WISeRKNet: wide slice residual Kronecker network for lung cancer detection based on CT images.
1/5 보강
Lung cancer poses a serious health risk, making early diagnosis essential for better survival outcomes.
APA
Shanthi A, Satheesh Kumar S, Koppu S (2026). WISeRKNet: wide slice residual Kronecker network for lung cancer detection based on CT images.. Scientific reports, 16(1). https://doi.org/10.1038/s41598-026-39793-w
MLA
Shanthi A, et al.. "WISeRKNet: wide slice residual Kronecker network for lung cancer detection based on CT images.." Scientific reports, vol. 16, no. 1, 2026.
PMID
41714764 ↗
Abstract 한글 요약
Lung cancer poses a serious health risk, making early diagnosis essential for better survival outcomes. Detection of lung cancer involves a series of medical evaluations and imaging techniques to identify cancerous cells in the lungs. Computed Tomography (CT) images are most frequently used to recognize lung cancer since it has high resolution, enhanced clarity, and minimal noise and distortions. However, accurate detection of lung cancer is complex owing to variations in nodule size, shape, and boundary definition. Therefore, an innovative model named Wide Slice Residual Kronecker Network (WISeRKNet) has been developed to diagnose lung cancer from CT images. Initially, image pre-processing is applied by using homomorphic filtering. Subsequently, the extraction of nodules in the lung is performed by the Link-Net model. Subsequently, augmentation of the image is conducted, and then the process of feature extraction is applied to refine shape-based features. At last, diagnosing lung cancer is executed by the WISeRKNet and which combines the Wide Slice Residual Network (WISeR) and the Deep Kronecker Network (DKN). Moreover, the developed WISeRKNet model demonstrated superior performance, by achieving improved value in accuracy as 91.686%, True Positive Rate (TPR) as 90.485%, True Negative Rate (TNR) as 92.727%, Precision as 90.980% and F1 score as 90.484% on the Lung Cancer Computed Tomography Images database using 90% of the data for training.
🏷️ 키워드 / MeSH 📖 같은 키워드 OA만
📖 전문 본문 읽기 PMC JATS · ~126 KB · 영문
Introduction
Introduction
The imaging tool of biomedical, like CT imaging, is used for accurate determination and localization of lung cancer. CT is also known as a digital method for cross-sectional imaging that allows for independent manipulation of image acquisition. The high exposure of CT improves image quality as the noise level decreases. Radiologists often prefer low-noise, high-quality CT images that are visually appealing. As a result, they may hesitate to reduce the radiation dose, which typically involves lowering the X-ray exposure settings during CT acquisition and can lead to increased image noise1. CT images are produced by utilizing X-rays to generate detailed cross-sectional views of the body2. CT scans are employed to view internal organs, bones, blood vessels, and various other structures within the body. CT imaging3can reveal both suspected as well as unsuspected nodules in lung cancer. However, variations in CT scans and potential misinterpretations of anatomical structures can complicate the identification of cancerous cells for radiologists4. Cancer is a more dangerous disease which is spreading increasingly worldwide. Moreover, lung cancer is the leading cause of cancer-related mortality worldwide5. Despite advancements in treatment and early detection methods, survival rates remain low, highlighting the urgent need for improved diagnostic tools and therapeutic strategies6.
The cells’ uncontrolled growth can spread over the lungs through metastasis, affecting nearby tissues or other areas of the body7. In the first two stages, cancer is formed in the lungs, then it spreads to other parts of the organs8. Segmenting anatomical structures of lungs is a crucial task for Computer-Assisted Diagnosis (CAD)9systems using CT scans of the chest. There are five lobes in the lungs: two in the left lung and three in the right lung. These are separated by lobar fissures. Segmentation of the lobe is simple since fissures are distinctly outlined in a CT scan. However, segmentation is often not the case due to incomplete fissures, the presence of additional structures, and surrounding lung parenchymal abnormalities1,7. Furthermore, prolonged smoking of tobacco is a key reason for 85% of lung cancer cases. Approximately 10–15% cases occur in individuals who have never smoked, primarily due to factors such as secondhand smoke, asbestos, air pollution and radon gas exposure10. The majority of diagnoses occur at advanced stages of disease due to the absence of early signs of the condition. Yet, detection in an earlier stage can significantly improve survival ability11.
The Machine Learning (ML) technique is a commonly applied technique for the classification of images. In ML-based approach, features are manually mined from images for identification purposes. The Support Vector Machine (SVM) technique is an ML technique primarily used for classification tasks12,13. In recent years, Deep Learning (DL) has become highly effective for lung cancer screening because it can automatically learn useful features and help detect the disease in its early stages14. DL algorithms can integrate imaging modalities, such as Magnetic Resonance Imaging (MRI)15and CT to enhance disease detection and support diagnosis as well as treatment planning7,16,17. Convolutional Neural Networks (CNNs), such as GoogleNet, AlexNet, ResNet, DenseNet, and Visual Geometry Group 16 (VGG 16), are widely used to identify visual patterns in medical images and have been applied to lung nodule detection and classification1,18.
The fatality rate of the lung cancer is primarily because of increasing prevalence of smoking. However, development in technology has made timely detection and diagnosis, potentially saving millions of lives worldwide. While CT scans are a well-established and widely used method in medicine, accurately diagnosing cancer on CT imaging remains a challenging task, even for experienced doctors and specialists. Therefore, an innovative model is needed to surmount the challenges associated with the identification of lung cancer. This research offers the following key contributions:
Introduction of WISeRKNet: A new diagnostic framework, WISeRKNet, is introduced by integrating Wide Slice Residual features with Kronecker-based representations to effectively capture variations in lung nodule size, shape, and boundary definition.
Improved nodule feature representation: The combination of wide-slice residual learning and Kronecker feature mapping enables the extraction of more discriminative and robust nodule features, supporting accurate differentiation between malignant and non-malignant nodules.
Enhanced diagnostic performance: The proposed WISeRKNet demonstrates improved precision, accuracy, TPR, F1-score, and TNR, showing its effectiveness over existing approaches.
This work is structured according to the following sections: An analysis of current models in the literature for diagnosing lung cancer is specified in Sect. “Literature survey”. Section “Proposed wide slice residual kronecker network for lung cancer detection ” provides an overview of the WISeRKNet to diagnose lung cancer using a CT image. The evaluation of the results of the devised approach is given in Sect. “Result and discussion”. The conclusion and an outline for further work are presented in Sect. “Conclusion”.
The imaging tool of biomedical, like CT imaging, is used for accurate determination and localization of lung cancer. CT is also known as a digital method for cross-sectional imaging that allows for independent manipulation of image acquisition. The high exposure of CT improves image quality as the noise level decreases. Radiologists often prefer low-noise, high-quality CT images that are visually appealing. As a result, they may hesitate to reduce the radiation dose, which typically involves lowering the X-ray exposure settings during CT acquisition and can lead to increased image noise1. CT images are produced by utilizing X-rays to generate detailed cross-sectional views of the body2. CT scans are employed to view internal organs, bones, blood vessels, and various other structures within the body. CT imaging3can reveal both suspected as well as unsuspected nodules in lung cancer. However, variations in CT scans and potential misinterpretations of anatomical structures can complicate the identification of cancerous cells for radiologists4. Cancer is a more dangerous disease which is spreading increasingly worldwide. Moreover, lung cancer is the leading cause of cancer-related mortality worldwide5. Despite advancements in treatment and early detection methods, survival rates remain low, highlighting the urgent need for improved diagnostic tools and therapeutic strategies6.
The cells’ uncontrolled growth can spread over the lungs through metastasis, affecting nearby tissues or other areas of the body7. In the first two stages, cancer is formed in the lungs, then it spreads to other parts of the organs8. Segmenting anatomical structures of lungs is a crucial task for Computer-Assisted Diagnosis (CAD)9systems using CT scans of the chest. There are five lobes in the lungs: two in the left lung and three in the right lung. These are separated by lobar fissures. Segmentation of the lobe is simple since fissures are distinctly outlined in a CT scan. However, segmentation is often not the case due to incomplete fissures, the presence of additional structures, and surrounding lung parenchymal abnormalities1,7. Furthermore, prolonged smoking of tobacco is a key reason for 85% of lung cancer cases. Approximately 10–15% cases occur in individuals who have never smoked, primarily due to factors such as secondhand smoke, asbestos, air pollution and radon gas exposure10. The majority of diagnoses occur at advanced stages of disease due to the absence of early signs of the condition. Yet, detection in an earlier stage can significantly improve survival ability11.
The Machine Learning (ML) technique is a commonly applied technique for the classification of images. In ML-based approach, features are manually mined from images for identification purposes. The Support Vector Machine (SVM) technique is an ML technique primarily used for classification tasks12,13. In recent years, Deep Learning (DL) has become highly effective for lung cancer screening because it can automatically learn useful features and help detect the disease in its early stages14. DL algorithms can integrate imaging modalities, such as Magnetic Resonance Imaging (MRI)15and CT to enhance disease detection and support diagnosis as well as treatment planning7,16,17. Convolutional Neural Networks (CNNs), such as GoogleNet, AlexNet, ResNet, DenseNet, and Visual Geometry Group 16 (VGG 16), are widely used to identify visual patterns in medical images and have been applied to lung nodule detection and classification1,18.
