MRI Image Linked Pixel Edge Segmentation with Least Correlated Weight Factor for Lung Tumor Identification Using Machine Learning Technique

The use of medical imaging in actual patient care has been expanding in the last several years. Radiographs of the lungs are crucial for medical diagnosis. The precision with which surgeons can remove a tumor depends on their knowledge of its precise location, volume, and size. Consequently, computer-aided design (CAD) software is crucial for managing and interpreting numerous images of lung tumors. Using a hybrid attention mechanism and parallel deep learning, Hu et al. [2] proposed a method for segmenting images of lung tumors that could handle the complexity and self-adaptation of such images. Identifying lung parenchyma was aided by early stages of image processing. As part of a hybrid attention mechanism, the pictures were subsequently fed into a DenseNet module. There are four possible approaches to the verification that include adjusting the quantity of dense blocks used in the convolution. The long-awaited realization of the perfect network layout has arrived.

A computer-aided detection system (CAD) for breast malignancies is meaningless without automated segmentation of breast ultrasound pictures. By combining a feature pyramid nonlocal network (FPNN) with transform modal ensemble learning (TMEL), Tang et al. [3] were able to effectively separate breast tumors from ultrasound pictures. The FPNN combines the nonlocal module with the feature pyramid network to fuse multilayer features while addressing long-range dependencies. The TMEL is also employed to direct two iFPNNs that extract specific tumor-related properties. Dataset-Cairo University and Dataset-Merge, two publicly available datasets, were important in making this possible.

Chronic NSCLC accounts for 84% of all lung cancer diagnosis. In order to differentiate between the two kinds of NSCLC, adenocarcinoma (ADC) and squamous cell carcinoma (SqCC), Bicakci et al. [4] conducted a thorough investigation into the application of deep learning-based classification methods. Eighteen F-FDG PET scans were administered to 94 study participants' tumors in a total of 1457 separate slices. The function of non-tumor regions in cancer subtype categorization was investigated in three experiments using positron emission tomography (PET) images. In these tests, three distinct image types were used to assess convolutional neural network (CNN) models, including the multilayer perceptron (MLP), SqueezeNet, and VGG16/VGG19. There are three types of image segments: first, those that have been cropped to include the cancer; second, those that have been randomly sliced and matched to tumors; and third, the raw slices, which have not been edited in any way. Each model was fine-tuned for diagnostic classification using a combination of optimizers and regularisation methods. We used F-score and area-under-curve (AUC) measures to evaluate the classification models' performance after training and verifying them with stratified 10-fold cross validation. The author draws the conclusion that including peritumoral areas and tissues entirely eliminates the need for manual segmentation and increases model performance based on the data.

The clinical implications of lung tumor segmentation in PET-CT images are substantial, as they facilitate precise diagnosis and treatment. This remains an uphill battle, nonetheless, in the realm of medical image processing. The lung tumor appears to change substantially between PET and CT images as a result of breathing and movement. Lung tumors in PET-CT scans appear to be of varying sizes and shapes, even though they were collected and aligned at around the same time. Despite these obstacles, a modality-specific segmentation network is still the way to go for PET-CT data processing when trying to identify lung tumors. It is possible to use MoSNet in tandem with PET and CT scans to detect lung cancer. In order to train MoSNet to recognize lung tumors in PET and CT scans, the network first learns a representation that is distinct to each modality, which shows how the two scans differ. Then, it learns a representation that is fused between modality, which shows how the two scans share features. In order to reduce the approximative modality gap, Xiang et al. [5] suggest an adversarial technique that utilizes a modality discriminator and a reserved modality-common representation in conjunction with an adversarial objective. The network's representational power is enhanced, allowing for modality-aware segmentation of lung cancer in PET and CT scans.

Tahmasebi et al. [6] created a GPU-based system that can accurately follow tumors in real time, making automated radiation therapy for cancer treatments more successful. In order to make it suitable for use in clinical settings, the authors reimplemented it on a cheap parallel GPU-based computer architecture, which significantly improved the processing performance. It is highly recommended to use GPUs for certain parts of the registration technique, such as computing the similarity metric, because they are well-suited to parallel processing. Utilizing the GPU's significantly quicker shared memory allowed for the rapid generation of the mesh deformation by speeding up the solution of a partial differential equation. Applying the suggested technique to an NVIDIA Tesla K40c GPU resulted in computational acceleration that was almost five times faster than a central processing unit.

