Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2s (2025) 

 

423 https://internationalpubls.com 

WTBAN-CNN-TL: Wavelet Transform-Based Feature Extraction and 

Bonferroni Adjusted ANOVA-CNN with Transfer Learning for Lung 

X-Ray Image Classification 

 

Kavitha S1 and Dr.H.Hannah Inbarani2 
1Research Scholar, Department of Computer Science, Periyar University, Salem, Tamilnadu, India. 

kavitha.chanjhana@gmail.com 

2Associate Professor, Department of Computer Science, Periyar University, Salem, TamilNadu, India. 

hhinba@periyaruniversity.ac.in 

 

Article History: 

Received: 20-09-2024 

Revised: 28-10-2024 

Accepted: 08-11-2024 

Abstract:  

A comprehensive approach for lung X-ray image classification is proposed, combining 

wavelet transform-based feature extraction with Bonferroni ANOVA feature selection and 

deep learning techniques. Features are initially extracted from lung X-ray images using 

wavelet transform, capturing crucial information across multiple scales. These features are 

concatenated to form a comprehensive feature set. To optimize this feature set, ANOVA is 

employed with Bonferroni adjusted feature selection to control the false positive rate and to 

identify the most relevant features, thereby reducing the dimensionality of the input data. 

This optimization can result in expedited training times and reduced computational 

resources required by the CNN. The selected features are subsequently fed into a 

Convolutional Neural Network (CNN) for classification. To further boost the CNN's 

performance, transfer learning is utilized by leveraging pre-trained models. Additionally, 

the performance of ANOVA-based feature selection is also compared with that of a genetic 

algorithm-based method. This integrated method, referred to as WTBAN-CNN-TL, aims to 

reduce the training times and resource requirements while improving the accuracy and 

efficiency of lung X-ray image classification. 

Keywords: Wavelet Transform, Bonferroni ANOVA (Analysis of Variance), Feature 

extraction, Feature selection, WTBAN-CNN-TL (Wavelet Transform Bonferroni ANOVA 

Enhanced Convolutional Neural Network with Transfer Learning) 

 

Introduction 

Medical imaging, particularly the analysis of lung X-ray images, holds paramount importance in the 

diagnosis and treatment of various respiratory ailments, including infections, tumors, Covid-19 and 

Pneumonia related diseases[1][2]. In recent years, the fusion of advanced computational techniques 

with traditional diagnostic methods has led to significant advancements in the field of medical image 

analysis [3][4]. This paper introduces a comprehensive approach aimed at refining the classification 

accuracy and efficiency of lung X-ray images through a synergy of wavelet transform-based feature 

extraction, Bonferroni adjusted ANOVA feature selection, and state-of-the-art deep learning 

methodologies. 

The proposed methodology capitalizes on the inherent advantages of wavelet transform, a powerful 

signal processing technique renowned for its ability to decompose complex signals into frequency 

sub-bands at different scales. By applying wavelet transform [5] to lung X-ray images, we can 

mailto:hhinba@periyaruniversity.ac.in


Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2s (2025) 

 

424 https://internationalpubls.com 

capture crucial information embedded within the images across multiple resolutions, thereby 

facilitating the extraction of discriminative features indicative of various pulmonary conditions [6]. 

These features, extracted at different frequency scales, are subsequently concatenated to construct a 

comprehensive feature set that encapsulates the intricate details essential for accurate classification. 

In addition to feature extraction, the proposed approach integrates ANOVA [7] feature selection 

algorithm with Bonferroni correction to refine the feature set further. To enhance the efficiency and 

effectiveness of the classification process, ANOVA (Analysis of Variance) with Bonferroni 

correction is utilized to control the false positive rate for feature selection. This statistical method 

identifies the most relevant features, thereby reducing the dimensionality of the input data and 

optimizing the computational resources required by the CNN. This reduction in dimensionality not 

only expedites training times but also mitigates the risk of overfitting, thereby improving the overall 

performance of the model. The optimized feature set, curated through the joint efforts of wavelet 

transform and ANOVA, serves as the input to a Convolutional Neural Network (CNN), a class of 

deep learning models renowned for their ability in extracting hierarchical representations from 

complex data. The CNN is tasked with learning discriminative patterns and relationships inherent in 

the feature-rich lung X-ray images, ultimately facilitating accurate classification across diverse 

pathological conditions. 