The fatality rate of the lung cancer is primarily because of increasing prevalence of smoking. However, development in technology has made timely detection and diagnosis, potentially saving millions of lives worldwide. While CT scans are a well-established and widely used method in medicine, accurately diagnosing cancer on CT imaging remains a challenging task, even for experienced doctors and specialists. Therefore, an innovative model is needed to surmount the challenges associated with the identification of lung cancer. This research offers the following key contributions:
Introduction of WISeRKNet: A new diagnostic framework, WISeRKNet, is introduced by integrating Wide Slice Residual features with Kronecker-based representations to effectively capture variations in lung nodule size, shape, and boundary definition.
Improved nodule feature representation: The combination of wide-slice residual learning and Kronecker feature mapping enables the extraction of more discriminative and robust nodule features, supporting accurate differentiation between malignant and non-malignant nodules.
Enhanced diagnostic performance: The proposed WISeRKNet demonstrates improved precision, accuracy, TPR, F1-score, and TNR, showing its effectiveness over existing approaches.
This work is structured according to the following sections: An analysis of current models in the literature for diagnosing lung cancer is specified in Sect. “Literature survey”. Section “Proposed wide slice residual kronecker network for lung cancer detection ” provides an overview of the WISeRKNet to diagnose lung cancer using a CT image. The evaluation of the results of the devised approach is given in Sect. “Result and discussion”. The conclusion and an outline for further work are presented in Sect. “Conclusion”.
Literature survey
Literature survey
Maleki, N. and Niaki, S.T.A19., developed an Artificial Neural Network (ANN) to recognize lung cancer. The linear discriminant and principal component analyses were utilized for the dimensionality reduction. This approach had the ability to analyze large numbers of CT scans quickly and efficiently, making them suitable for use in large-scale healthcare settings. However, the evaluation was performed using a limited feature set, which may affect the overall performance. SU, A., et al.7devised Bag of Visual Words-Convolutional Recurrent Neural Network (BoVW-CRNN) to detect lung cancer based on CT images. Here, the segmentation was done using the Guaranteed Convergence Particle Swarm Optimization (GCPSO). This approach effectively reduced human error and improved the overall accuracy of lung tumor detection. This model was unsuccessful in identify the specific locations of lung nodules and did not tune hyperparameters. Althubiti, S.A., et al.20developed gradient boosting for lung cancer detection using CT images. Here, an adaptive histogram equalization was established for image contrast improvement. This model effectively handled class imbalance through an ensemble of multiple predictive approaches, which improved detection accuracy for rare cases. The high-quality of labeled data for lung cancer detection was low because of limited training. Nazir, I., et al.5devised Laplacian Pyramid-Adaptive Sparse Representation (LP + ASR) to detect lung cancer based on a CT image. Here, the Adaptive Sparse Representation (ASR) along with the Laplacian Pyramid (LP) decomposition was utilized for the image fusion. Efficient pre-processing and segmentation speeded up the processing time, enabling faster diagnosis and treatment planning. This method was only applied to a single database.
Shakeel, P.M., et al.21introduced an improved deep neural network and ensemble classifier to detect lung cancer based on CT images. Here, to enhance the lung image quality, the multilevel brightness-preserving scheme was applied to examine every pixel and remove the noise. This method enabled precise detection, and automatic processes for effective clinical workflow. The model failed to detect the false positives that led to unnecessary biopsies. Naseer, I., et al.13developed LungNet-SVM for lung cancer detection using CT images. This method created distinct decision boundaries and showed resilience to noise, also it enabled efficient detection of cancer. It relied on a particular framework may fail to entirely reflect intricate features to other advanced deep learning architectures. Alamgeer, M., et al.22developed Deep Learning-based Computer-Aided Diagnosis for Lung Cancer using Biomedical CT images (DLCADLC-BCT) to detect lung cancer based on CT images. Here, the gray level co-occurrence matrix (GLCM) approach was utilized for feature extraction and the Long Short-Term Memory (LSTM) approach was utilized for lung cancer classification. This approach effectively handled complex patterns in CT images, allowing for precise detection of nodules in the lung. It failed to develop a segmentation method to enhance the classification of the image. Shafi, I., et al.23devised a Convolutional Neural Network-Support Vector Machine (CNN-SVM) for lung cancer detection using CT images. This scheme was initially trained by comparing and recording the selected profile values in CT images. This model effectively diagnosed lung cancer, enhancing accuracy and efficiency in processing CT images by managing complex patterns and variations. The model had intensive computational requirements as it was difficult to implement effectively in clinical settings with limited resources. Anindita Saha, et al.24established a novel transfer learning model (VER-Net) by stacking three various transfer learning approaches for lung cancer detection. This model was trained and analyzed by a multiclass classification of chest CT images. This model had better accuracy, yet, the model can only be applied to detecting lung cancer. Lavina Jean Crasta, et al.25established a DL approach, named 3D-ResNet for the detection and classification of lung cancer. Also, in this model, a Three-Dimensional Voxel-based Network (3D-VNet) structure was implemented for accurate pulmonary nodules segmentation. This model had reduced false positives. However, more analysis was required to prove its generalizability.
Maleki, N. and Niaki, S.T.A19., developed an Artificial Neural Network (ANN) to recognize lung cancer. The linear discriminant and principal component analyses were utilized for the dimensionality reduction. This approach had the ability to analyze large numbers of CT scans quickly and efficiently, making them suitable for use in large-scale healthcare settings. However, the evaluation was performed using a limited feature set, which may affect the overall performance. SU, A., et al.7devised Bag of Visual Words-Convolutional Recurrent Neural Network (BoVW-CRNN) to detect lung cancer based on CT images. Here, the segmentation was done using the Guaranteed Convergence Particle Swarm Optimization (GCPSO). This approach effectively reduced human error and improved the overall accuracy of lung tumor detection. This model was unsuccessful in identify the specific locations of lung nodules and did not tune hyperparameters. Althubiti, S.A., et al.20developed gradient boosting for lung cancer detection using CT images. Here, an adaptive histogram equalization was established for image contrast improvement. This model effectively handled class imbalance through an ensemble of multiple predictive approaches, which improved detection accuracy for rare cases. The high-quality of labeled data for lung cancer detection was low because of limited training. Nazir, I., et al.5devised Laplacian Pyramid-Adaptive Sparse Representation (LP + ASR) to detect lung cancer based on a CT image. Here, the Adaptive Sparse Representation (ASR) along with the Laplacian Pyramid (LP) decomposition was utilized for the image fusion. Efficient pre-processing and segmentation speeded up the processing time, enabling faster diagnosis and treatment planning. This method was only applied to a single database.
Shakeel, P.M., et al.21introduced an improved deep neural network and ensemble classifier to detect lung cancer based on CT images. Here, to enhance the lung image quality, the multilevel brightness-preserving scheme was applied to examine every pixel and remove the noise. This method enabled precise detection, and automatic processes for effective clinical workflow. The model failed to detect the false positives that led to unnecessary biopsies. Naseer, I., et al.13developed LungNet-SVM for lung cancer detection using CT images. This method created distinct decision boundaries and showed resilience to noise, also it enabled efficient detection of cancer. It relied on a particular framework may fail to entirely reflect intricate features to other advanced deep learning architectures. Alamgeer, M., et al.22developed Deep Learning-based Computer-Aided Diagnosis for Lung Cancer using Biomedical CT images (DLCADLC-BCT) to detect lung cancer based on CT images. Here, the gray level co-occurrence matrix (GLCM) approach was utilized for feature extraction and the Long Short-Term Memory (LSTM) approach was utilized for lung cancer classification. This approach effectively handled complex patterns in CT images, allowing for precise detection of nodules in the lung. It failed to develop a segmentation method to enhance the classification of the image. Shafi, I., et al.23devised a Convolutional Neural Network-Support Vector Machine (CNN-SVM) for lung cancer detection using CT images. This scheme was initially trained by comparing and recording the selected profile values in CT images. This model effectively diagnosed lung cancer, enhancing accuracy and efficiency in processing CT images by managing complex patterns and variations. The model had intensive computational requirements as it was difficult to implement effectively in clinical settings with limited resources. Anindita Saha, et al.24established a novel transfer learning model (VER-Net) by stacking three various transfer learning approaches for lung cancer detection. This model was trained and analyzed by a multiclass classification of chest CT images. This model had better accuracy, yet, the model can only be applied to detecting lung cancer. Lavina Jean Crasta, et al.25established a DL approach, named 3D-ResNet for the detection and classification of lung cancer. Also, in this model, a Three-Dimensional Voxel-based Network (3D-VNet) structure was implemented for accurate pulmonary nodules segmentation. This model had reduced false positives. However, more analysis was required to prove its generalizability.
Proposed wide slice residual Kronecker network for lung cancer detection
Proposed wide slice residual Kronecker network for lung cancer detection
A block diagram of WISeRKNet for the detection of lung cancer is depicted in Fig. 1.Initially, a CT image acquired from the Lung Cancer Computed Tomography Images database26is given to pre-processing in order to eliminate unwanted noise by means of homomorphic filtering27. After that, lung nodule extraction is performed using Link-Net28. Then, augmentation is performed based on Rotation, Random erasing and flipping methods. Here, Shape-Based Features like Nodule Irregularity Index, nodules area, Nodule Solidity, nodule perimeter, Nodule Equivalent29are extracted. Finally, detection of lung cancer is done using the proposed WISeRKNet, which is the integration of WISeR30and DKN31.
