t)+f .\left({Pr}_a(t)-{Pr}_b(t)\right)$          (19)

where, Pr(t) denotes the current parameter set, Pr(t+1) denotes the updated parameter set, ϕ represents the constant motion, a and b indicates two randomly selected indices to choose the exploration direction. In some instances, the optimization process drops and selects the local optima. To prevent this, the SWO algorithm explores the surrounding search space with a small step size, as expressed in Eq. (20):

${Pr}(t+1)={Pr}(t)+\phi \cdot \vec{L}_u+\vec{r}\left(\vec{K}-\vec{L}_u\right)$          (20)

where, $\vec{r}$ defines the random vector [0, 1]. In searching phase, the updated hyperparameters in the search space are determined using Eq. (21):

${Pr}(t+1)=\left\{\begin{array}{l}{Pr}(t)+f \cdot\left({Pr}_a(t)-{Pr}_b(t)\right) ; \text { if } c<g \\ {Pr}(t)+\phi \cdot \vec{L}_u+\vec{r}\left(\vec{K}-\vec{L}_u\right) ;  { otherwise }\end{array}\right.$          (21)

where, χ and γ defines the random number ranging [0, 1], representing the probability of obtaining local optima in the searching phase.

$\begin{gathered}{Pr}(t+1)={Pr}(t) +G \times\left|2 \times r_e \times P r_a(t)-P r_b(t)\right|\end{gathered}$          (22)

where, G denotes the deviation between the current solution and the best solution, and r_e indicates the randomness in escaping in range [0, 1]. The formula for calculating G is expressed in Eq. (23):

$G=r_e \times\left(2-2 \times\left(\frac{t}{T_m}\right)\right)$          (23)

where, t indicates the current iteration, and T_m denotes the maximum iteration. This stage is the beginning of the exploitation stage in SWO approach.

v. Hunting phase

This phase follows how the female wasps move towards the region where the best spider (prey) can be found. The hyperparameter enters this stage only when its fitness is less than 0.3 (near optimal) and they are updated by exploiting locally. Instead of searching randomly, the algorithm adjusts current solutions by shifting them closer to the high-quality solutions, as formulated in Eq. (24):

$\begin{gathered}{Pr}(t+1)={Pr}(t)^{\prime}  +\cos (2 p l) \times\left({Pr}(t)^{\prime}-{Pr}(t)\right)\end{gathered}$          (24)

where, Pr(t)' defines the best available solution. After updating hyperparameters, the fitness value was determined for them for selection purpose.

vi. Parameter selection and termination

If the fitness of the updated sequence is greater than the old fitness, the updated sequence was selected for modeling CNN. If the new fitness is less than the old fitness, no changes in CNN design the parameter selection is formulated in Eq. (25):

${Pr}= \begin{cases}{ if }\left(f_t^*<f_t\right) ,\ {Pr}(t+1) \\ { else } ,\ {Pr}(t)\end{cases}$          (25)

where, $f_t^*$ denotes the fitness of the updated hyperparameter, and f_t indicates the fitness of the current solution. This process continues until it reaches maximum iteration. Algorithm 1 presents the working of the developed classifier in pseudocode format. Figure 5 presents the flowchart of the SWO algorithm for CNN parameter update.

5.png

Figure 5. The flowchart of the SWO algorithm for CNN parameter update