Moreover, to augment the CNN's learning capabilities and foster generalization across datasets, 

transfer learning is leveraged. This approach involves transferring knowledge from pre-trained 

models trained on large-scale datasets to our specific task of lung X-ray image classification. By 

initializing the CNN with pre-learned parameters and fine-tuning its architecture on our dataset, the 

wealth of information encapsulated within the pre-trained models is utilized, thereby enhancing the 

CNN's performance and accelerating the convergence of the training process. 

2. Related works 

Numerous studies have investigated various techniques to enhance the accuracy and efficiency of 

medical image classification, especially for lung X-rays. A significant emphasis has been placed on 

leveraging deep learning methods and feature extraction algorithms to improve diagnostic 

capabilities. Table I summarizes these related works. 

Table I – Related works 

            Authors                Techniques                    Result 

 

Reduwan, N.H. et 

al (2024) [8] 

 

 

 

 

Keshta, I et al. 

(2023) [9] 

 

. 

Feature selection techniques combined 

with deep learning models for external root 

resorption identification. 

 

Integration of ANOVA, RReliefF and 

Random Forest feature selection algorithm 

for cancer detection in Microarray data 

 

Discrete Wavelet Transform and Capsule 

 

Recursive Feature selection with 

VGG model achieved higher 

accuracy. 

 

 

 

The proposed approach 

demonstrates that feature 

selection improves the 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2s (2025) 

 

425 https://internationalpubls.com 

 

 

 

Monday HN et al 

(2022) [10] 

 

 

 

 

Nasiri, Hamid & 

Alavi, Seyed Ali. 

(2022) [11] 

 

 

 

Samee, N.A. et al. 

(2022) [12] 

 

 

 

 

Zheng Chen et al. 

(2020) [13] 

 

 

 

 

Liu et al. (2021) 

[14] 

 

 

 

 

Sarhan et al. 

(2020) [15]  

Rasheed et 

al.(2021 [16] 

Network for classifying pneumonia related 

diseases using XRay images. 

 

DenseNet169 and ANOVA Feature 

selection and XGBoost for classification of 

chest X-Ray images 

 

 

AlexNet, VGG, and GoogleNet for feature 

extraction from the INbreast mammograms 

and univariate approach for feature 

selection 

 

Feature selection integrated with Deep 

Neural Networks. 

 

 

 

Wavelet based convolutional wavelet 

neural network by replacing the fully 

connected neural network with wavelet 

neural network. 

 

Wavelet decomposition and CNN for lung 

cancer detection using CT images. 

 

 

Wavelets transform and Transfer learning 

to detect cancer in mammograms. 

 

 

classification accuracy. 

 

 

Accuracy 99.6% 

 

 

 

 

Two class classification accuracy 

– 98.72% and Multiclass 

classification accuracy – 92% 

 

 

 

Classification accuracy 98.50% 

 

 

 

SVM-RFE (Recursive Feature 

Elimination) Feature selection 

achieved better accuracy. 

 

 

Average accuracy – 96.57% 

 

 

 

 

Accuracy 99.5% 

 

 

 

 

Resnet50 achieved better 

accuracy than VGG16, 

GoogleNet and AlexNet. 

