Lung cancer diagnosis with GAN supported deep learning models

Abstract

Background

Lung cancer is a leading cause of cancer-related deaths worldwide, making early diagnosis crucial for improving treatment success and survival rates. Traditional diagnostic methods, such as biopsy and manual CT image interpretation, are time-consuming and prone to variability, highlighting the need for more efficient and accurate tools. Advances in deep learning offer promising solutions by enabling faster and more objective medical image analysis.

Objective

This study aims to classify benign, malignant, and normal lung CT images using advanced deep learning techniques, including a specially developed CNN model, to improve diagnostic accuracy.

Methods

A dataset of 1097 lung CT images was balanced using GANs and preprocessed with techniques like histogram equalization and noise reduction. The data was split into 70% training and 30% testing sets. Models including VGG19, AlexNet, InceptionV3, ResNet50, and a custom-designed CNN were trained. Additionally, Faster R-CNN-based region proposal methods were integrated to enhance detection performance.

Results

The custom CNN model achieved the highest accuracy at 99%, surpassing other architectures like VGG19, which reached 97%. The Faster R-CNN integration further improved sensitivity and classification precision.

Conclusion

The results demonstrate the effectiveness of GAN-supported deep learning models for lung cancer classification, highlighting their potential clinical applications for early detection and diagnosis.

Keywords

GAN-supported lung cancer diagnosis classification with CNN deep learning-based diagnostic methods R-CNN data augmentation

Introductıon

Lung cancer is one of the leading causes of mortality worldwide.¹ Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) are recognized as standard medical procedures that extend patients’ lifespans.^2,3 Early diagnosis can significantly improve treatment success and enhance patient survival rates. Early detection of lung cancer is a crucial means of providing effective treatment options.⁴ Traditional medical imaging methods are manually evaluated by expert radiologists, a time-consuming process prone to subjective interpretation variances. Hence, the need for automation and increased accuracy in medical image analysis has led to the growing utilization of deep learning techniques in this field.

Deep learning has the potential to revolutionize medical image analysis due to its ability to automatically learn complex features from large datasets. Particularly, CNN (Convolutional Neural Network) and the ResNet50, Inceptionv3, VGG19, and AlexNet architectures used in this study have shown high success rates in recognizing and classifying important features in images. These architectures, along with the designed CNN model, have been optimized for classification tasks. Additionally, the use of Faster R-CNN for region proposal aims to reduce processing load and increase success rates. Furthermore, the imbalance in the dataset has been addressed by using GAN (Generative Adversarial Networks) to balance benign, malignant cancer tomographies, and normal tomographies in the dataset. The application of these techniques to lung tomography images could lead to significant advancements in the detection and classification of lung cancer.

This study was conducted using a retrospective approach on a population consisting of individuals diagnosed with and without lung cancer. The population used in the study comprises 60% individuals diagnosed with lung cancer and 40% healthy individuals. These proportions were taken into account to minimize the impact of potential imbalances in the dataset on model accuracy, and data augmentation was performed using GAN (Generative Adversarial Network) to ensure the balance of the dataset. This approach enabled the achievement of high accuracy rates with sufficient population representation and improved classification performance.

The aim of this study is to develop a machine learning model for automatically detecting and classifying normal, benign, and malignant lesions in lung tomography images. This model has been trained on a dataset consisting of lung tomography images that have been balanced using GAN. After undergoing medical image processing techniques, the augmented dataset was used for training and testing with deep learning architectures. The dataset has been classified into three categories: normal, benign, and malignant. The model training was conducted using a training set covering 70% of the images, while the remaining 30% was utilized as a test set to evaluate the model's performance.

This study addresses the potential contributions of using a machine learning-based approach in the diagnosis of lung cancer. Some of the various contributions provided by this approach include:

Object Detection and Classification Capability: The deep learning models utilized in this study can effectively be employed for both object detection and classification tasks. This is beneficial for identifying the presence of lesions in lung tomography images as well as characterizing their types. Particularly, distinguishing between benign cancer images and normal images can be challenging. However, the designed machine learning model overcomes this hurdle with a high success rate.

Early Diagnosis Capability: While traditional medical imaging methods typically focus on detecting the presence of specific lesions, deep learning-based systems have the potential to detect early-stage cancer lesions. This is of significant importance in preventing disease progression and initiating treatment earlier.

Fast and Automated Assessment: Machine learning-based systems can automatically analyze images and produce results rapidly. This can reduce the workload of radiologists and expedite the diagnostic process, enabling the initiation of treatment for patients more quickly.