Image acquisition
Consider a CT image, which is taken from U dataset and it is described as follows,
Where, defines the quantity of CT images that are presented in the dataset U and implies CT image that is chosen to detect lung cancer.
Homomorphic filtering-based image pre-processing
The objective of mage pre-processing is to improve the visibility of structures and features within the images while minimizing artifacts or noise. Here, homomorphic filtering is employed to limit noise interference and improve the quality of the input image. Homomorphic filtering27is a technique and it is employed to counteract effects like illumination of uneven images while simultaneously enhancing their appearance through variable intensity and contrast adjustments during image compression. This step ensures that important features in CT images are more distinguishable, which leads to better segmentation and feature extraction. The homomorphic filtering is computed by,
Thus, the multiplication of the image result is denoted as, is a component of. The pre-processed result is specified as.
Lung lobe extraction using Link-Net
Lung lobe extraction in a CT image is a computational or manual process of isolating or segmenting the individual lobes of the lung from the surrounding structures within a CT scan. The purpose of this extraction is to facilitate improved analysis, visualization, and quantification of the lung lobes for various clinical applications. Hence, the Link-Net28is used to extract the lung lobe by using a pre-processed image.
Link-Net
Link-Net28is a DL-based architecture designed for semantic segmentation, primarily aimed at tasks that require precise pixel-level classification of images. It is very helpful for medical imaging where accurate segmentation of anatomical structures is critical, such as lung lobe extraction. In Link-Net refers to convolution, and full convolution is specified as. Furthermore, the factor of down sampling 2 is denoted as to achieve stride convolution, and the up-sampling feature is the factor of 2. The block is applied in spatial max pooling over an area with a size for a stride 2. Link-Net employ ResNet18 as a lightweight network that still provides superior performance. Then, the full convolution technique is utilized in the decoder. Therefore, alland applications have three parameters. Here, the kernel size is defined as and the input map and output map are denoted as. Therefore, the extracted lung lobe is specified as . The architecture of Link-Net is specified in Fig. 2.
Image augmentation based on lung nodule extraction
Image augmentation techniques help to enhance the performance of diagnostic models by increasing the dimensionality of data. Thus, the lung lobe extracted image is employed for the image augmentation process. The different techniques used in the image augmentation are specified as below,
Rotation
Rotation32specifies the process to rotate the image around the center by a particular angle. This technique involves modifying original photos to increase their size artificially, and it is used to enhance robustness.
Here, the location of each pixel after undergoing rotation is implied as and, the coordination of the image is represented as r, s. Thus, the rotation outcome is denoted as
Random erasing
This technique32is used to enhance robustness, and a randomly chosen rectangular section of the image has its pixel values deleted, usually by setting them to zero or replacing them with random noise and a constant value. It enhances the generalization of the model and avoids overfitting. The random erasing outcome is defined as.
Flipping
Flipping32is a common technique that creates a mirror image of the original image by flipping it along a particular axis. It also helps to enhance the image by increasing its variability and promoting better training for the model. The flipping process has different kinds of flipping, which are provided below.
Vertical flipping
Vertical flipping32is flipped along with the horizontal axis to create a mirror image of itself from top to bottom. The and value is the initial coordinates of every pixel after the flipping process along the vertical axis, and it is expressed as,
Horizontal flipping
Horizontal flipping32creates a mirrored version of the original image from left to right. It is illustrated as follows,
Hence, flipping is denoted as. Accordingly, the image augmentation formulation is given as,
Feature extraction
This is a crucial step in medical image analysis, as these features can help in diagnosing diseases, evaluating treatment responses, and predicting patient outcomes. The augmented image , and thus, the extracted features are Shape-Based Features (SBF).
SBF
SBF analyzes characteristics at the pixel level and possess inherent properties that are instinctive and optical in nature. The shape-based features are explained below:
Nodule perimeter
Nodule perimeter29is the total length of the boundary or edge of a nodule that is used to quantify the contour of a nodule, providing critical information about its geometry. The perimeter is typically derived from a segmented region of interest (ROI) representing a nodule. The nodule perimeter is described as,
Thus, length is given as G, width is specified as F and regional pixel intensity is implied as .
Nodule area
Nodule area29measures a 2-D surface that is occupied by a nodule in a cross-sectional image; it is typically represented in square millimeters. It is used for nodule size estimation from surrounding tissue in the image.
Where, suspicious region of pixel value is represented as.
Nodule irregularity index
Nodule Irregularity Index29is used to measure the degree of irregularity or deviation from a perfectly circular shape. This feature helps in discriminating between benign and malignant nodules using geometric properties.
Thus, area is denoted as and the perimeter of the nodule is specified as.
Nodule solidity
Nodule Solidity29is particularly employed to calculate the compactness or fill of a nodule relative to its convex hull, which is the smallest convex shape that can completely enclose the nodule. Hence, nodule solidity is calculated as,
Nodule equivalent
Nodule Equivalent29is known as a geometric measurement, and it is used as a standardized representation of size and shape. It is particularly relevant in the context of identification in imaging, and it is computed as,
Hence, the expression for feature extraction is given as,
Lung cancer detection using Wiserknet model
In this research, a novel model called WISeRKNet has been developed to recognize lung cancer. Here, WISeRKNet is the combination of WISeR and DKN network. The tuning process of the hybrid network is performed based on Adam optimization. Figure 3 specifies a general outline of the WISeRKNet model. The process begins with an input CT image is fed into the WISeR model. Then, the outcome and feature extracted outcome is considered as input for the WISeRKNet model, such as . Furthermore, the result of WISeRKNet is implied as and it employed fusion as well as a regression phase to capture more complex patterns to improve prediction accuracy. At last, the result from WISeRKNet and the input image is subjected as input to the DKN, like that. Thus, the final result is .
WISeR
WISeR30excels in detection tasks within CT images as it has the ability to enhance image quality. The WISeR preserves spatial dependencies across larger regions, which enables better identification of subtle variations in nodule structures. The residual learning architecture addresses the vanishing gradient issue, enabling the use of deeper networks as it converges more quickly while preserving the performance. The broader architecture facilitates the capture of intricate features, which is essential in medical imaging, as subtle variations can signify important changes in pathological. The model seeks to accomplish this objective by identifying structural nuances of the image through integration of a slice convolution layer through residual learning.
Residual learning
Here, J is given to shallow network through a few stacked with non-linear layers that are estimated as a mapping function and are denoted as. The network has the same structure as it is approximately for the residual function. Moreover, the learning approximation of the mapping function and residual function is specified as, which is feasible. The simplicity of this process varies greatly. The mapping function is not encountered in the network to train the approximate residual function.
Wide residual blocks
The network utilizes residual learning after every few stacked layers. Therefore, the residual block for identity mapping is specified as,
Here, , and represent a group of corresponding parameters with the residual block, u specifies the layer count in a residual block. The residual learning objective is to find the parameters that approximate of function . Thus, the resultant output is implied as . The architecture of WS-ResNet is specified in Fig. 4.
WISeRKNet model
The lung cancer detection is performed using the WISeRKNet model. This model captures more information and variability in the CT image, as it is particularly suitable for complex lung cancer detection. The WISeR outcome and the outcome from is given as input for the WISeRKNet model. This input is subsequently processed through fusion and regression operations, resulting in an output as .
where, denotes the weighting parameter associated with the feature, and e implies total number of feature components included in the fusion operation. Thus, output from interval is considered as B. Based on Fractional Calculus (FC),
where, denotes the feature representation at the current interval m, signifies feature representation at the previous interval, and implies feature representation at the next interval, and indicates the Fractional coefficient. Let us assume,
Thus, Eq. (15) becomes,
Therefore, the attained result is implied as .
DKN model
DKN31is a type of neural network architecture that employs the Kronecker product to efficiently model high-dimensional data. It is particularly useful in scenarios where the image can be decomposed into lower-dimensional representations that can be combined to capture complex interactions. The sample y is observed with a matrix representation of images along with scalar responses. Let us consider the is tracked by generalized linear method31as,
Here, specifies the unknown target of the coefficient’s matrix, and defines a univariate function. The outcome obtained from DKN is illustrated as . The architecture of DKN is illustrated in Fig. 5.
The pseudocode of the WISeRKNet model is given in Algorithm 1.
A block diagram of WISeRKNet for the detection of lung cancer is depicted in Fig. 1.Initially, a CT image acquired from the Lung Cancer Computed Tomography Images database26is given to pre-processing in order to eliminate unwanted noise by means of homomorphic filtering27. After that, lung nodule extraction is performed using Link-Net28. Then, augmentation is performed based on Rotation, Random erasing and flipping methods. Here, Shape-Based Features like Nodule Irregularity Index, nodules area, Nodule Solidity, nodule perimeter, Nodule Equivalent29are extracted. Finally, detection of lung cancer is done using the proposed WISeRKNet, which is the integration of WISeR30and DKN31.
Image acquisition
Consider a CT image, which is taken from U dataset and it is described as follows,
Where, defines the quantity of CT images that are presented in the dataset U and implies CT image that is chosen to detect lung cancer.
Homomorphic filtering-based image pre-processing
The objective of mage pre-processing is to improve the visibility of structures and features within the images while minimizing artifacts or noise. Here, homomorphic filtering is employed to limit noise interference and improve the quality of the input image. Homomorphic filtering27is a technique and it is employed to counteract effects like illumination of uneven images while simultaneously enhancing their appearance through variable intensity and contrast adjustments during image compression. This step ensures that important features in CT images are more distinguishable, which leads to better segmentation and feature extraction. The homomorphic filtering is computed by,
Thus, the multiplication of the image result is denoted as, is a component of. The pre-processed result is specified as.