 

 

 

3. Research motivation 

         The motivation for this research is driven by the critical need for precise and efficient 

diagnostic tools in medical imaging, particularly for lung diseases such as pneumonia, COVID-19, 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2s (2025) 

 

426 https://internationalpubls.com 

and other lung opacities. Accurate diagnosis is crucial for effective treatment, yet traditional methods 

relying on radiologist interpretation can be slow and variable. While deep learning, especially 

Convolutional Neural Networks (CNNs), shows promise in automating diagnosis, challenges such as 

high computational demands and the need for large datasets persist. This study proposes a novel 

approach that integrates wavelet transform-based feature extraction with modified ANOVA feature 

selection, and deep learning techniques to address these issues. Wavelet transforms effectively 

capture multi-scale information from lung X-rays, and Bonferroni adjusted ANOVA help in selecting 

the most relevant features with less false positive rate, reducing data dimensionality and 

computational load. Furthermore, transfer learning leverages pre-trained CNN models to enhance 

classification accuracy and efficiency. This integrated method aims to provide a robust, scalable, and 

accurate tool for lung X-ray classification, improving diagnostic consistency and patient outcomes. 

4. Methodology 

The proposed methodology integrates wavelet transform-based feature extraction and concatenation, 

Bonferroni adjusted ANOVA feature selection, and deep learning techniques for lung X-ray image 

classification. The following steps outline the comprehensive approach: 

4.1 Data Acquisition and Preprocessing: 

The experimental dataset consists of 4,000 lung X-ray images sourced from the COVID-19 

Radiography Database on Kaggle [17]. The images are organized into four categories: COVID-19 

pneumonia (Class 0), Normal (Class 1), Viral pneumonia (Class 2), and Lung opacity (Class 3), with 

each category containing 1,000 images, ensuring a balanced dataset for comprehensive analysis and 

model training. All images are standardized to a resolution of 299x299 pixels. For the purposes of 

model training and evaluation, the dataset has been divided into training and testing sets with a 90-10 

split, respectively. Figure 1 illustrates representative samples from each category used in the 

analysis. 

 

Fig.1 Sample Images from Lung X-Ray data set (a) Covid (b) Normal (C) Viral Pneumonia (D) Lung 

Opacity 

The preprocessing of the dataset involved two main steps: normalization and resizing. First, the input 

images were normalized to scale the pixel values between 0 and 1. This normalization enhances the 

performance of pre-trained models by standardizing the input data. The normalization is performed 

using the formula: 

                                                  𝑋𝑖 =
𝑋𝑖

max (𝑋)
                                             (1) 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2s (2025) 

 

427 https://internationalpubls.com 

where 𝑋𝑖 represents an individual pixel value and max(X) is the maximum pixel value in the image. 

Following normalization, the images were resized to 224x224 pixels to match the input dimensions 

required by the pre-trained models. 

4.2 Wavelet Transform-Based Feature Extraction and Concatenation: 

The Haar wavelet transform [18] is a type of discrete wavelet transform (DWT) that is particularly 

effective for image processing tasks due to its simplicity and efficiency. The Haar wavelet 

decomposes an image into different frequency components, allowing the capture of both spatial and 

frequency information. This decomposition helps in identifying various levels of detail within the 

image, which are essential for distinguishing between different classes of lung conditions.  

Each X-ray image from the dataset is initially decomposed into four sub-bands using two level 

‘Haar’ wavelet decomposition. The first level Haar wavelet transform decomposes X into four 

subbands: approximation (𝐴(1)), horizontal detail(𝐻(1)), vertical detail(𝑉(1)), and diagonal 

detail(𝐷(1)). The (𝐴(1)) sub-band captures the low-frequency information, representing the coarse 

approximation of the image. In contrast, the(𝐻(1)), (𝑉(1)), and (𝐷(1)) sub-bands contain high-

frequency information, detailing the edges and textures of the image. The extracted coefficients from 

these sub-bands encapsulate the image's structure and texture across different scales and orientations. 