High Sensitivity and Specificity: Deep learning models can provide high sensitivity and specificity, meaning they have high capabilities in detecting cancer lesions and accurately classifying normal tissues. This contributes to reducing false positive or false negative results and facilitates more accurate diagnoses.

Scalability: Machine learning-based systems can be trained on large datasets and utilized in various clinical settings. This enhances the model's generalization ability and facilitates its usage across different hospitals or clinics.

Guiding Treatment: Early diagnosis plays a critical role in determining treatment options. Accurate and rapid diagnosis by machine learning-based systems enables better formulation of treatment plans and access to personalized treatment regimens for patients.

These contributions highlight the potential of machine learning-based lung cancer diagnosis systems in clinical applications and their potential to enhance patients’ quality of life.

Literature review

There are many studies in the literature on the classification and detection of lung cancer. These studies encompass various approaches, including different imaging techniques and deep learning methods.

Asuntha and colleagues investigated the role of lung cancer in fatal cases worldwide in their study. They focused on the diagnosis of lung diseases using CT (Computed Tomography) scans. By employing feature extraction techniques such as HoG (Histogram of Oriented Gradients) and wavelet transform-based features, they detected cancerous lung nodules. The results indicated that the new FPSOCNN method was more successful compared to other techniques.⁵

Yiwen Xu and colleagues investigated the prediction of clinical outcomes using deep learning based on CT images of stage III NSCLC (non-small cell lung cancer) patients in their study. The model predicted pathological responses by categorizing patients into low and high mortality risk groups. The results indicate that deep learning could improve predictions of clinical outcomes and that machine learning-based radiomic biomarkers could have a significant impact.⁶

Marjolein A. Heuvelmans and colleagues aimed to validate the LCP-CNN model trained in the United States in Europe in their study. The model was used to determine the malignancy scores of lung nodules, and in retrospective analyses conducted in Europe, an AUC value of 94.5% and sensitivity of 99.0% were achieved. These results obtained in their study could assist in avoiding follow-up scans for patients.⁷

Lakshmanaprabu and colleagues stated that early diagnosis of lung cancer could increase the chances of survival in their study. The study presents an automatic diagnosis classification method for CT images. Using features extracted from CT images with deep neural networks and linear discriminant analysis, the aim was to classify lung nodules as malignant or benign. The proposed classifier achieved 96.2% sensitivity, 94.2% specificity, and 94.56% accuracy.⁸

She Y. and colleagues developed and validated a deep learning model to predict the survival of lung cancer patients in their study. This model demonstrated more promising results compared to tumor, node, and metastasis staging systems, and it revealed superior survival rates for patients receiving the recommended treatments. The results of the study were visualized with a user-friendly graphical interface, showing that the deep learning model could be used to provide survival predictions and treatment recommendations to healthcare professionals.⁹

Chaunzwa and colleagues proposed a radiomic approach using deep learning methods to predict lung cancer tumor histology in their study. CNNs trained with data obtained from Massachusetts General Hospital achieved a 71% AUC success rate on two common histological types. It was also determined that the AUC obtained with machine learning classifiers reached up to 0.71. The best-performing CNN served as a reliable classifier and provided qualitatively interpretable predictions.¹⁰

Coudray and colleagues evaluated deep learning methods for the classification of lung tumor histopathology images and mutation prediction in their study. A deep neural network trained with the InceptionV3 model successfully classified lung cancer subtypes such as LUAD and LUSC and could predict the ten most common gene mutations in LUAD. These findings from the study demonstrate that deep learning models could assist pathologists in cancer diagnosis.¹¹

Teramoto and colleagues utilized CNN for the automatic classification of lung cancer types in microscopic images in their study. The CNN architecture used consists of three convolutional layers, three pooling layers, and two fully connected layers. The developed scheme was trained by augmenting collected images through rotation, flipping, and filtering. Classification accuracy was determined to be 71% with three-fold cross-validation. These results indicate that the scheme achieves similar accuracy to cytotechnologists and pathologists.¹²

Cengil and Cinar utilized deep learning methods for the classification of lung nodules in their study. This method has become increasingly popular in recent years for the classification process, particularly being widely implemented with libraries such as TensorFlow and architectures like 3D convolutional neural networks. The study was conducted on CT images from the SPIE-AAPM-LungX dataset, yielding successful results. Such studies demonstrate the effectiveness of deep learning in biomedical image classification.¹³