Models based on fully convolutional networks (FCNs) and unified neural networks (U-Nets) are now considered to be the best for medical picture segmentation. Although these models work, there are two big problems with them: first, the best depth to use is unknown in advance, so finding an architecture that works takes a long time or using an ensemble of prototypes with different depths doesn't work. Second, the fusion scheme that these models use is too restrictive because of their bypassing connections. As a result, you can only aggregate maps that are the same size in the encoder and decoder sub-networks. Zhou et al. [7] offered three solutions to the issue of unknown network depth: first, a highly flexible feature fusion scheme that uses a redesign of skip connections to aggregate characteristics with different semantic scales at the decoder sub-networks; second, an efficient ensemble of U-Nets with different depths that share an encoder and co-learn simultaneously using deep supervision; and third, UNet++, an architecture that uses this redesign for semantic and instance segmentation.

In order to separate several organs and UDAs from CT images, Jiang et al. [8] created a unique hybrid network called PSIGAN, which combines probabilistic segmentation with image distribution matching. The UDA method uses a new structure discriminator to show how the images and their segmentation are dependent on each other through a joint probability distribution. The ability to automatically segment liver tumors is crucial for improving the detection and management of liver cancer. Annotations at the pixel level have been the basis for many deep learning techniques that have improved picture segmentation in recent times. Nevertheless, due to the extensive annotations that need to be obtained, deep neural networks' performance is limited when applied to medical image segmentation. When presenting data linked to liver cancer, radiologists frequently employ the Couinaud segment because of its usefulness in describing tumor location. A new method for segmenting liver tumors using convolutional networks and annotations from Couinaud segments was introduced by Lyu et al. [9] in their publication. Numbers 1 through 8 placed on images form the Couinaud segments, which each stand for a distinct section of the liver. In order to train a fully supervised tumor segmentation model, the author presented a model called Couinaud Net. This model can estimate pseudo tumour masks using only the Couinaud segment annotations as pixel-wise supervision.

Brain tumour segmentation has attracted the attention of both academics and businesses due of the hope it holds for the future of improved diagnosis and treatment. Present methods, however, frequently disregard clinical practise standards in favour of treating the problem as a simple semantic segmentation exercise. A frequent clinical practice that can uncover areas of cancer that were previously undiscovered is comparing volume data across multiple imaging modalities. After dividing the tumor into its most distinguishable piece, they also search for the other two sections. To segment brain tumors, Zhang et al. [10] first presented a new task-structured network (TS net) built on the task-task structure, and then they presented a new task-structured network (TSBTS net) built on the task-modality structure. The author developed a modality-aware feature embedding method to examine the task-modality relationship by deducing the important weights of the modality data during network training. The author performed tumor area prediction as contingently dependent sub-tasks and embedded this reliance in the network stream to examine the task-task structure.

While detection and segmentation have garnered a lot of attention in computer-assisted lung cancer diagnosis, volume estimation and grading of malignant nodules have gotten comparatively little attention. Furthermore, due to the fact that lung cancer segmentation methods are currently semi-automatic, radiologists are still required to identify the cancerous regions on each slice. This leads to longer wait times for diagnoses and makes them more subjective. The use of conventional convolution in these methods also leads to erroneous segmentation of the cancerous nodule's real boundaries. It is also important to have a system in place that can detect irregularities in screenings and assign a grade to lung cancer based on actual criteria. Enhanced performance in cancer diagnosis, segmentation, volume estimate, grading, and early warning system are described in this study work by Sathe et al. [11]. The system is fully automated and end-to-end and uses five primary models. So that the real shape of the cancerous nodule can be preserved, the traditional convolutional method is tweaked. Segmenting lymph nodes, trachea, and cancer all at once is possible with the use of a focus module and a modified loss function, and it achieves an accuracy of 92.09% while also eliminating redundancy.

One of the deadliest cancers in the world, lung cancer, has a much better prognosis when caught early. Pulmonary nodules are discovered most frequently by CT scans and are an early indicator of lung cancer. More and more medical professionals are turning to computer-aided diagnostic technologies to help them spot illnesses. Pulmonary nodule segmentation accuracy is affected by both internal heterogeneity and external data sources. Jiang et al. [12] propose a new mixed manual feature network to address the segmentation problems of subtle, mixed, adhesion-type, benign, and unclear nodule types by increasing sensitivity and accuracy. This approach incorporates feature information using a multi-dimensional fusion module and a dual-branch network structure.