It is calculated as: 

                             𝐴(1) =
𝑋2𝑖−1,2𝑗−1+𝑋2𝑖−1,2𝑗+𝑋2𝑖,2𝑗−1+𝑋2𝑖,2𝑗

4
                                     (2)                                   

                             𝐻(1) =
𝑋2𝑖−1,2𝑗−1−𝑋2𝑖−1,2𝑗+𝑋2𝑖,2𝑗−1−𝑋2𝑖,2𝑗

4
                                     (3)                                

                             𝑉(1) =
𝑋2𝑖−1,2𝑗−1+𝑋2𝑖−1,2𝑗−𝑋2𝑖,2𝑗−1−𝑋2𝑖,2𝑗

4
                                      (4)                              

                            𝐷(1) =
𝑋2𝑖−1,2𝑗−1−𝑋2𝑖−1,2𝑗−𝑋2𝑖,2𝑗−1+𝑋2𝑖,2𝑗

4
                                       (5)                                 

Here, the indices i and j run from 1 to m/2 and n/2 respectively. The second level transform is 

calculated as: 

                             𝐴(2) =
𝐴
2𝑖−1,2𝑗−1

(1)
+𝐴

2𝑖−1,2𝑗
(1)

+𝐴
2𝑖,2𝑗−1

(1)
+𝐴

2𝑖,2𝑗
(1)

4
                          (6) 

                             𝐻(2) =
𝐴
2𝑖−1,2𝑗−1

(1)
−𝐴2𝑖−1,2𝑗

(1)
+𝐴2𝑖,2𝑗−1

(1)
−𝐴

2𝑖,2𝑗
(1)

4
                          (7) 

                            𝑉(2) =
𝐴
2𝑖−1,2𝑗−1

(1)
+𝐴

2𝑖−1,2𝑗
(1)

−𝐴2𝑖,2𝑗−1

(1)
−𝐴

2𝑖,2𝑗
(1)

4
                           (8) 

                           𝐷(2) =
𝐴
2𝑖−1,2𝑗−1

(1)
−𝐴2𝑖−1,2𝑗

(1)
−𝐴

2𝑖,2𝑗−1

(1)
+𝐴

2𝑖,2𝑗
(1)

4
                            (9) 

To form a comprehensive feature set, the coefficients from the LL, LH, HL, and HH sub-bands are 

concatenated. This, results in a unified feature vector comprising 32768 features. This extensive 

feature set provides a holistic representation of the image, encompassing both the overall shape and 

structure of the lungs (from the LL coefficients) and the detailed edges and textures (from the LH, 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2s (2025) 

 

428 https://internationalpubls.com 

HL, and HH coefficients). This comprehensive approach ensures that the feature set includes a wide 

range of information, making it robust for further analysis or machine learning tasks. 

The concatenation of wavelet coefficients involves combining all the coefficients obtained from the 

decomposition into a single feature vector.  Discrete Wavelet Transform is applied on each feature to 

obtain a set of coefficients for each feature. Mathematically, for feature i and level j: 

                       𝑐𝑜𝑒𝑓𝑓𝑠𝑖 = [𝐴𝑖,1, 𝐷𝑖,1, 𝐴𝑖,2, 𝐷𝑖,2, … . . 𝐴𝑖,𝑗, 𝐷𝑖,𝑗]                          (10) 

For all features, the transformed dataset 𝑋𝑡𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚𝑒𝑑 is: 

           𝑋𝑡𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚𝑒𝑑 = [𝑐𝑜𝑒𝑓𝑓𝑠1 ||   𝑐𝑜𝑒𝑓𝑓𝑠2  || … … || 𝑐𝑜𝑒𝑓𝑓𝑠𝑚]               (11) 

   Here, || denotes concatenation.  

4.3 Bonferroni-Adjusted ANOVA Feature Selection: 

After the wavelet transform-based feature extraction and concatenation process, a substantial number 

of features are generated. The concatenated feature set, comprising 32,768 features, exhibits high 

dimensionality, potentially impeding subsequent processing steps. To address this challenge, it 

becomes imperative to manage the dimensionality effectively to ensure efficient analysis. Employing 

ANOVA (Analysis of Variance) becomes crucial in this context to select the most pertinent features 

from the extracted feature set. ANOVA meticulously evaluates the statistical significance of each 

feature concerning the target variable, discerning those features that contribute significantly to the 

classification task.          For a single feature 𝑋𝑖, the ANOVA F-value is calculated as follows: 