Lung lobe extraction using Link-Net
Lung lobe extraction in a CT image is a computational or manual process of isolating or segmenting the individual lobes of the lung from the surrounding structures within a CT scan. The purpose of this extraction is to facilitate improved analysis, visualization, and quantification of the lung lobes for various clinical applications. Hence, the Link-Net28is used to extract the lung lobe by using a pre-processed image.
Link-Net
Link-Net28is a DL-based architecture designed for semantic segmentation, primarily aimed at tasks that require precise pixel-level classification of images. It is very helpful for medical imaging where accurate segmentation of anatomical structures is critical, such as lung lobe extraction. In Link-Net refers to convolution, and full convolution is specified as. Furthermore, the factor of down sampling 2 is denoted as to achieve stride convolution, and the up-sampling feature is the factor of 2. The block is applied in spatial max pooling over an area with a size for a stride 2. Link-Net employ ResNet18 as a lightweight network that still provides superior performance. Then, the full convolution technique is utilized in the decoder. Therefore, alland applications have three parameters. Here, the kernel size is defined as and the input map and output map are denoted as. Therefore, the extracted lung lobe is specified as . The architecture of Link-Net is specified in Fig. 2.
Image augmentation based on lung nodule extraction
Image augmentation techniques help to enhance the performance of diagnostic models by increasing the dimensionality of data. Thus, the lung lobe extracted image is employed for the image augmentation process. The different techniques used in the image augmentation are specified as below,
Rotation
Rotation32specifies the process to rotate the image around the center by a particular angle. This technique involves modifying original photos to increase their size artificially, and it is used to enhance robustness.
Here, the location of each pixel after undergoing rotation is implied as and, the coordination of the image is represented as r, s. Thus, the rotation outcome is denoted as
Random erasing
This technique32is used to enhance robustness, and a randomly chosen rectangular section of the image has its pixel values deleted, usually by setting them to zero or replacing them with random noise and a constant value. It enhances the generalization of the model and avoids overfitting. The random erasing outcome is defined as.
Flipping
Flipping32is a common technique that creates a mirror image of the original image by flipping it along a particular axis. It also helps to enhance the image by increasing its variability and promoting better training for the model. The flipping process has different kinds of flipping, which are provided below.
Vertical flipping
Vertical flipping32is flipped along with the horizontal axis to create a mirror image of itself from top to bottom. The and value is the initial coordinates of every pixel after the flipping process along the vertical axis, and it is expressed as,
Horizontal flipping
Horizontal flipping32creates a mirrored version of the original image from left to right. It is illustrated as follows,
Hence, flipping is denoted as. Accordingly, the image augmentation formulation is given as,
Feature extraction
This is a crucial step in medical image analysis, as these features can help in diagnosing diseases, evaluating treatment responses, and predicting patient outcomes. The augmented image , and thus, the extracted features are Shape-Based Features (SBF).
SBF
SBF analyzes characteristics at the pixel level and possess inherent properties that are instinctive and optical in nature. The shape-based features are explained below:
Nodule perimeter
Nodule perimeter29is the total length of the boundary or edge of a nodule that is used to quantify the contour of a nodule, providing critical information about its geometry. The perimeter is typically derived from a segmented region of interest (ROI) representing a nodule. The nodule perimeter is described as,
Thus, length is given as G, width is specified as F and regional pixel intensity is implied as .
Nodule area
Nodule area29measures a 2-D surface that is occupied by a nodule in a cross-sectional image; it is typically represented in square millimeters. It is used for nodule size estimation from surrounding tissue in the image.
Where, suspicious region of pixel value is represented as.
Nodule irregularity index
Nodule Irregularity Index29is used to measure the degree of irregularity or deviation from a perfectly circular shape. This feature helps in discriminating between benign and malignant nodules using geometric properties.
Thus, area is denoted as and the perimeter of the nodule is specified as.
Nodule solidity
Nodule Solidity29is particularly employed to calculate the compactness or fill of a nodule relative to its convex hull, which is the smallest convex shape that can completely enclose the nodule. Hence, nodule solidity is calculated as,
Nodule equivalent
Nodule Equivalent29is known as a geometric measurement, and it is used as a standardized representation of size and shape. It is particularly relevant in the context of identification in imaging, and it is computed as,
Hence, the expression for feature extraction is given as,
Lung cancer detection using Wiserknet model
In this research, a novel model called WISeRKNet has been developed to recognize lung cancer. Here, WISeRKNet is the combination of WISeR and DKN network. The tuning process of the hybrid network is performed based on Adam optimization. Figure 3 specifies a general outline of the WISeRKNet model. The process begins with an input CT image is fed into the WISeR model. Then, the outcome and feature extracted outcome is considered as input for the WISeRKNet model, such as . Furthermore, the result of WISeRKNet is implied as and it employed fusion as well as a regression phase to capture more complex patterns to improve prediction accuracy. At last, the result from WISeRKNet and the input image is subjected as input to the DKN, like that. Thus, the final result is .
WISeR
WISeR30excels in detection tasks within CT images as it has the ability to enhance image quality. The WISeR preserves spatial dependencies across larger regions, which enables better identification of subtle variations in nodule structures. The residual learning architecture addresses the vanishing gradient issue, enabling the use of deeper networks as it converges more quickly while preserving the performance. The broader architecture facilitates the capture of intricate features, which is essential in medical imaging, as subtle variations can signify important changes in pathological. The model seeks to accomplish this objective by identifying structural nuances of the image through integration of a slice convolution layer through residual learning.
Residual learning
Here, J is given to shallow network through a few stacked with non-linear layers that are estimated as a mapping function and are denoted as. The network has the same structure as it is approximately for the residual function. Moreover, the learning approximation of the mapping function and residual function is specified as, which is feasible. The simplicity of this process varies greatly. The mapping function is not encountered in the network to train the approximate residual function.
Wide residual blocks
The network utilizes residual learning after every few stacked layers. Therefore, the residual block for identity mapping is specified as,
Here, , and represent a group of corresponding parameters with the residual block, u specifies the layer count in a residual block. The residual learning objective is to find the parameters that approximate of function . Thus, the resultant output is implied as . The architecture of WS-ResNet is specified in Fig. 4.
WISeRKNet model
The lung cancer detection is performed using the WISeRKNet model. This model captures more information and variability in the CT image, as it is particularly suitable for complex lung cancer detection. The WISeR outcome and the outcome from is given as input for the WISeRKNet model. This input is subsequently processed through fusion and regression operations, resulting in an output as .
where, denotes the weighting parameter associated with the feature, and e implies total number of feature components included in the fusion operation. Thus, output from interval is considered as B. Based on Fractional Calculus (FC),
where, denotes the feature representation at the current interval m, signifies feature representation at the previous interval, and implies feature representation at the next interval, and indicates the Fractional coefficient. Let us assume,
Thus, Eq. (15) becomes,
Therefore, the attained result is implied as .
DKN model
DKN31is a type of neural network architecture that employs the Kronecker product to efficiently model high-dimensional data. It is particularly useful in scenarios where the image can be decomposed into lower-dimensional representations that can be combined to capture complex interactions. The sample y is observed with a matrix representation of images along with scalar responses. Let us consider the is tracked by generalized linear method31as,
Here, specifies the unknown target of the coefficient’s matrix, and defines a univariate function. The outcome obtained from DKN is illustrated as . The architecture of DKN is illustrated in Fig. 5.
The pseudocode of the WISeRKNet model is given in Algorithm 1.
Result and discussion
Result and discussion
The effectiveness of WISeRKNet is analyzed by altering training data and K-fold using various metrics.
Experimental setup
The proposed WISeRKNet is implemented by Python tool version 3.9.11 on Windows 10 OS, 4GB RAM, more than 5 GB ROM, and 1.7 GHz CPU. The parameters of WISeRKNet are depicted in Table 1.
Dataset description
Lung cancer computed tomography images database (Dataset-1):
The dataset used for implementing the model is the Lung Cancer Computed Tomography Images database26for lung cancer detection. The dataset has CT images of lungs potentially categorized into different classes such as, Benign tumors, Malignant tumors and Healthy lung tissue. All images were stored using the Digital Imaging and Communications in Medicine (DICOM) standard.
Iraq-Oncology Teaching Hospital/National Center for Cancer Diseases (IQ-OTH/NCCD) Lung Cancer Dataset (Dataset-2):
The IQ-OTH/NCCD lung cancer dataset34is collected over a three-month period in fall 2019 at the Iraq Oncology Teaching Hospital and the National Center for Cancer Diseases. It consists of 1190 CT scan slices from 110 cases, including 40 malignant, 15 benign, and 55 normal subjects. CT images were obtained using a Siemens SOMATOM scanner in DICOM format with a 120 kV protocol, 1 mm slice thickness, window width ranging from 350 to 1200 HU, and window center from 50 to 600 HU, acquired during full-inspiration breath-hold. Each patient scan contains 80–200 slices, annotated by expert oncologists and radiologists. The dataset includes diverse individuals varying in age, gender, and socioeconomic background, primarily from central Iraqi provinces such as Baghdad, Wasit, Diyala, Salahuddin, and Babylon. In this study, a patient-level split is followed to prevent data leakage. For training (90%), data from 36 malignant, 13 benign, and 49 normal patients are used, while the remaining patients are reserved for testing.