The way people live their lives now has caused or exacerbated many ailments. Cancer, the deadliest of all cancers, is one of these illnesses. It is possible that early cancer identification made possible by Computer Aided Diagnosis (CAD) technology might save lives. When it comes to finding lung tumors, computed tomography (CT) scans are highly regarded. Location, form, and poor image quality from CT scans are just a few of the challenges that clinicians face when attempting to diagnose malignant cancers. Because of this, deep learning algorithms have a lot of respect among experts. A new model for tumor and nodule segmentation in CT scan images was introduced by Rehman et al. [13] using a convolutional neural network (CNN) approach. The proposed model incorporates pre- and post-processing stages to ensure precise nodule segmentation. Filtering is used in preprocessing to enhance images, and morphological operators are used in postprocessing to fine-tune segmentation. Finally, the active counter method showed that tumors and nodules could be accurately detected.

Although airway segmentation is a major pain to do manually, it is necessary for studying, diagnosing, and prognosing the progression of lung diseases. In an attempt to do away with this time-consuming and subjective human procedure, researchers have proposed methods to automatically split airways using CT images. Certain small-sized airway branches, such bronchus and terminal bronchioles, make automatic segmentation using machine learning models even more challenging. Because voxel values vary and there is a huge data imbalance in airway branches, the computational module is prone to false-negative and discontinuous predictions, especially for groups with different lung illnesses. The attention mechanism has proven capable of segmenting complex structures, and fuzzy logic can reduce uncertainty in feature representations. Since the fuzzy attention layer incorporates fuzzy theory, combining deep attention networks with it should yield a more robust and generalizable solution. In their study, Nan et al. [14] laid out a method for airway segmentation that works. With the use of a comprehensive loss function and a novel fuzzy attention neural network (FANN), it enhances the spatial continuity of airway segmentation. A deep fuzzy set is built upon a set of feature map voxels and a learnable Gaussian membership function. The proposed channel-specific fuzzy attention approach deviates from the existing attention method in order to address the issue of channels with diverse attributes.

The use of automated detection methods, namely nodule segmentation approaches, has recently become popular in lung cancer screening, thanks to developments in computing technology. These techniques distinguish between solid and soft tissues, normal and cancerous nodules, and more by means of thresholding. Recognizing nodules in close proximity to vital lung structures including blood vessels, bronchi, and the pleura is challenging and calls for more sophisticated procedures to enhance the accuracy of diagnoses. According to Gunawan et al. [15], combination processing filters should be used to prepare the data before using one of the modified Convolutional Neural Networks (CNNs) as the classifier. Solid, semi-solid, or ground glass nodule targets with improved filters can be used for low-stage cancer (cancer screening data) or high-stage cancer. Furthermore, two additional investigations addressing juxta-pleural nodules were incorporated; these investigations used three-dimensional domains for pre-processing and classification rather than the conventional two-dimensional domain.

Segmenting the lung parenchyma accurately is critical for computer-aided lung cancer diagnosis. When presented with large, clearly defined regions, existing networks for segmenting lung parenchyma perform admirably. However, when confronted with smaller, less clearly defined regions, they become extremely slow. Zhu et al. [16] proposed an improved network setup to improve segmentation performance of small and fuzzy lung parenchyma sections while maintaining accuracy for big and clear regions. U-Net has been upgraded to become the proposed network. A shape stream branch and multi-scale convolutional blocks are now part of the network. The suggested network takes computed tomography (CT) pictures as input and returns matching binary masks as output. The Open Source Imaging Consortium's Pulmonary Fibrosis Progression dataset provided the major CT images utilized in this study.

In PET images because the process of segmenting sick tissue in these images is tedious, inaccurate, and takes a lot of time. The framework designed by Guan et al. [17] began by selecting whole-body pictures that contain lesion tissue using a differential activation filter, which takes PET image features into consideration. Secondly, a novel neural network with residual connections is suggested for PET image denoising and reconstruction; this network outperforms the standard FCN network in terms of generalization. Lastly, a proprietary density-based clustering approach is employed to differentiate between normal and lesion tissues during lesion tissue segmentation. In comparison to other algorithms, the entire automated system outperforms them in terms of performance and time cost when it comes to screening PET lesion images, denoising images, and segmenting lesion tissues, as demonstrated by tests conducted on real medical PET scans. There is hope for future research and practical use of the framework.