                                   𝐹𝑖 =
𝐵𝑒𝑡𝑤𝑒𝑒𝑛 𝑔𝑟𝑜𝑢𝑝 𝑣𝑎𝑟𝑖𝑎𝑛𝑐𝑒 𝑓𝑜𝑟 𝑒𝑎𝑐ℎ 𝑓𝑒𝑎𝑡𝑢𝑟𝑒 𝑋𝑖

𝑊𝑖𝑡ℎ𝑖𝑛 𝑔𝑟𝑜𝑢𝑝 𝑣𝑎𝑟𝑖𝑎𝑛𝑐𝑒 𝑓𝑜𝑟 𝑓𝑒𝑎𝑡𝑢𝑟𝑒 𝑋𝑖
                                  (12) 

Where 𝐹𝑖 is the ANOVA F-value for feature 𝑋𝑖, Between-group variance for feature 𝑋𝑖 Measures the 

variance of feature 𝑋𝑖 across different classes of the target variable. Within-group variance for 

feature 𝑋𝑖 Measures the variance of feature 𝑋𝑖 within each class of the target variable. The ANOVA 

F-value is computed for each feature, and the corresponding p-value (Probability Value) for each 

feature is returned. The Bonferroni correction is applied to adjust the p-values for multiple 

comparisons to control the error rate. It is modified as:                                          

                                       𝑝𝑎𝑑𝑗𝑢𝑠𝑡𝑒𝑑 = min (𝑝. 𝑚, 1)                                                      (13) 

where m is the number of features. By exclusively retaining the most informative features, 

Bonferroni-Adjusted ANOVA effectively control the false positive rate and curtails the 

dimensionality of the input data to 16384 features. This reduction in dimensionality not only 

expedites the training process but also minimizing computational resources while retaining the most 

informative features.  In addition to ANOVA, a genetic algorithm (GA) based feature selection 

method is employed for comparison. GA iteratively evolves a population of candidate feature subsets 

using principles inspired by natural selection. The fitness of each candidate subset is evaluated based 

on its performance in a deep learning models trained on the selected features.       

 

           



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2s (2025) 

 

429 https://internationalpubls.com 

4.4 Proposed WTAN-CNN with Transfer Learning  

In this study, transfer learning model VGG19 [19] is employed to train on features selected using the 

Bonferroni-Adjusted ANOVA feature selection method. These features are derived from the 

concatenation of features extracted using the Haar wavelet transform. Utilizing the pre-trained 

weights of the model has proven to be a more effective strategy, significantly accelerating the 

training process. VGG19, a CNN model proposed by Zisserman and his team, is 19 layers deep. It 

excels at capturing detailed features by utilizing 5x5 and 3x3 receptive fields, thereby enhancing 

prediction accuracy. 

In the proposed work, a common approach is adopted where the top layers of these models are frozen 

and then modified by adding three dense layers and one dropout layer. The final dense layer consists 

of four neurons, corresponding to our four target output classes: Class 0 (COVID), Class 1 (Normal), 

Class 2 (Viral Pneumonia), and Class 3 (Lung Opacity). Additionally, a deep learning model, 

specifically a CNN, has been trained on the same features to compare its performance with the 

transfer learning models. Both the transfer learning models and the CNN have been fine-tuned on the 

dataset to optimize their performance in classifying the lung X-ray images into their respective 

categories. 

Fig.2. illustrates the comprehensive structure of the proposed model, showcasing the integration of 

wavelet-based feature extraction, Bonferroni-Adjusted ANOVA feature selection, and the 

architecture modifications of the transfer learning models. This integrated approach aims to improve 

the accuracy and efficiency of lung X-ray image classification, providing a robust tool for medical 

diagnostics. 

 

Fig.2. Structure of proposed WTBAN-CNN-TL 

 

 

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2s (2025) 

 

430 https://internationalpubls.com 

 

Algorithm: Wavelet-Bonferroni ANOVA-CNN with Transfer Learning for Lung X-Ray Image 

Classification 

Input:  

  Set of lung X-ray images 

Output: 

  WTBANEN-CNN: Wavelet ANOVA Enhanced CNN model 

Steps: 

1. Feature Extraction using Wavelet Transform: 

    For each X-ray image in data set: 

      a. Apply two level  Haar wavelet transform to the image to extract features across multiple 

scales. 

      b. Concatenate the wavelet features to form a comprehensive feature set. 