In this research, the dataset is split into training and testing sets. Training data ranges from 50% to 90%, with the remaining data used for testing. For instance, when 50% of the data is used for training, 50% is used for testing, and similarly for 60–90% training. The corresponding test results are plotted in the graphs, with the X-axis representing the training data percentage.
Also, in this study, k-fold cross-validation was employed to ensure a reliable and unbiased evaluation of the WISeRKNet. The dataset was divided into k equal parts, with the value of k varying from 5 to 9. During each iteration, a single fold was used for testing and the remaining k − 1 folds for training. The process continued until all folds had been used once as the test set.
Evaluation metrics
The WISeRKNet is analyzed using different metrics, such as accuracy, TPR and TNR.
Accuracy
Accuracy represents the fraction of correctly classified instances out of the total instances in the WISeRKNet. Accuracy33is specified as,
where, specifies true positive and implies true negative, denotes false positive and expresses false negative.
TPR
TPR33measures the proportion of actual positive instances or true cases that are correctly identified by the model as positive.
Hence, TPR is defined as .
TNR
TPR33calculates the proportion of actual negative instances that are correctly identified by the model as negative.
Thus, TNR is specified as .
Precision
It is the ratio of true positives to the total number of positive predictions.
F1-score
The F1-Score merges Precision and Recall into one metric through their harmonic mean, ensuring a balance between them.
Experimental outcomes
The experimental results of images 1 and 2 obtained by the WISeRKNet is represented in Fig. 6. Figure 6a) represents the input lung CT images-1 and 2, Fig. 6b) implies pre-processed result of images-1 and 2, nodule extracted images-1 and 2 is illustrated in Fig. 6c), Fig. 6d) represents the rotated images-1 and 2 result and random erased result of images-1 and 2 is implied in Fig. 6e), Fig. 6f) implies flipped images-1 and 2, and the detected outcomes of images-1 and 2 are given in Fig. 6g).
Performance analysis for WISeRKNet
The performance analysis of the WISeRKNet model is conducted based on altering training data and K-fold value by using a CT image.
Evaluation of performance based on training data
Figure 7 provides a comprehensive evaluation of the efficacy of the WISeRKNet model by varying training data. The analysis of WISeRKNet using accuracy is depicted in Fig. 7a). The WISeRKNet obtained accuracy of 80.604%, 83.307%, 85.950%, 86.755% and 91.508% with epochs 50, 60, 70, 80 and 90, by considering training data as 90%. Figure 7b) presents the efficiency of WISeRKNet utilizing TPR. For training data = 90%, the TPR acquired by the developed WISeRKNet of 80.604%, 83.307%, 85.950%, 86.755% and 90.408% with epochs 50, 60, 70, 80 and 90. The assessment for WISeRKNet based on TNR is specified in Fig. 7c). The TNR attained by WISeRKNet of 78.597% for epochs 50, 82.394% for epochs 60, 84.731% for epochs 70, 88.902% for epochs 80 and 92.700% for epochs 90. The effectiveness of WISeRKNet using precision is seen in Fig. 7d). With epochs 50, 60, 70, 80, and 90, the precision obtained by the constructed WISeRKNet was 78.655%, 82.951%, 84.616%, 88.966%, and 90.863%. Figure 7e details the F1 score-based evaluation for WISeRKNet. For the 50th, 60th, 70th, 80th, and 90th epochs, WISeRKNet achieved a F1 score of 80.604%, 83.307%, 85.949%, 86.754%, and 90.407%, respectively.
Comparative analysis
The comparative assessment for WISeRKNet is performed using traditional models, such as ANN19, BoVW-CRNN7, Gradient boosting20, LP + ASR5, VER-Net24, and 3D-ResNet25. All comparative models were evaluated under the same experimental conditions, including the same dataset, training–testing split, preprocessing steps, and evaluation metrics. This section describes a comparative evaluation of WISeRKNet by modifying a few-shot samples, training data and k-fold.
Evaluation using dataset-1
Few shot sample generation
Few-shot samples are first generated by dividing the minimum number of samples from each class. The dataset consists of 928 samples, categorized into two classes: tumor and non-tumor. For 10% of the minimum class, a total of approximately 40 samples (20 tumor and 20 non-tumor) are selected for training. Similarly, using 15% of the minimum class, 60 samples are chosen, consisting of roughly 30 tumor and 30 non-tumor samples. For 20% of the minimum class, 80 samples are selected, with about 40 tumor and 40 non-tumor samples. Finally, when 25% of the minimum class is used, a total of 100 samples are taken, equally divided between tumor (50) and non-tumor (50) for training.
Assessment for WISeRKNet based on few shot samples
The analysis of WISeRKNet used for lung cancer detection by changing few-shot sample using dataset-1 is specified in Fig. 8. In Fig. 8a), the assessment of the WISeRKNet model with traditional models in terms of accuracy is evaluated. For 100 few shot sample, the testing accuracy acquired by WISeRKNet is 90.944% and conventional models, such as VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR measured accuracy as 84.137%, 86.144%, 79.507%, 82.134%, 85.854% and 87.902%.
Moreover, the effectiveness WISeRKNet model regarding TPR is specified as Fig. 8b). For 100 few shot sample, the TPR attained by WISeRKNet is 89.911% and the existing models, such as VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR measured TPR as 82.562%, 84.589%, 80.026%, 83.638%, 84.247% and 86.316%. Accordingly, the developed model in terms of TNR is implied as Fig. 8c). The TNR attained by WISeRKNet is 92.088% and the existing models, such as VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR acquired TNR as 83.239%, 85.449%, 78.929%, 82.342%, 84.938% and 87.193%. Figure 8d compares the precision of the WISeRKNet model to that of conventional models. Conventional models like VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, and LP + ASR tested precision as 84.666%, 82.864%, 79.846%, 82.777%, 84.555%, and 86.393% for a 100 few shot sample, but WISeRKNet obtained testing precision of 90.059%. The F1-score comparison of the WISeRKNet model with conventional models is shown in Fig. 8e). WISeRKNet achieved an F1-Score of 89.911% for a 100 few shot sample, whereas VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, and LP + ASR measured an F1-score of 84.589%, 82.562%, 80.025%, 83.638%, 84.246%, and 86.315%.
Assessment for WISeRKNet based on training data
The evaluation of the WISeRKNet model by modifying training data using dataset-1 is illustrated in Fig. 9. The accuracy assessment of the WISeRKNet model is implied in Fig. 9a). The WISeRKNet acquired an accuracy of 91.686%, while considering training data = 90%. Then, the accuracy measured by existing models, like VER-Net is 83.296%, 3D-ResNet is 85.504%, ANN is 79.471%, BoVW-CRNN is 82.127%, Gradient boosting is 84.995% and LP + ASR is 87.249% while using training data as 90%. Specifically, WISeRKNet outperforms VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR by a significant value of 9.151%, 6.743%, 13.323%, 10.426%, 7.298% and 4.840%.
In Fig. 9b), the evaluation of WISeRKNet based on TPR is presented. The TPR value acquired by WISeRKNet as 90.485%, Thus, prior models like VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR attained TPR of 83.557%, 85.317%, 80.569%, 83.760%, 85.262% and 87.058%. Figure 9c) demonstrates analysis of the WISeRKNet model concerning TNR, when training data = 90%. The WISeRKNet attained TNR as 92.727%, wherein TNR obtained by existing approaches like VER-Net is 83.298%, 3D-ResNet is 85.827%, ANN is 78.030%, BoVW-CRNN is 82.874%, Gradient boosting is 84.998% and LP + ASR is 87.578% while using training data as 90%. Analysis of the WISeRKNet model with respect to Precision is shown in Fig. 9d) when the training data is 90%. With 90% training data, the WISeRKNet achieved a Precision of 90.980%, compared to 86.913% for VER-Net, 83.216% for 3D-ResNet, 80.940% for ANN, 83.292% for BoVW-CRNN, 84.913% for gradient boosting, and 90.980% for LP + ASR. When training data reaches 90%, Fig. 9e) shows the examination of the WISeRKNet model with respect to F1 score. F1 score for the WISeRKNet was 90.484%, compared to 85.317% for VER-Net, 83.557% for 3D-ResNet, 80.569% for the ANN, 83.759% for the BoVW-CRNN, 85.262% for gradient boosting, and 87.058% for LP + ASR when 90% training data was used.
Assessment for WISeRKNet based on K-fold
The evaluation of the WISeRKNet model by modifying K-fold using dataset-1 is given in Fig. 10.