Anirudh, R. and colleagues investigated the detection of lung nodules using a 3D CNN trained with weakly labeled data in their study. The importance of early detection of lung nodules was emphasized, and the effectiveness of 3D CNN was discussed. It was noted that the network trained with unsupervised segmentation method provided high sensitivity with low false positive rates.¹⁴

Song and colleagues examined the effectiveness of deep learning methods for the classification of lung nodules in their study. Deep neural networks such as CNN, DNN, and SAE were used and evaluated on the LIDC-IDRI database. The study demonstrated that CNN performed best in classifying lung nodules.¹⁵

Salaken and colleagues addressed the classification of lung cancer on a small-sample high-dimensional dataset in their study. In contrast to traditional methods, automatic feature extraction and classification were performed using deep learning. It was demonstrated that the deep learning-based classifier outperformed other methods, and this improvement was statistically significant.¹⁶

The importance of deep learning methods in the classification of lung cancer is increasingly recognized in the field. Many studies in the literature demonstrate the effectiveness of such approaches. However, our study stands out in terms of novelty by integrating various image processing techniques on a balanced dataset using GAN. The support of these approaches has enabled us to achieve high success rates in network architectures, leading to precise and reliable results. This study not only highlights the potential of deep learning techniques in lung cancer diagnosis but also underscores the importance of advanced artificial intelligence techniques like GANs in clinical applications.

Methods and materıals

Figure 1.

Study Flowchart. A detailed flowchart illustrating the steps involved in the lung cancer diagnosis study, from data preprocessing to model evaluation.

Data pre-processing techniques

Before the training and testing processes, the dataset underwent data dimension resizing and medical image processing techniques. In Figure 1, the main flow diagram of the study is presented in detail. The CT images were resized to 224 × 224 dimensions to fit the machine learning architectures and were balanced by augmentation using GAN. Subsequently, various medical image processing techniques were applied to the CT images to enhance the success rates. These processes were conducted to improve the accuracy and reliability of the results.

GAN (generative adversarial network)

GANs are significant models in the field of deep learning. They consist of two networks - a generator and a discriminator - competing in a binary deep network model. During training, the generator network attempts to produce more realistic data by trying to deceive the discriminator network, while the discriminator network learns to distinguish between real and fake examples. As these two networks work in opposition, at the end of the process, the generator network learns to produce better sample data to capture the real data distribution. When training is completed after a certain number of iterations, the generator network is trained to generate real data.^17,18 These steps are shown sequentially in Figure 2.

Figure 2.

GAN Workflow Diagram. A schematic representation of the Generative Adversarial Network (GAN) workflow used for data augmentation and balancing the dataset.

Medical image processing techniques

Histogram Equalization

Histogram equalization is a method used in image processing to enhance the contrast of an image. The histogram of an image shows the frequency of each brightness value in the image. Histogram equalization adjusts the distribution of these brightness values to be equal across the entire brightness range of the image. This process particularly improves the visibility of details in low-contrast images. The method assigns new brightness values to each pixel using the cumulative distribution function of the image. As a result of this process, details in low-contrast regions become more pronounced, and the image generally achieves a more balanced contrast.^19,20 In this study, Otsu's method was used, where the most suitable threshold value is determined by analysing the histogram of the image.²¹ The formulas below illustrate the histogram equalization formulas. Formula 1 represents histogram calculation, formula 2 represents cumulative distribution function, formula 3 represents equalized values, and formula 4 represents the creation of the new image. Figure 3 depicts the histogram equalized CT images and graphs.

p_{r} (r) = \frac{n_{r}}{mn}

(1)

s_{k} = \sum_{j = 0}^{k} p_{r} (r_{j})

(2)

s_{k} = round ((L - 1) . s_{k})

(3)

I^{'} (x, y) = s_{I (x, y)}^{^{'}}

(4)

Figure 3.

Histogram Equalization. Comparison of original and processed CT images after applying histogram equalization to enhance contrast and visibility of lung regions.

Noise Reduction

Noise reduction is a set of techniques used to reduce unwanted distortions in images. Sensor errors, lighting conditions, errors occurring during transmission, etc., can create noise in the image. In this study, the Wiener filter was used for noise reduction. The Wiener filter is a technique used in image processing and signal processing to reduce noise, based on statistical methods. Developed by Norbert Wiener, this method uses the statistical properties of noise and signal to make the best estimate of the original signal or image. The Wiener filter is designed especially for linear, time-invariant systems and provides optimization for both reducing random noise and signal processing.²² As shown in Figure 4, noise reduction has been applied to the image.