2. Feature Selection using Bonferroni-Adjusted ANOVA: 

      a. Perform Bonferroni-Adjusted ANOVA feature selection on the concatenated feature set to 

identify the most relevant    

          features and Reduce the dimensionality of the feature set. 

3. Convolutional Neural Network (CNN) Initialization and Transfer Learning: 

      a. Initialize CNN model architecture suitable for image classification. 

      b. Incorporate transfer learning by adding pre-trained layers from VGG19. 

4. Model Training: 

      a.Train the CNN model using the selected features and corresponding labels: 

      b.Split the dataset into training and validation sets. 

5. Evaluation: 

      a. Evaluate the trained model on the test data: 

      b. Compute evaluation metrics such as accuracy, precision, recall, and F1-score. 

5. Result Analysis 

In this study, lung x-ray images were utilized to extract a comprehensive set of features, which were 

subsequently processed to form a robust feature set. ANOVA feature selection algorithm was 

employed to perform feature selection, aiming to enhance the performance of the classification 

model by identifying the most significant features. The selected features were then used to classify 

the lung x-ray images. The following analysis discusses the results obtained from this process. 

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2s (2025) 

 

431 https://internationalpubls.com 

5.1 Results from Feature Extraction and Concatenation 

The application of wavelet transform on lung X-ray images effectively captures essential features 

across multiple scales. A total of 8,000 images are used for four-class classification, with 2,000 

images per class. For two-class classification, 5,000 images are used, with 2,500 images from each 

class being used for analysis. This multi-resolution analysis helps in extracting features that are both 

spatially and frequency localized, offering a rich set of information from the X-ray images. These 

features were concatenated into a single feature set, providing a rich representation of the image data 

with 32768 features extracted from a two-level Haar wavelet transform, as shown in Fig. 3. The 

high-dimensional feature set, while comprehensive, posed a challenge due to the potential presence 

of irrelevant or redundant features. 

 

Fig.3. Concatenated Features (32768) extracted using Haar Wavelet transform 

5.2. Analysis of Bonferroni-Adjusted ANOVA for Feature Selection 

Employing Bonferroni-Adjusted ANOVA for feature selection proves beneficial in identifying the 

most relevant 16,384 features from the concatenated set of 32,768 features shown in Fig.4. Also 

includes Bonferroni correction to adjust for multiple tests, ensuring a stricter control of the false 

positive rate. This step not only reduces dimensionality but also helps eliminate redundant and less 

informative features. The reduction in dimensionality leads to faster training times and less 

computational load on the subsequent CNN model. 

 

Fig.4. Features (16384) selected through Bonferroni-Adjusted ANOVA Feature Selection 

5.3. Results of Proposed WTAN-CNN with Transfer Learning 

The selected features are fed into a CNN for classification. The CNN, trained with these optimized 

features, shows significant improvement in accuracy and robustness. The reduction in input feature 

dimensionality ensures that the network focuses on the most informative aspects, enhancing learning 

efficiency. Leveraging transfer learning by using pre-trained model such as VGG19 further boosts 

the CNN's performance. The results, demonstrated in Figures 5 and 6, include an accuracy graph and 

a classification report. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2s (2025) 

 

432 https://internationalpubls.com 

 

Fig.5.Accuracy graph of   a) WTBAN-CNN (4 class)   b) WTBAN-CNN (2 Class) 

 

Fig.6.Accuracy graph of c) WTBAN-CNN-TL (4 Class)   d) WTBAN-CNN-TL (2 Class) 

Fig.5 and Fig.6 illustrates that the proposed WTBAN-CNN-TL with VGG19 achieves an accuracy of 

88% for two class classification and 86% for four-class classification. The WTBAN-CNN approach 

without transfer learning achieves an accuracy of 82 % for four-class classification and 86% for 

binary classification of Lung X-Ray data set. For the four-class classification using WTBAN-CNN, 

out of 8,000 images, 6,400 are used for training and 1,600 for testing. In the transfer learning 

approach, 10,000 images are utilized, with 8,000 for training and 2,000 for testing. For the two-class 

classification, from a total of 5,000 images, 4,000 are allocated for training and 1,000 for testing in 

both approaches, as shown in Fig.7. 