Figure 10a) demonstrates analysis of the WISeRKNet model based on accuracy, while k-fold = 9. Then, the devised WISeRKNet approach obtained an accuracy of 91.209%, whereas accuracy attained by traditional models like VER-Net is 84.084%, 3D-ResNet is 87.049%, ANN is 79.470%, BoVW-CRNN is 82.433%, Gradient boosting is 85.800% and LP + ASR is 88.825% while considering training data as 90%. The analysis of the WISeRKNet method as well as other traditional methods using TPR is specified in Fig. 10b). The devised WISeRKNet model outperforms existing methods with TPR of 90.160%, when k-fold = 9. Then, TPR acquired by conventional model, like VER-Net is 83.811%, 3D-ResNet is 84.950%, ANN is 80.222%, BoVW-CRNN is 83.575%, Gradient boosting is 85.521%and LP + ASR is 86.684%. Specifically, WISeRKNet outperforms VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR by a significant value of 7.042%, 5.779%, 11.023%, 7.304%, 5.146% and 3.856%. In Fig. 10c), the evaluation of WISeRKNet based on TNR is presented. The TNR value acquired by WISeRKNet as 92.395%, while existing models, such as VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR attained TNR of 83.658%, 86.094%, 79.225%, 83.549%, 85.365% and 87.851%, with k-fold = 9. Figure 10d) shows the WISeRKNet assessment based on Precision. WISeRKNet achieved a precision value of 90.979%, whereas other models, including VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, and LP + ASR, achieved TNR values of 84.812%, 83.561%, 79.617%, 82.743%, 85.266%, and 86.542%, respectively, using k-fold = 9. The F1 score-based assessment of WISeRKNet is shown in Fig. 10e. While current models like VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, and LP + ASR achieved an F1-score of 84.950%, 83.811%, 80.222%, 83.575%, 85.520%, and 86.684%, with k-fold = 9, WISeRKNet achieved an F1 score of 90.160%.
Evaluation using dataset-2
The comparative evaluation on Dataset-2 with varying training data proportions is presented in Fig. 11, where accuracy, TPR, TNR, precision, and F1-score are shown in Figs. 11(a), 11(b), 11(c), 11(d), and 11(e), respectively. When 90% of the data is used for training, the accuracy achieved by VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, LP + ASR, and the proposed WISeRKNet is 82.463%, 84.649%, 78.677%, 81.306%, 84.145%, 86.377%, and 90.769%, respectively. Similarly, the corresponding TPR values are 82.722%, 84.464%, 79.764%, 82.922%, 84.410%, 86.188%, and 89.580%, while the TNR values are 82.465%, 84.968%, 77.250%, 82.046%, 84.148%, 86.702%, and 91.800%. The precision values recorded are 86.044%, 82.383%, 80.131%, 82.460%, 84.065%, 87.800%, and 90.070%, and the F1-scores obtained are 84.464%, 82.722%, 79.764%, 82.922%, 84.410%, 86.188%, and 89.580%, respectively. These results indicate that the proposed WISeRKNet delivers superior performance across all evaluation metrics when using 90% of the training data in Dataset-2.
Ablation study
The ablation study is shown in Fig. 12, highlighting how the proposed model performs relative to its variants. The ablation study using dataset-1 is given in Fig. 12a) and dataset-2 is Fig. 12b). When the training data is 90%, the accuracy recorded by the WISeRKNet without homomorphic filtering, WISeRKNet without Link-Net, WISeRKNet without data augmentation, WISeRKNet without feature extraction, WISeR, DKN, and WISeRKNet is 85.116%, 86.258%, 87.258%, 88.258%, 89.256%, 90.249%, and 91.526% using dataset-1 and 84.264%, 85.396%, 86.386%, 87.376%, 88.363%, 89.347%, and 90.611% using dataset-2. The results clearly indicate that each component contributes to performance improvement, with the full WISeRKNet achieving the highest accuracy.
Analysis by varying noise
Figure 13 presents the performance analysis under varying noise levels, with the results for Dataset-1 shown in Fig. 13(a) and Dataset-2 displayed in Fig. 13(b). When applying 25% noise, the accuracy gained by the VER-Net is 70.579%, 3D-ResNet is 72.572%, ANN is 67.586%, BoVW-CRNN is 70.029%, Gradient boosting is 72.658%, LP + ASR is 74.589%, and proposed WISeRKNet is 76.258% using dataset-1, while considering training data as 90%. Similarly, for dataset-2, the VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, LP + ASR, and the proposed WISeRKNet is obtained the accuracy of 69.873%, 71.846%, 66.910%, 69.329%, 71.932%, 73.843%, and 75.496% at noise level 25%. Although the overall performance decreases as the noise level increases, the proposed WISeRKNet consistently outperforms all comparative models. This demonstrates its robustness and noise-handling capability in lung cancer detection.
Feature extraction analysis
The feature extraction analysis is shown in Fig. 14. Figure 14a) shows the feature extraction analysis using dataset-1 and Fig. 14b) depicts the analysis using dataset-2. Using 90% of the dataset for training, the model attained an accuracy of WISeRKNet without feature extraction, WISeRKNet with Deep features, and WISeRKNet with shape-based features is 86.258%, 88.260%, and 91.006%, using dataset-1 and 85.396%, 87.377%, and 90.095% using dataset-2. These results confirm that feature extraction, particularly shape-based features, plays a significant role in improving the detection capability of WISeRKNet.
T-Test
The performance of the WISeRKNet was statistically validated using a T-test by comparing it with VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, and LP + ASR using dataset-1. The corresponding results are reported in Table 2. The T-values obtained were all higher than the critical value at a 97% confidence level, highlighting a clear performance difference between WISeRKNet and the other models. Additionally, the P-values ranging from 0.02 to 0.03 were below the 0.05 significance threshold, confirming that the improvements achieved by WISeRKNet are statistically significant.
Variability analysis
The variability analysis of the WISeRKNet and existing models in terms of mean, standard deviation (SD), variance, and confidence interval (CI) using dataset-1 is presented in Table 3. These values are derived from k-fold analysis using various evaluation metrics, with the CI calculated at a 97% confidence level. The proposed WISeRKNet demonstrates higher mean and CI values while maintaining lower SD and variance compared to existing methods.
Comparative discussion
The developed WISeRKNet was tested for lung cancer detection on two datasets, evaluated using multiple metrics, and its results were compared with state-of-the-art methods, as summarized in Table 4.With training data = 90%, the WISeRKNet achieved a high accuracy rate of 91.686%, when compared to VER-Net is 83.296%, 3D-ResNet is 85.504%, ANN is 79.471%, BoVW-CRNN is 82.127%, Gradient boosting is 84.995% and LP + ASR is 87.249%. The performances attained by WISeRKNet for TPR as 90.485%, where TPR for VER-Net is 83.557%, 3D-ResNet is 85.317%, ANN is 80.569%, BoVW-CRNN is 83.760%, Gradient boosting is 85.262% and LP + ASR is 87.058%. Furthermore, TNR acquired by the WISeRKNet model as 92.727%, when compared to VER-Net is 83.298%, 3D-ResNet is 85.827%, ANN is 78.030%, BoVW-CRNN is 82.874%, Gradient boosting is 84.998% and LP + ASR is 87.578%. The highest precision attained by WISeRKNet is 90.980%, where precision for VER-Net is 86.913%, 3D-ResNet is 83.216%, ANN is 80.940%, BoVW-CRNN is 83.292%, Gradient boosting is 84.913% and LP + ASR is 88.686%. F1-score acquired by the developed WISeRKNet model as 90.484%, when compared to VER-Net is 85.317%, 3D-ResNet is 83.557%, ANN is 80.569%, BoVW-CRNN is 83.759%, Gradient boosting is 85.262% and LP + ASR is 87.058%.
The effectiveness of WISeRKNet is analyzed by altering training data and K-fold using various metrics.
Experimental setup
The proposed WISeRKNet is implemented by Python tool version 3.9.11 on Windows 10 OS, 4GB RAM, more than 5 GB ROM, and 1.7 GHz CPU. The parameters of WISeRKNet are depicted in Table 1.
Dataset description
Lung cancer computed tomography images database (Dataset-1):
The dataset used for implementing the model is the Lung Cancer Computed Tomography Images database26for lung cancer detection. The dataset has CT images of lungs potentially categorized into different classes such as, Benign tumors, Malignant tumors and Healthy lung tissue. All images were stored using the Digital Imaging and Communications in Medicine (DICOM) standard.
Iraq-Oncology Teaching Hospital/National Center for Cancer Diseases (IQ-OTH/NCCD) Lung Cancer Dataset (Dataset-2):
The IQ-OTH/NCCD lung cancer dataset34is collected over a three-month period in fall 2019 at the Iraq Oncology Teaching Hospital and the National Center for Cancer Diseases. It consists of 1190 CT scan slices from 110 cases, including 40 malignant, 15 benign, and 55 normal subjects. CT images were obtained using a Siemens SOMATOM scanner in DICOM format with a 120 kV protocol, 1 mm slice thickness, window width ranging from 350 to 1200 HU, and window center from 50 to 600 HU, acquired during full-inspiration breath-hold. Each patient scan contains 80–200 slices, annotated by expert oncologists and radiologists. The dataset includes diverse individuals varying in age, gender, and socioeconomic background, primarily from central Iraqi provinces such as Baghdad, Wasit, Diyala, Salahuddin, and Babylon. In this study, a patient-level split is followed to prevent data leakage. For training (90%), data from 36 malignant, 13 benign, and 49 normal patients are used, while the remaining patients are reserved for testing.
In this research, the dataset is split into training and testing sets. Training data ranges from 50% to 90%, with the remaining data used for testing. For instance, when 50% of the data is used for training, 50% is used for testing, and similarly for 60–90% training. The corresponding test results are plotted in the graphs, with the X-axis representing the training data percentage.
Also, in this study, k-fold cross-validation was employed to ensure a reliable and unbiased evaluation of the WISeRKNet. The dataset was divided into k equal parts, with the value of k varying from 5 to 9. During each iteration, a single fold was used for testing and the remaining k − 1 folds for training. The process continued until all folds had been used once as the test set.