Figure 4.

Noise Reduction. Visual demonstration of noise reduction applied to a lung CT image using the Wiener filter for improved clarity.

Edge Detection

Edge detection is an important technique used to determine the boundaries of objects in an image. This method works by detecting sudden changes in brightness values in the image. There are various edge detection filters such as Sobel, Prewitt, Roberts, and Canny. These filters detect edges by taking derivatives in the horizontal and vertical directions on the image. The Canny edge detector used in this study is often preferred due to its high performance; it reduces the influence of noise while detecting edges and determines edges sharply and precisely.²³

Formula 5 shows Gaussian filtering, formula 6 indicates gradient magnitude, and formula 7 represents gradient direction. Here, Gx and Gy are the derivatives of the gradients in the x and y directions, respectively. As shown in Figure 5, edge detection has been applied to the image.

G (x, y) = \frac{1}{2 π σ^{2}} e^{\frac{x^{2} + y^{2}}{2 σ^{2}}}

(5)

G = \sqrt{G_{x}^{2} + G_{y}^{2}}

(6)

θ = \arctan (\frac{G_{y}}{G_{x}})

(7)

Figure 5.

Edge Detection. Application of the Canny edge detection algorithm on lung CT images to identify boundaries of cancerous lesions.

Erosion

Erosion is one of the morphological operations and is typically applied to binary images. Erosion narrows the boundaries of an image and eliminates small objects. This operation is performed using a structuring element that is traversed over the image. When the structuring element does not exactly match an object in the image, it results in the erosion of that object. Erosion is used to remove unwanted small pieces from images or to open up gaps between objects.²⁴ As shown in Figure 6, erosion has been applied to the image.

Figure 6.

Erosion. Illustration of morphological erosion on binary lung CT images, demonstrating removal of small artifacts.

Segmentation

Segmentation is the process of dividing an image into meaningful components or objects. This process is used to determine the location, number, size, and shape of objects in the image. Among segmentation techniques are thresholding, region growing, watershed, and k-means clustering. Thresholding separates the image into objects and background by selecting pixels above or below a certain brightness value.^25,26 In this study, segmentation was performed using the k-means clustering algorithm.²⁷ [In Formula 8, K cluster centers are determined. In Formula 9, the Euclidean distance between each data point x and the cluster center is calculated. In Formula 10, the center of each cluster is updated as the average of the data points within the cluster. As shown in Figure 7, segmentation has been applied to the image.

C_{i} i = 1, 2, \dots, K

(8)

d (x, C_{i}) = \sqrt{\sum_{j = 1}^{n} {(x_{j} - C_{ij})}^{2}}

(9)

C_{i} = \frac{1}{| S_{i} |} \sum_{x \in S_{i}} x

(10)

Figure 7.

Segmentation. Results of k-means clustering-based segmentation on CT images, highlighting cancerous regions.

Faster r-CNN

Faster R-CNN (Region-based Convolutional Neural Networks) is a popular deep learning model especially used in object detection tasks. Faster R-CNN is developed from previous versions, namely R-CNN and Fast R-CNN, and it provides significant improvements in both accuracy and speed. In Faster R-CNN, the input image first passes through convolutional and other layers, resulting in the creation of a feature map of the image. The feature map contains the characteristics of objects within the image.²⁸ The feature map obtained from the convolutional layers then enters a network called RPN (Region Proposal Network). RPN scans the feature map for thousands of possible object locations with boxes called anchors. Anchors are predetermined boxes with fixed sizes and aspect ratios used in some object detection algorithms like Faster R-CNN. These anchors used by RPN are employed to detect objects of various shapes and sizes. Essentially, these boxes, placed at fixed intervals on an image and with different scales and aspect ratios, assist the model in identifying potential object candidates within the image. It computes the probability of each anchor being an object and performs bounding box regression. After this stage, the positions of potential object candidates within the image are determined.²⁹ Figure 8 illustrates the bounding boxes determined by the anchors within a tomography image with extracted features.

Figure 8.

RPN Region Proposals. Bounding box predictions generated by the Region Proposal Network (RPN) as part of the Faster R-CNN model applied to lung CT images.

ResNet50

ResNet50 is a deep neural network model based on the concept of Residual Networks. It was introduced by Kaiming He and his colleagues in 2015. ResNet50 is a convolutional neural network with 50 layers, utilizing ReLU as the activation function. It incorporates skip connections and pooling layers. ResNet50 is trained on the ImageNet dataset, which consists of 1000 different classes. It is commonly used in various image processing applications, including object recognition and other deep learning-based tasks.³⁰ As shown in Figure 9, the ResNet50 architecture is depicted.