 

Fig.7.Classification Report of a) WTBAN-CNN (4 class) b) WTBAN-CNN (2 Class) 

c) Proposed WTBAN-CNN-TL (4 Class)   d) proposed WTBAN-CNN-TL (2 Class) 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2s (2025) 

 

433 https://internationalpubls.com 

5.4. Comparison with Genetic Algorithm-Based Feature Selection 

In the comparative analysis, ANOVA demonstrates a more straightforward and computationally 

efficient approach for feature selection. While the genetic algorithm (GA) offers a robust search 

mechanism, it is generally more resource-intensive and time-consuming compared to ANOVA. The 

experiments indicate that ANOVA selected features achieve comparable results in terms of 

classification accuracy and computational efficiency. Table II shows the comparison results. 

Table II Accuracy results of proposed WTAN-CNN-TL, WTAN-CNN and GA selected features 

 

6. Conclusion 

The proposed (WTBAN-CNN-TL) comprehensive approach for lung X-ray image classification 

effectively combines wavelet transform-based feature extraction, Bonferroni ANOVA feature 

selection, and deep learning techniques. By using wavelet transform, essential features are captured 

across multiple scales and concatenated to form a rich feature set. Bonferroni ANOVA is then 

employed to select the most relevant features with less false positive rate, reducing the feature set 

from 32,768 to 16,384 features. This dimensionality reduction not only streamlines the dataset but 

also enhances the efficiency of the subsequent CNN model, leading to faster training times and lower 

computational resource requirements. The comparative analysis with genetic algorithm-based feature 

selection highlights the efficiency and simplicity of Bonferroni adjusted ANOVA, which achieves 

comparable results with less computational overhead. Feeding the selected features into a CNN 

demonstrates significant improvements in classification accuracy, and the use of transfer learning 

with pre-trained models further boosts performance. 

References 

[1] Farhan AMQ, Yang S. Automatic lung disease classification from the chest X-ray images using hybrid deep 

learning algorithm. Multimed Tools Appl. 2023 Mar 22:1-27.  

[2] Goyal S, Singh R. Detection and classification of lung diseases for pneumonia and Covid-19 using machine and 

deep learning techniques. J Ambient Intell Humaniz Comput. 2023;14(4):3239-3259.  

Method  
Train Accuracy 

(%) 

Test Accuracy 

(%) 

Epochs 

WTBAN-CNN (2 Class) 88 86 20 

WTBAN-CNN (4 Class) 83 82 20 

Proposed WTBAN-CNN- TL (VGG19)(2 

Class) 
98 88 

     

       10 

Proposed WTBAN-CNN- TL (VGG19)(4 

Class) 
95 86 

 

10 

WT with Genetic algorithm based CNN(2 Class) 49 

 

50 

 

10 

WT with Genetic algorithm based CNN (4 Class) 81 

 

76 

 

10 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2s (2025) 

 

434 https://internationalpubls.com 

[3] El-Shafai, W., Ghandour, C. & El-Rabaie, S. Improving traditional method used for medical image fusion by deep 

learning approach-based convolution neural network. J Opt 52, 2253–2263 (2023). https://doi.org/10.1007/s12596-

023-01123-y. 

[4] Shreyesh Doppalapudi, Robin G. Qiu, Youakim Badr, Lung cancer survival period prediction and understanding: 

Deep learning approaches, International Journal of Medical Informatics, Volume 148, 2021, 104371, ISSN 1386-

5056. 

[5] Qiao, Yu‐Long & Zhao, Yue & Song, Chun‐Yan & Zhang, Kai‐Ge & Xiang, Xue‐Zhi. (2021). Graph wavelet 

transform for image texture classification. IET Image Processing. 15. 10.1049/ipr2.12220.  