A deterministic result, ignoring the underlying uncertainty, is generally produced by deep learning systems, notwithstanding their success in multi-modality segmentation tasks. In safety-critical clinical applications in particular, the lack of uncertainty could cause overconfidence in forecasts, which could have disastrous results. Because it provides a metric for the certainty of machine choices, uncertainty estimate has been getting a lot of interest as of late. There has been very little research into the uncertainty of multi-modality networks since most of the current uncertainty estimation methods concentrate on single-modality networks. Kang et al. [18] provided the first comprehensive look into multi-modality uncertainty as they pertain to PET/CT tumor segmentation. For this purpose, the author compared four popular uncertainty estimating methods on a number of criteria, such as segmentation accuracy, uncertainty quality, correlation to contradicting information across modalities, and comparison to uncertainties for a single modality. By combining qualitative and quantitative methods, the author was able to better understand the advantages of multi-modality uncertainty, the information they capture, and their correlation with data from individual modalities.

When working with massive labelled datasets, supervised deep learning techniques have achieved state-of-the-art performance. However, due to privacy concerns, high annotation costs, and a lack of medical professionals, medical image analysis faces challenges when trying to access big labeled datasets. In a so-called pretext task, a subfield of self-supervised learning uses unlabeled data to train a network to prioritize a legitimate latent representation of the input. After learning features in earlier tasks, the model applies them to a downstream job with less annotations. Cheke et al. [19] presented PatchLoc, a new kind of pretext task, whose goal is to use an image as a supervisor and determine where a particular patch is located. Using three separate medical datasets, we proved that PatchLoc was effective on a downstream segmentation job.

Among the many obstacles to overcome in the creation of CADs is the automated segmentation of data from multimodality positron emission tomography—computed tomography experiments. Here, methods based on CNNs are seen as cutting. But when they only look at local patterns, these CNN-based approaches struggle to learn the global context or co-learn the complementing PET-CT picture elements. Methods: For the purpose of PET-CT image segmentation, Bi et al. [20] presented a cascaded CNN-transformer network (CCNN-TN). Because TNs may build global context via self-attention and embedding picture patches. By using a cascade of TNs and CNNs, the author was able to learn both the global and local contexts, expanding the TN definition. Additionally, the author included a hyper fusion branch that merges the complementary picture features that were extracted individually in an iterative fashion. Using three datasets, one for soft tissue sarcoma (STS), one for nonsmall cell lung cancer (NSCLC), and one for our method, compared it against the state-of-the-art CNN approaches. To enhance cancer survival prediction, Cai et al. [21] presented a novel computational method called CRESCENT. It is a GCN based on prior knowledge graphs including protein-protein interactions (PPI). In order to forecast whether patients will survive, CRESCENT uses gene expression networks instead of gene expression levels.

4.1 Weighted discriminative extreme learning machine design for lung cancer detection by an electronic nose system

All sensors are treated uniformly and their data is utilized consistently when features are extracted. Features obtained from sensors that are unable to detect lung cancer biomarkers will ultimately be useless for classification, though, because not all sensors will have this capability [27]. Thus, it is commendable to use feature selection to further optimize features. To achieve this goal, the TRC algorithm is utilized to evaluate characteristics. The TRC framework is a universal graph-based feature selection tool that finds the best subset-level score rather than maximizing the scores of individual features. To compare the accuracy of feature subsets of different sizes, the WDELM classifier and the TRC method are used in leave-one-out cross validation. By combining the ideas of accuracy with G-mean optimality, we can get the five-feature optimal feature subset. In order to facilitate the subsequent construction of the sensor array, the characteristics are correlated with the WSP2110, MP901, and SP3S-AQ2-01 sensors.

The feature data and the proposed classification method are utilized to construct the discriminant model. Detailed three-dimensional principal component analysis (PCA) diagram showing all features and the features that were selected for study. (a) A full feature principal component analysis (PCA). (b) The attributes that have been specified will undergo principal component analysis (PCA). The effectiveness of the model is evaluated using the leave-one-out cross-validation. The next thing to do is arrange the samples by when they were collected. Next, the model's robustness is tested by using the remaining samples as a training set and the final data as a test set. Using the min-max technique, the data is normalized to remove disparities in feature magnitudes. In addition, the proposed classification method is tested using a gas sensor array with a drift dataset obtained from the UCI machine learning repository [28]. Batch processing is utilized due to the enormous quantity of information contained in this dataset. Class 4 has the smallest sample size overall, yet each of the six classes has its own unique sample size. For this reason, datasets with class 4 data and datasets with data from other classes are used to compare WDELM to other classical algorithms. Every dataset has its data standardized using the z-score method and subjected to tenfold cross validation. After that, for comparison, the average recognition results are displayed.