Figure 9.

ResNet50 Architecture.³⁰ Diagram of the ResNet50 architecture utilized for feature extraction and classification of lung CT images.

VGG19

VGG19, an convolutional neural network used in image recognition and classification in the field of deep learning. It consists of 16 convolutional layers and 3 fully connected layers, utilizing ReLU as the activation function. VGG19 is commonly employed in transfer learning with pre-trained weights trained on large datasets like ImageNet. It is utilized in visual recognition and classification tasks that require high computational resources.³¹ As shown in Figure 10, the VGG19 architecture is depicted.

Figure 10.

VGG19 Architecture. Depiction of the VGG19 architecture, showcasing its convolutional and fully connected layers.

InceptionV3 (GoogleNet)

The InceptionV3 architecture developed by Google consists of a total of 48 layers, including convolutional, pooling, normalization, and fully connected layers. InceptionV3 is a model trained with ImageNet data with over 1000 different classes. It is commonly used for transfer learning. The Inception modules contained within it are important structural elements commonly used in CNNs and are synonymous with the Inception architecture. These modules are formed by combining filters of different sizes (1 × 1, 3 × 3, 5 × 5) and pooling layers. Inception modules are designed to make deep learning models more effective and efficient.³² As shown in Figure 11, the Inception V3 architecture is depicted.

Figure 11.

InceptionV3 Architecture. Illustration of the InceptionV3 architecture, emphasizing its unique inception modules with multi-scale convolutional filters.

Alexnet

AlexNet is a landmark convolutional neural network model in deep learning and computer vision research. This neural network model comprises 60 million parameters and 500,000 neurons. The model includes five convolutional layers, some of which are followed sequentially by maximum pooling layers. At the end of the model, there are two globally fully connected layers with a softmax function outputting 1000 classes.³³ As shown in Figure 12, the AlexNet architecture is depicted.

Figure 12.

Alexnet Architecture.³³ Structure of AlexNet, highlighting its convolutional and pooling layers for image classification tasks.

The designed CNN

The CNN designed for this study comprises a total of 15 layers. Softmax is used as the classification function in the final layer. ReLU is utilized as the activation function after each convolutional layer. To enhance the speed and stability of the network, batch normalization layers are applied after each convolutional layer. Following the convolutional layers, three fully connected layers are employed. The architecture of the network is illustrated in Figure 13.

Figure 13.

Designed CNN Architecture. Custom-designed 15-layer CNN architecture used in this study for lung cancer classification.

The developed CNN model has been optimized to achieve high accuracy in lung tomography images. This optimization process encompasses not only the model's complex architecture but also the effective utilization of learning algorithms. Compared to commonly used deep learning models such as VGG19, AlexNet, and ResNet50, the model significantly increases diagnostic accuracy by offering lower error rates. In particular, the model's unique layer structure enables it to capture detailed features of lung cancer lesions. As a result, lesions can be classified with high accuracy, allowing for early detection of the disease, which presents a considerable advantage for clinical applications.

Additionally, inter-class imbalances in the dataset have been addressed through GAN-supported data augmentation. This approach enables the model to identify different classes more reliably and enhances overall classification performance. During GAN training, critical parameters such as learning rate and number of epochs were optimized to ensure that synthetic images were produced realistically and reflected the target classes. These optimizations play a crucial role in enabling the model to generalize better and achieve more compatible results with real-world data. Consequently, the developed model offers an innovative solution for lung cancer diagnosis, making significant contributions to the field of healthcare.

Training and testing procedures

The dataset has been split into 70% for training and 30% for testing. The CNN network architecture has been trained as shown in Figure 13. Additionally, testing procedures have been conducted on the CNN model trained with pre-trained network architectures such as VGG19, AlexNet, InceptionV3, and ResNet50, which were previously trained on an average of 1000 different categories. The parameters of the used network architectures are shown in Table 1.

Table 1.

Network parameters.

Batch Size	32
Optimizer	SGD
Learning Rate	0.001
Momentum	0.9
Loss Function	Cross Entropy Loss
Epochs	10
Image Size	224 × 224
Resolution	4096
Number of Classes	3
Training Data Ratio	0.7
Test Data Ratio	0.7
Number of Training Samples	1178
Number of Test Samples	505

Performance metrics

Performance evaluation is a critical step to understand how well the trained model performs. In this study, metrics such as accuracy, precision, recall, and F1 score have been used to evaluate the performance of the model. Additionally, examining the macro and weighted average values allows for the investigation of performance imbalances among classes.