[6] Wulandari M, Chai R, Basari B, Gunawan D. Hybrid Feature Extractor Using Discrete Wavelet Transform and 

Histogram of Oriented Gradient on Convolutional-Neural-Network-Based Palm Vein Recognition. Sensors (Basel). 

2024 Jan 6;24(2):341. doi: 10.3390/s24020341. PMID: 38257434; PMCID: PMC10820403. 

[7] Balamurugan, D. & Seshadri, s.Aravinth & Reddy, P. & Rupani, Ajay & Manikandan, A.. (2022). Multiview 

Objects Recognition Using Deep Learning-Based Wrap-CNN with Voting Scheme. Neural Processing Letters. 54. 

1-27. 10.1007/s11063-021-10679-4. 

[8] Reduwan, N.H., Abdul Aziz, A.A., Mohd Razi, R. et al. Application of deep learning and feature selection 

technique on external root resorption identification on CBCT images. BMC Oral Health 24, 252 (2024). 

https://doi.org/10.1186/s12903-024-03910-w. 

[9] Annamalai, Manikandan & Muthiah, Ponni. (2022). An Early Prediction of Tumor in Heart by Cardiac Masses 

Classification in Echocardiogram Images Using Robust Back Propagation Neural Network Classifier. Brazilian 

Archives of Biology and Technology. 65. 10.1590/1678-4324-2022210316. 

[10] Monday HN, Li J, Nneji GU, Nahar S, Hossin MA, Jackson J. COVID-19 Pneumonia Classification Based on 

NeuroWavelet Capsule Network. Healthcare. 2022; 10(3):422. 

[11] Manikandan, Annamalai, M,Ponni Bala. (2023). Intracardiac Mass Detection and Classification Using Double 

Convolutional Neural Network Classifier. Journal of Engineering Research. 11(2A). 272-280. 10. 36909/jer.12237. 

[12] Ali, R., Manikandan, A. & Xu, J. A Novel framework of Adaptive fuzzy-GLCM Segmentation and Fuzzy with 

Capsules Network (F-CapsNet) Classification. Neural Comput & Applic (2023). https://doi.org/10.1007/s00521-

023-08666-y 

[13] Ali, R., Manikandan, A., Lei, R. et al. A novel SpaSA based hyper-parameter optimized FCEDN with adaptive 

CNN classification for skin cancer detection. Sci Rep 14, 9336 (2024). https://doi.org/10.1038/s41598-024-57393-

4. 

[14] Li, Qiufu & Shen, Linlin & Guo, Sheng & Lai, Zhihui. (2020). Wavelet Integrated CNNs for Noise-Robust Image 

Classification. 7243-7252. 

[15] Sarhan, Ahmad. (2020). A Novel Lung Cancer Detection Method Using Wavelet Decomposition and Convolutional 

Neural Network. Journal of Biomedical Science and Engineering. 13. 81-92. 

[16] Rasheed, A., Younis, M.S., Qadir, J., & Bilal, M. (2021). Use of Transfer Learning and Wavelet Transform for 

Breast Cancer Detection. ArXiv, abs/2103.03602. 

[17] https://www.kaggle.com/datasets/tawsifurrahman/covid19-radiography-database 

[18] B. Cui and H. Jiang, "An Image Edge Detection Method Based on Haar Wavelet Transform," 2020 International 

Conference on Artificial Intelligence and Computer Engineering (ICAICE), Beijing, China, 2020, pp. 250-254, doi: 

10.1109/ICAICE51518.2020.00054. 

[19] Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint 

arXiv:1409.1556, 2014.   

 

 

 

https://doi.org/10.1007/s12596-023-01123-y
https://doi.org/10.1007/s12596-023-01123-y
https://doi.org/10.1186/s12903-024-03910-w
https://doi.org/10.1038/s41598-024-57393-4
https://doi.org/10.1038/s41598-024-57393-4
https://www.kaggle.com/datasets/tawsifurrahman/covid19-radiography-database