4.2 Metabolic imaging based sub-classification of lung cancer

For the purposes of this inquiry, three independent experiments were conducted using three distinct datasets. The main dataset for the initial experiment was a series of raw FDG-PET images obtained from the scans, with a resolution of 168 × 168 pixels. A different dataset, including ROIs meticulously extracted from each FDG-PET slice to include tumor and peritumoral tissue, is utilized in the second study. Instead of specifying a typical bounding box size, the boxes were individually cropped to fit the tumor. Because the tumors were not uniform in shape, the number of peritumoral areas differed among slices [29]. In order to ensure that the input photos used by deep learning models are uniform in size, a standard and appropriate picture size is defined. Since further processing wasn't needed for the initial trial, all of the photographs were the same size. We utilized datasets with a maximum subimage size of 64 × 64 pixels, which allowed us to include all tumors in each slice, for the second and third tests. Everyone in the bounding box had their pixel value set to zero in order to achieve zero padding, with the exception of the ROIs in the second experiment and the segmented sections in the last experiment. Ultimately, the datasets were input into the deep learning models, which were subsequently trained and tested.

As part of this technique, nodes are removed from the network at random during training in accordance with a fixed ratio. Finally, this work trained the models' fully-connected layers using the dropout with a ratio of 0.5. These fine-tuned convolutional neural network (CNN) designs take 64 × 64 pictures. Using 64 neurons and varying the number of hidden layers, this research aimed to compare MLP models. Deepening the structure of the MLP model improved its classification performance. Studying the optimal performance of MLP variants with3,7,11, and 15 hidden layers helped reach this objective. There is no known academic study that specifies a minimum or maximum number of neurons that should be used in MLP architectures at this time. The dropout strategy, which maximizes the number of hidden layers by employing alternative values, improves the number of neurons by randomly removing nodes from the network during training based on the 0.5 ratio.

The lung cancer is dreaded in many parts of the world, despite the fact that lung cancer has a relatively high survival rate if discovered in its early stages. The disease known as lung cancer can be fatal. The detection of cancer remains a difficult task for doctors. Unfortunately, neither the precise origin nor a comprehensive cure for cancer has been found. If detected early enough, cancer can be successfully treated. Locating areas of the lung affected by cancer requires the application of image processing techniques including noise mitigation, extraction of features, recognizing damaged regions, and maybe contrasting them with data on previous diagnoses of lung cancer.  There are two main categories of cancer cells: benign and malignant.  Malignant tumours cause the growth of several cancerous cells deep within a person's lungs, while benign tumours are easily detectable.  The key to survival is finding malignant cancer cells when they're still relatively easy to treat. The benign and damaging modules differ in terms of where cancer cells develop and how they look and feel.  In order to analyze and make sense of visual data, segmentation is a necessary first step. Several methods have been proposed in the existing literature.

When segmenting an image, each pixel is labeled with a category. In order to do any additional analysis on an image, segmentation must first be performed. Recent developments in imaging and sequencing technologies have allowed for the methodical advancement of clinical research into lung cancer. However, there is a limit to what the human brain can do in terms of effectively handling and making full use of such enormous data sets. Machine learning-based approaches play a significant role in integrating and evaluating these massive and complicated datasets, which have effectively described lung cancer from a variety of viewpoints using the acquired data.

Google Colab development in Python requires a comprehensive setup that begins with basic hardware requirements, including a web browser, a stable internet connection, and minimum 4GB RAM on the local machine, along with a Google account for accessing Google Drive. The platform offers free tier specifications including 12GB RAM, 68GB disk space, and GPU access (NVIDIA K80/T4/P4/V100, subject to availability) with a maximum session duration of 12 hours and an idle timeout of 90 minutes. The software setup involves installing essential Python libraries for data processing (numpy, pandas), visualization (matplotlib, seaborn), machine learning (sklearn, tensorflow, torch, keras), and image processing (cv2, PIL), along with proper configuration of GPU acceleration and memory management tools.