The materials and methods section, by employing performance metrics, provides a deeper understanding of the techniques and methodology used, thereby enhancing the reliability and reproducibility of the research.

Precision

Precision indicates how many of the samples predicted as positive by the model are truly positive. Precision aims to reduce the number of false positives.

Precision = \frac{TP}{TP + FP}

Recall

Recall shows how accurately the true positives are detected. Recall aims to reduce the number of false negatives.

Recall = \frac{TP}{TP + FN}

F-Score

F-score is the harmonic mean of precision and recall. F-score balances both precision and recall values.

F - Score = \frac{Precision x Recall}{Precision / Recall}

Results

In this study, lung CT images were classified into three categories (benign cancer, malignant cancer, and normal) as shown in Figure 1. Initially, GANs were used to balance the dataset. GANs are deep neural networks used to generate synthetic data that closely resembles real data. After balancing the dataset, biomedical image processing techniques such as histogram equalization, segmentation, and edge detection were applied to make the cancerous regions in the tomography images clearer.

Subsequently, a 15-layer CNN architecture was employed, utilizing various models, including ResNet50, AlexNet, InceptionV3 (GoogleNet), and VGG19, trained across multiple categories. The dataset was divided into two parts: 70% for training and 30% for testing, and training and testing procedures were carried out. The performance metrics of the models in cancer classification are presented in Table 2, and the model success rates are presented in Table 3. Among the pre-trained network architectures, VGG19 achieved an accuracy of 97%, while the accuracy of the designed CNN architecture reached 99%. Confusion matrices for both VGG19 and the designed CNN architecture are shown in Figures 14 and 15. It can be observed that the designed CNN architecture achieved high success rates, particularly in distinguishing between benign cancer and normal images. Macro and weighted average performance metrics are shown in Tables 4 and 5. The performance metrics of all the network models used are shown in Figure 16.

Figure 14.

Confusion Matrix for VGG19. Confusion matrix showing the classification performance of the VGG19 architecture on the test dataset.

Figure 15.

Confusion Matrix for Designed-CNN. Confusion matrix demonstrating the classification performance of the custom-designed CNN on the test dataset.

Figure 16.

Success Graphs. Graphs comparing the performance metrics (accuracy, precision, recall, and F1-score) of various architectures used in the study.

Table 2.

Class-Wise precision, recall, and F1-score.

Model	Class	Precision	Recall	F1-Score
ResNet50	Normal	0.8637	0,9477	0,8898
	Bening	0,7266	0,7167	0,8026
	Malignant	0,9693	0,9875	0,9798
AlexNet	Normal	0,9654	0,8550	0,9216
	Bening	0,7307	0,9790	0,8444
	Malignant	0,9720	0,9806	0,9915
InceptionV3	Normal	0,9420	0,9848	0,9629
	Bening	0,8823	0,8094	0,8333
	Malignant	0,9990	0,9875	0,9937
VGG19	Normal	0,9672	0,9752	0,9711
	Bening	0,9210	0,9210	0,9210
	Malignant	0,9990	0,9941	0,9970
Designed-CNN	Normal	0,9920	0,9920	0,9920
	Bening	0,9722	0,9722	0,9722
	Malignant	0,9990	0,9990	0,9990

Table 3.

Model accuracy.

Model	Accuracy
ResNet50	0,8818
AlexNet	0,9484
InceptionV3	0,9636
VGG19	0,9787
Designed-CNN	0,9939

Table 4.

Macro avg precision, recall, and F1-score.

Model	Macro avg precision	Macro avg recall	Macro avgF1-score
ResNet50	0,8132	0,7704	0,7849
AlexNet	0,9046	0,9515	0,9192
InceptionV3	0,9414	0,9206	0,9300
VGG19	0,9627	0,9634	0,9631
Designed-CNN	0,9880	0,9880	0,9880

Table 5.

Weighted avg Precision, Recall, and F1-Score.

Model	Weighted avg precision	Weighted avg recall	Weighted avgF1-score
ResNet50	0,8690	0,8818	0,8566
AlexNet	0,9600	0,9484	0,9498
InceptionV3	0,9632	0,9636	0,9629
VGG19	0,9778	0,9787	0,9788
Designed-CNN	0,9939	0,9939	0,9939

The R-CNN integration was applied only to the proposed CNN model. This integration was designed to enable more accurate detection of lung cancer lesions, aiming to enhance classification success. The results obtained indicate that our model integrated with R-CNN achieved higher accuracy rates compared to popular models like VGG19.