Evaluation metrics
The WISeRKNet is analyzed using different metrics, such as accuracy, TPR and TNR.
Accuracy
Accuracy represents the fraction of correctly classified instances out of the total instances in the WISeRKNet. Accuracy33is specified as,
where, specifies true positive and implies true negative, denotes false positive and expresses false negative.
TPR
TPR33measures the proportion of actual positive instances or true cases that are correctly identified by the model as positive.
Hence, TPR is defined as .
TNR
TPR33calculates the proportion of actual negative instances that are correctly identified by the model as negative.
Thus, TNR is specified as .
Precision
It is the ratio of true positives to the total number of positive predictions.
F1-score
The F1-Score merges Precision and Recall into one metric through their harmonic mean, ensuring a balance between them.
Experimental outcomes
The experimental results of images 1 and 2 obtained by the WISeRKNet is represented in Fig. 6. Figure 6a) represents the input lung CT images-1 and 2, Fig. 6b) implies pre-processed result of images-1 and 2, nodule extracted images-1 and 2 is illustrated in Fig. 6c), Fig. 6d) represents the rotated images-1 and 2 result and random erased result of images-1 and 2 is implied in Fig. 6e), Fig. 6f) implies flipped images-1 and 2, and the detected outcomes of images-1 and 2 are given in Fig. 6g).
Performance analysis for WISeRKNet
The performance analysis of the WISeRKNet model is conducted based on altering training data and K-fold value by using a CT image.
Evaluation of performance based on training data
Figure 7 provides a comprehensive evaluation of the efficacy of the WISeRKNet model by varying training data. The analysis of WISeRKNet using accuracy is depicted in Fig. 7a). The WISeRKNet obtained accuracy of 80.604%, 83.307%, 85.950%, 86.755% and 91.508% with epochs 50, 60, 70, 80 and 90, by considering training data as 90%. Figure 7b) presents the efficiency of WISeRKNet utilizing TPR. For training data = 90%, the TPR acquired by the developed WISeRKNet of 80.604%, 83.307%, 85.950%, 86.755% and 90.408% with epochs 50, 60, 70, 80 and 90. The assessment for WISeRKNet based on TNR is specified in Fig. 7c). The TNR attained by WISeRKNet of 78.597% for epochs 50, 82.394% for epochs 60, 84.731% for epochs 70, 88.902% for epochs 80 and 92.700% for epochs 90. The effectiveness of WISeRKNet using precision is seen in Fig. 7d). With epochs 50, 60, 70, 80, and 90, the precision obtained by the constructed WISeRKNet was 78.655%, 82.951%, 84.616%, 88.966%, and 90.863%. Figure 7e details the F1 score-based evaluation for WISeRKNet. For the 50th, 60th, 70th, 80th, and 90th epochs, WISeRKNet achieved a F1 score of 80.604%, 83.307%, 85.949%, 86.754%, and 90.407%, respectively.
Comparative analysis
The comparative assessment for WISeRKNet is performed using traditional models, such as ANN19, BoVW-CRNN7, Gradient boosting20, LP + ASR5, VER-Net24, and 3D-ResNet25. All comparative models were evaluated under the same experimental conditions, including the same dataset, training–testing split, preprocessing steps, and evaluation metrics. This section describes a comparative evaluation of WISeRKNet by modifying a few-shot samples, training data and k-fold.
Evaluation using dataset-1
Few shot sample generation
Few-shot samples are first generated by dividing the minimum number of samples from each class. The dataset consists of 928 samples, categorized into two classes: tumor and non-tumor. For 10% of the minimum class, a total of approximately 40 samples (20 tumor and 20 non-tumor) are selected for training. Similarly, using 15% of the minimum class, 60 samples are chosen, consisting of roughly 30 tumor and 30 non-tumor samples. For 20% of the minimum class, 80 samples are selected, with about 40 tumor and 40 non-tumor samples. Finally, when 25% of the minimum class is used, a total of 100 samples are taken, equally divided between tumor (50) and non-tumor (50) for training.
Assessment for WISeRKNet based on few shot samples
The analysis of WISeRKNet used for lung cancer detection by changing few-shot sample using dataset-1 is specified in Fig. 8. In Fig. 8a), the assessment of the WISeRKNet model with traditional models in terms of accuracy is evaluated. For 100 few shot sample, the testing accuracy acquired by WISeRKNet is 90.944% and conventional models, such as VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR measured accuracy as 84.137%, 86.144%, 79.507%, 82.134%, 85.854% and 87.902%.
Moreover, the effectiveness WISeRKNet model regarding TPR is specified as Fig. 8b). For 100 few shot sample, the TPR attained by WISeRKNet is 89.911% and the existing models, such as VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR measured TPR as 82.562%, 84.589%, 80.026%, 83.638%, 84.247% and 86.316%. Accordingly, the developed model in terms of TNR is implied as Fig. 8c). The TNR attained by WISeRKNet is 92.088% and the existing models, such as VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR acquired TNR as 83.239%, 85.449%, 78.929%, 82.342%, 84.938% and 87.193%. Figure 8d compares the precision of the WISeRKNet model to that of conventional models. Conventional models like VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, and LP + ASR tested precision as 84.666%, 82.864%, 79.846%, 82.777%, 84.555%, and 86.393% for a 100 few shot sample, but WISeRKNet obtained testing precision of 90.059%. The F1-score comparison of the WISeRKNet model with conventional models is shown in Fig. 8e). WISeRKNet achieved an F1-Score of 89.911% for a 100 few shot sample, whereas VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, and LP + ASR measured an F1-score of 84.589%, 82.562%, 80.025%, 83.638%, 84.246%, and 86.315%.
Assessment for WISeRKNet based on training data
The evaluation of the WISeRKNet model by modifying training data using dataset-1 is illustrated in Fig. 9. The accuracy assessment of the WISeRKNet model is implied in Fig. 9a). The WISeRKNet acquired an accuracy of 91.686%, while considering training data = 90%. Then, the accuracy measured by existing models, like VER-Net is 83.296%, 3D-ResNet is 85.504%, ANN is 79.471%, BoVW-CRNN is 82.127%, Gradient boosting is 84.995% and LP + ASR is 87.249% while using training data as 90%. Specifically, WISeRKNet outperforms VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR by a significant value of 9.151%, 6.743%, 13.323%, 10.426%, 7.298% and 4.840%.
In Fig. 9b), the evaluation of WISeRKNet based on TPR is presented. The TPR value acquired by WISeRKNet as 90.485%, Thus, prior models like VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR attained TPR of 83.557%, 85.317%, 80.569%, 83.760%, 85.262% and 87.058%. Figure 9c) demonstrates analysis of the WISeRKNet model concerning TNR, when training data = 90%. The WISeRKNet attained TNR as 92.727%, wherein TNR obtained by existing approaches like VER-Net is 83.298%, 3D-ResNet is 85.827%, ANN is 78.030%, BoVW-CRNN is 82.874%, Gradient boosting is 84.998% and LP + ASR is 87.578% while using training data as 90%. Analysis of the WISeRKNet model with respect to Precision is shown in Fig. 9d) when the training data is 90%. With 90% training data, the WISeRKNet achieved a Precision of 90.980%, compared to 86.913% for VER-Net, 83.216% for 3D-ResNet, 80.940% for ANN, 83.292% for BoVW-CRNN, 84.913% for gradient boosting, and 90.980% for LP + ASR. When training data reaches 90%, Fig. 9e) shows the examination of the WISeRKNet model with respect to F1 score. F1 score for the WISeRKNet was 90.484%, compared to 85.317% for VER-Net, 83.557% for 3D-ResNet, 80.569% for the ANN, 83.759% for the BoVW-CRNN, 85.262% for gradient boosting, and 87.058% for LP + ASR when 90% training data was used.
Assessment for WISeRKNet based on K-fold
The evaluation of the WISeRKNet model by modifying K-fold using dataset-1 is given in Fig. 10.
Figure 10a) demonstrates analysis of the WISeRKNet model based on accuracy, while k-fold = 9. Then, the devised WISeRKNet approach obtained an accuracy of 91.209%, whereas accuracy attained by traditional models like VER-Net is 84.084%, 3D-ResNet is 87.049%, ANN is 79.470%, BoVW-CRNN is 82.433%, Gradient boosting is 85.800% and LP + ASR is 88.825% while considering training data as 90%. The analysis of the WISeRKNet method as well as other traditional methods using TPR is specified in Fig. 10b). The devised WISeRKNet model outperforms existing methods with TPR of 90.160%, when k-fold = 9. Then, TPR acquired by conventional model, like VER-Net is 83.811%, 3D-ResNet is 84.950%, ANN is 80.222%, BoVW-CRNN is 83.575%, Gradient boosting is 85.521%and LP + ASR is 86.684%. Specifically, WISeRKNet outperforms VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR by a significant value of 7.042%, 5.779%, 11.023%, 7.304%, 5.146% and 3.856%. In Fig. 10c), the evaluation of WISeRKNet based on TNR is presented. The TNR value acquired by WISeRKNet as 92.395%, while existing models, such as VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient boosting and LP + ASR attained TNR of 83.658%, 86.094%, 79.225%, 83.549%, 85.365% and 87.851%, with k-fold = 9. Figure 10d) shows the WISeRKNet assessment based on Precision. WISeRKNet achieved a precision value of 90.979%, whereas other models, including VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, and LP + ASR, achieved TNR values of 84.812%, 83.561%, 79.617%, 82.743%, 85.266%, and 86.542%, respectively, using k-fold = 9. The F1 score-based assessment of WISeRKNet is shown in Fig. 10e. While current models like VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, and LP + ASR achieved an F1-score of 84.950%, 83.811%, 80.222%, 83.575%, 85.520%, and 86.684%, with k-fold = 9, WISeRKNet achieved an F1 score of 90.160%.