The lung cancer dataset known as the Iraq-Oncology Teaching Hospital/National Center for Cancer Diseases (IQ-OTH/NCCD) was amassed over the course of three months in the fall of 2019 in the aforementioned specialist hospitals. The dataset contains MRI scans of both healthy subjects and patients diagnosed with lung cancer at various stages. The oncologists and radiologists working at the IQ-OTH/NCCD centers annotated the slides used to create the dataset. The dataset is organized into three categories: normal, benign, and malignant. Out of these, 40 cases are classified as malignant, 15 cases as benign, and 55 cases as normal. The original format of the MRI scans was DICOM. The dataset is considered from the link https://www.kaggle.com/datasets/adityamahimkar/iqothnccd-lung-cancer-dataset. This research proposes a Linked Pixel Edge Segmentation with Least Correlated Weight Factor (LPES-LCWF) using machine learning for Lung Tumor Detection. The proposed model is compared with the traditional Weighted Discriminative Extreme Learning Machine Design for Lung Cancer Detection by an Electronic Nose System (WDEL-ENS) and Metabolic Imaging Based Sub-Classification of Lung Cancer (MIbSCLC).

Transforming an image into a set of mask or label representable pixel regions is what image segmentation is all about. If an image is segmented, then only the relevant parts of the image need to be processed, rather than the complete image. The proposed model performs segmentation on MRI images to consider the relevant portion to perform pixel extraction. The Table 1 and Figure 5 shows the Image Segmentation Accuracy Levels of the proposed and existing models and the Image Segmentation Accuracy Levels compared with Deep Learning Models are shown in Table 2 and Figure 6.

Table 1. Image segmentation accuracy levels

4-5.png

Figure 5. Image segmentation accuracy levels

4-6.png

Figure 6. Image segmentation accuracy levels compared with deep learning models

Table 2. Image segmentation accuracy levels compared with deep learning models

Image data is segmented into useful and useless parts throughout the processing phase. On the image segments, the pixel extraction is applied. In order to detect edges, the collected pixels will undergo a pixel similarity check. The pixel extraction time levels of the proposed and existing models are shown in Table 3 and Figure 7.

Table 3. Pixel extraction time levels

4-7.png

Figure 7. Pixel extraction time levels

4-8.png

Figure 8. Edge detection accuracy levels

Table 4. Edge detection accuracy levels

Table 5. Edge detection accuracy levels compared with deep learning models

Image processing algorithms are often tasked with locating the edges of objects in MRI images. The way it works is by sensing when there are abrupt changes in light levels. Edge detection is used for picture segmentation and data extraction in domains including computer vision, machine vision, and image analysis. The edge detection accuracy levels of the existing and proposed models are depicted in Table 4 and Figure 8. Table 5 and Figure 9 show the edge detection accuracy levels compared with deep learning models.

4-9.png

Figure 9. Edge detection accuracy levels compared with deep learning models

4-10.png

Figure 10. Feature extraction time levels

Table 6. Feature extraction time levels

Table 7. Feature extraction time levels compared with deep learning models

Table 8. Feature weight allocation time levels

By eliminating unnecessary information, feature extraction cleans up the dataset. In the end, data reduction aids model construction with less machine effort and boosts the pace of machine learning's learning and generalization phases. Table 6 and Figure 10 represent the feature extraction time levels of the proposed and existing models. Table 7 and Figure 11 show the feature extraction time levels compared with deep learning models.

4-11.png

Figure 11. Feature extraction time levels compared with deep learning models

4-12.png

Figure 12. Feature weight allocation time levels

The extracted features will undergo weight allocation. The weights are allocated based on the correlation factor. The weakly correlated features are allocated with highest weight. The training is performed based on the allocated weights. The Feature Weight Allocation Time Levels of the existing and proposed models are shown in Table 8 and Figure 12.

Table 9 represents the accuracy levels of least correlated weighted feature vector generation, and Table 10 represents the overall comparison of the proposed model with traditional models. The proposed model, from the weights allocated to the extracted features, generates the least correlated weighted feature set. The final feature vector is used to train the machine learning model for accurate lung cancer detection. The least correlated weighted feature vector generation accuracy levels of the existing and proposed models are shown in Table 9 and Figure 13. Table 10 presents the overall comparison of the proposed model with the traditional models.

Table 9. Least correlated weighted feature vector generation accuracy levels

Table 10. Overall Comparison

4-13.png

Figure 13. Least correlated weighted feature vector generation accuracy levels