The use of Faster R-CNN demonstrated an increase in the model's sensitivity, proving it to be an effective candidate in terms of diagnostic accuracy for clinical applications. This integration provides greater accuracy in identifying cancerous regions due to the advantages offered by the region proposal network. The features of R-CNN allow for more effective recognition of lung cancer lesions of various sizes and shapes, increasing the likelihood of early diagnosis.

Furthermore, this method exhibits superior performance in detecting early-stage lung cancer lesions compared to other models. This study, which significantly contributes to clinical applications, can aid in the earlier and more accurate treatment of patients. In conclusion, the integration of R-CNN enhances the overall performance of the proposed CNN model, offering innovative solutions in the healthcare field. This underscores the importance of timely interventions in lung cancer treatment and establishes a solid foundation for future research.

Conclusion

Lung cancer is one of the most common types of cancer worldwide and is a leading cause of cancer-related deaths. Millions of new cases of lung cancer are diagnosed each year. Smoking and genetic factors are significant risk factors for lung cancer.

The aim of this study is to overcome the challenges in distinguishing between benign cancer tomographies and normal tomographies. The dataset balanced with GANs, and cancerous regions became clearer with image processing techniques. As a result, the success rates of the network architectures reached high levels.

The results of this study demonstrate the effectiveness of deep learning models for classifying lung CT images. Balancing techniques such as GANs were utilized for the accurate classification of benign cancer, malignant cancer, and normal images, and biomedical image processing techniques were employed to enhance the clarity of cancerous regions. The obtained results were evaluated by comparing the designed CNN architecture with pretrained architectures (ResNet50, AlexNet, InceptionV3, and VGG19).

The examination of success rates indicates that the designed CNN architecture achieves higher accuracy compared to other architectures, with a 99% success rate. Particularly, compared to VGG19's 97% accuracy, the designed CNN architecture achieved 99% accuracy. These results demonstrate that the designed CNN architecture achieves high success, especially in distinguishing between benign cancer and normal images in lung CT scans.

One of the significant contributions of this study is to highlight the effectiveness and potential of deep learning models in medical image analysis. The high accuracy rates obtained indicate the potential usability of these models in clinical applications, enhancing the quality of life for patients in diagnosis and treatment processes.

Another aspect that could contribute to future studies is the use of larger and more diverse datasets and the analysis of data obtained from different clinical settings. This could improve the model's generalization ability and better adapt it to real-world applications.

Furthermore, the high success achieved by the designed CNN architecture in distinguishing between benign cancer and normal images encourages further exploration and development of this algorithm in future studies. This could facilitate the development of new and effective solutions in lung cancer diagnosis and classification.

Footnotes

Acknowledgments

The authors have no acknowledgments.

ORCID iD

Talip ÇAY

Funding

The author received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The author declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

References

Sahu

Dasari

, et al. A lightweight multi-section CNN for lung nodule classification and malignancy estimation. IEEE J Biomed Health Informatics 2019; 23: 960–968.

Ciompi

, et al. Towards automatic pulmonary nodule management in lung cancer screening with deep learning. Sci Rep 2017; 7: 46479.

Kuan

, et al. Deep Learning for Lung Cancer Detection: Tackling the Kaggle Data Science Bowl 2017 Challenge. 2017 May. Available from: https://arxiv.org/abs/1705.09435v1.

da Silva

GLF

Filho

AODC

Silva

, et al. Taxonomic indexes for differentiating malignancy of lung nodules on CT images. Res Biomed Eng 2016; 32: 263–272.

Alyasriy

. The IQ-OTHNCCD lung cancer dataset. 2020;1. doi: 10.17632/BHMDR45BH2.1.

Asuntha

Srinivasan

. Deep learning for lung cancer detection and classification. Multimed Tools Appl 2020 Mar; 79: 7731–7762.

, et al. Deep learning predicts lung cancer treatment response from serial medical imaging. Clin Cancer Res 2019 Jun; 25: 3266–3275.

Heuvelmans

van Ooijen

Ather

, et al. Lung cancer prediction by deep learning to identify benign lung nodules. Lung Cancer 2021; 154: –14.

Lakshmanaprabu

Mohanty

Shankar

, et al. Optimal deep learning model for classification of lung cancer on CT images. Future Gener Comput Syst 2019; 92: 374–382.