Evaluation using dataset-2
The comparative evaluation on Dataset-2 with varying training data proportions is presented in Fig. 11, where accuracy, TPR, TNR, precision, and F1-score are shown in Figs. 11(a), 11(b), 11(c), 11(d), and 11(e), respectively. When 90% of the data is used for training, the accuracy achieved by VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, LP + ASR, and the proposed WISeRKNet is 82.463%, 84.649%, 78.677%, 81.306%, 84.145%, 86.377%, and 90.769%, respectively. Similarly, the corresponding TPR values are 82.722%, 84.464%, 79.764%, 82.922%, 84.410%, 86.188%, and 89.580%, while the TNR values are 82.465%, 84.968%, 77.250%, 82.046%, 84.148%, 86.702%, and 91.800%. The precision values recorded are 86.044%, 82.383%, 80.131%, 82.460%, 84.065%, 87.800%, and 90.070%, and the F1-scores obtained are 84.464%, 82.722%, 79.764%, 82.922%, 84.410%, 86.188%, and 89.580%, respectively. These results indicate that the proposed WISeRKNet delivers superior performance across all evaluation metrics when using 90% of the training data in Dataset-2.
Ablation study
The ablation study is shown in Fig. 12, highlighting how the proposed model performs relative to its variants. The ablation study using dataset-1 is given in Fig. 12a) and dataset-2 is Fig. 12b). When the training data is 90%, the accuracy recorded by the WISeRKNet without homomorphic filtering, WISeRKNet without Link-Net, WISeRKNet without data augmentation, WISeRKNet without feature extraction, WISeR, DKN, and WISeRKNet is 85.116%, 86.258%, 87.258%, 88.258%, 89.256%, 90.249%, and 91.526% using dataset-1 and 84.264%, 85.396%, 86.386%, 87.376%, 88.363%, 89.347%, and 90.611% using dataset-2. The results clearly indicate that each component contributes to performance improvement, with the full WISeRKNet achieving the highest accuracy.
Analysis by varying noise
Figure 13 presents the performance analysis under varying noise levels, with the results for Dataset-1 shown in Fig. 13(a) and Dataset-2 displayed in Fig. 13(b). When applying 25% noise, the accuracy gained by the VER-Net is 70.579%, 3D-ResNet is 72.572%, ANN is 67.586%, BoVW-CRNN is 70.029%, Gradient boosting is 72.658%, LP + ASR is 74.589%, and proposed WISeRKNet is 76.258% using dataset-1, while considering training data as 90%. Similarly, for dataset-2, the VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, LP + ASR, and the proposed WISeRKNet is obtained the accuracy of 69.873%, 71.846%, 66.910%, 69.329%, 71.932%, 73.843%, and 75.496% at noise level 25%. Although the overall performance decreases as the noise level increases, the proposed WISeRKNet consistently outperforms all comparative models. This demonstrates its robustness and noise-handling capability in lung cancer detection.
Feature extraction analysis
The feature extraction analysis is shown in Fig. 14. Figure 14a) shows the feature extraction analysis using dataset-1 and Fig. 14b) depicts the analysis using dataset-2. Using 90% of the dataset for training, the model attained an accuracy of WISeRKNet without feature extraction, WISeRKNet with Deep features, and WISeRKNet with shape-based features is 86.258%, 88.260%, and 91.006%, using dataset-1 and 85.396%, 87.377%, and 90.095% using dataset-2. These results confirm that feature extraction, particularly shape-based features, plays a significant role in improving the detection capability of WISeRKNet.
T-Test
The performance of the WISeRKNet was statistically validated using a T-test by comparing it with VER-Net, 3D-ResNet, ANN, BoVW-CRNN, Gradient Boosting, and LP + ASR using dataset-1. The corresponding results are reported in Table 2. The T-values obtained were all higher than the critical value at a 97% confidence level, highlighting a clear performance difference between WISeRKNet and the other models. Additionally, the P-values ranging from 0.02 to 0.03 were below the 0.05 significance threshold, confirming that the improvements achieved by WISeRKNet are statistically significant.
Variability analysis
The variability analysis of the WISeRKNet and existing models in terms of mean, standard deviation (SD), variance, and confidence interval (CI) using dataset-1 is presented in Table 3. These values are derived from k-fold analysis using various evaluation metrics, with the CI calculated at a 97% confidence level. The proposed WISeRKNet demonstrates higher mean and CI values while maintaining lower SD and variance compared to existing methods.
Comparative discussion
The developed WISeRKNet was tested for lung cancer detection on two datasets, evaluated using multiple metrics, and its results were compared with state-of-the-art methods, as summarized in Table 4.With training data = 90%, the WISeRKNet achieved a high accuracy rate of 91.686%, when compared to VER-Net is 83.296%, 3D-ResNet is 85.504%, ANN is 79.471%, BoVW-CRNN is 82.127%, Gradient boosting is 84.995% and LP + ASR is 87.249%. The performances attained by WISeRKNet for TPR as 90.485%, where TPR for VER-Net is 83.557%, 3D-ResNet is 85.317%, ANN is 80.569%, BoVW-CRNN is 83.760%, Gradient boosting is 85.262% and LP + ASR is 87.058%. Furthermore, TNR acquired by the WISeRKNet model as 92.727%, when compared to VER-Net is 83.298%, 3D-ResNet is 85.827%, ANN is 78.030%, BoVW-CRNN is 82.874%, Gradient boosting is 84.998% and LP + ASR is 87.578%. The highest precision attained by WISeRKNet is 90.980%, where precision for VER-Net is 86.913%, 3D-ResNet is 83.216%, ANN is 80.940%, BoVW-CRNN is 83.292%, Gradient boosting is 84.913% and LP + ASR is 88.686%. F1-score acquired by the developed WISeRKNet model as 90.484%, when compared to VER-Net is 85.317%, 3D-ResNet is 83.557%, ANN is 80.569%, BoVW-CRNN is 83.759%, Gradient boosting is 85.262% and LP + ASR is 87.058%.
Conclusion
Conclusion
Lung cancer requires special attention because of its high impact on both men and women, resulting in a higher mortality rate. However, conventional lung cancer prediction techniques struggle with accuracy due to low-quality images that hinder the segmentation process. Therefore, an innovative model named WISeRKNet has been developed for detecting lung cancer. At first, the input CT image undergoes pre-processing by using homomorphic filtering. Subsequently, the lung nodule identification is conducted via the Link-Net model. Next, image augmentation is performed, and feature extraction is carried out by shape-based features. The final step involves detecting lung cancer by the WISeRKNet, which is the integration of WISeR and the DKN. Moreover, the developed WISeRKNet model demonstrated superior performance, by achieving improved value in accuracy as 91.686%, TPR as 90.485%, TNR as 92.727%, Precision as 90.980% and F1 score as 90.484%. However, the integration of WISeR and DKN increases model complexity and requires higher computational resources, which may limit real-time deployment in clinical settings. To address this, future work will focus on developing a lightweight architecture and employing optimization techniques to reduce computational demands and enable real-time clinical use.
Lung cancer requires special attention because of its high impact on both men and women, resulting in a higher mortality rate. However, conventional lung cancer prediction techniques struggle with accuracy due to low-quality images that hinder the segmentation process. Therefore, an innovative model named WISeRKNet has been developed for detecting lung cancer. At first, the input CT image undergoes pre-processing by using homomorphic filtering. Subsequently, the lung nodule identification is conducted via the Link-Net model. Next, image augmentation is performed, and feature extraction is carried out by shape-based features. The final step involves detecting lung cancer by the WISeRKNet, which is the integration of WISeR and the DKN. Moreover, the developed WISeRKNet model demonstrated superior performance, by achieving improved value in accuracy as 91.686%, TPR as 90.485%, TNR as 92.727%, Precision as 90.980% and F1 score as 90.484%. However, the integration of WISeR and DKN increases model complexity and requires higher computational resources, which may limit real-time deployment in clinical settings. To address this, future work will focus on developing a lightweight architecture and employing optimization techniques to reduce computational demands and enable real-time clinical use.
출처: PubMed Central (JATS). 라이선스는 원 publisher 정책을 따릅니다 — 인용 시 원문을 표기해 주세요.
🏷️ 같은 키워드 · 무료전문 — 이 논문 MeSH/keyword 기반
- Diagnostic accuracy of Ga-PSMA PET/CT versus multiparametric MRI for preoperative pelvic invasion in the patients with prostate cancer.
- Lung Cancer Screening in Adults: State-of-the-Art and Policy Mapping (2025).
- LCMS-Net: Deep Learning for Raw High Resolution Mass Spectrometry Data Applied to Forensic Cause-of-Death Screening.
- Novel silicon-based material decomposition images in diagnosis of pancreatic ductal adenocarcinoma: comparison with iodine-based and 50-keV virtual monoenergetic images.
- A novel real-world data methodology for lymphoma outcome classification: the real-world Lugano study.
- Independent Predictors Associated with Patient Refusal of Invasive Diagnostic Procedures After Positive LDCT Lung Cancer Screening.