10.

She

Jin

, et al. Development and validation of a deep learning model for non–small cell lung cancer survival. JAMA Netw Open 2020; 3: e205842.

11.

Chaunzwa

, et al. Deep learning classification of lung cancer histology using CT images. Sci Rep 2021; 11: 1–12.

12.

Coudray

, et al. Classification and mutation prediction from non–small cell lung cancer histopathology images using deep learning. Nat Med 2018; 24: 1559–1567.

13.

Teramoto

Tsukamoto

Kiriyama

, et al. Automated classification of lung cancer types from cytological images using deep convolutional neural networks. Biomed Res Int 2017; 2017: 4067832.

14.

Cengil

Çinar

. A Deep Learning Based Approach to Lung Cancer Identification. 2018 Int Conf Artif Intell Data Process (IDAP 2018) 2018. doi: https://doi.org/10.1109/idap.2018.8620723

15.

Anirudh

Thiagarajan

Bremer

, et al. Lung nodule detection using 3D convolutional neural networks trained on weakly labeled data. 2016; 9785: 791–796. doi: https://doi.org/10.1117/12.2214876.

16.

Song

Zhao

Luo

, et al. Using deep learning for classification of lung nodules on computed tomography images. J Healthc Eng 2017; 2017: 8314740.

17.

Salaken

Khosravi

Khatami

, et al. Lung cancer classification using deep learned features on low population dataset. Can Conf Electr Comput Eng 2017. doi: https://doi.org/10.1109/ccece.2017.7946700.

18.

Goodfellow

, et al. Generative adversarial nets. Adv Neural Inf Process Syst. 2014. Available from: https://proceedings.neurips.cc/paper/5423-generative-adversarial-nets.

19.

Radford

Metz

Chintala

. Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks. 4th Int Conf Learn Represent (ICLR 2016). Available from: https://arxiv.org/abs/1511.06434v2.

20.

Dhal

Das

Ray

, et al. Histogram equalization variants as optimization problems: a review. Arch Comput Methods Eng 2021 May; 28: 1471–1496.

21.

Somal

. Image enhancement using local and global histogram equalization technique and their comparison. Adv Intell Syst Comput 2020; 1045: 739–753.

22.

Otsu

. A threshold selection method from gray-level histograms. Automatica. 1975; Available from: dspace.tul.cz.

23.

Nuha

Absa

. Noise reduction and speech enhancement using Wiener filter. 2022 Int Conf Data Sci Its Appl (ICoDSA 2022) 2022: 177–180. doi: https://doi.org/10.1109/icodsa55874.2022.9862912.

24.

Sekehravani

Babulak

Masoodi

. Implementing Canny edge detection algorithm for noisy image. Bull Electr Eng Informatics 2020 Aug; 9: 1404–1410.

25.

Karimi

Salcudean

. Reducing the Hausdorff distance in medical image segmentation with convolutional neural networks. IEEE Trans Med Imaging 2020 Feb; 39: 499–513.

26.

Aydin

. Medical image segmentation using deep neural networks. SIU 2021 - 29th IEEE Conf Signal Process Commun Appl Proc 2021. doi: https://doi.org/10.1109/siu53274.2021.9477838.

27.

Ölmez

Akdoğan

Korkmaz

, et al. Automatic segmentation of meniscus in multispectral MRI using regions with convolutional neural network (R-CNN). J Digit Imaging 2020 Aug; 33: 916–929.

28.

Lloyd

. Least squares quantization in PCM. IEEE Trans Inf Theory 1982; 28: 129–137.

29.

Ren

Girshick

, et al. Faster R-CNN: towards real-time object detection with region proposal networks. Adv Neural Inf Process Syst. 2015. Available from: proceedings.neurips.cc.

30.

Zhong

Sun

Huo

. An anchor-free region proposal network for Faster R-CNN-based text detection approaches. Int J Doc Anal Recognit 2019 Sep; 22: 315–327.

31.

Zhang

Ren

, et al. Deep residual learning for image recognition. In: Proceedings; 2016. pp. 770-778. Available from: http://image-net.org/challenges/lsvrc/2015/.

32.

Simonyan

Zisserman

. Very deep convolutional networks for large-scale image recognition. 3rd Int Conf Learn Represent (ICLR 2015). Available from: https://arxiv.org/abs/1409.1556v6.

33.

Szegedy

Vanhoucke

Ioffe

, et al. Rethinking the Inception architecture for computer vision. 2016. pp. 2818-2826.