Advancing multi-categorization and segmentation in brain tumors using novel efficient deep learning approaches Authors: Nadenlla RajamohanReddy, G. Muneeswari Published: 2024 ABSTRACT BackgroundA brain tumor is the development of abnormal brain cells, some of which may progress to cancer. Early identification of illnesses and development of treatment plans improve patients’ quality of life and life expectancy. Brain tumors are most commonly detected by magnetic resonance imaging (MRI) scans. The range of tumor sizes, shapes, and locations in the brain makes the existing approaches inadequate for accurate classification. Furthermore, using the current model takes a lot of time and yields results that are not as accurate. The primary goal of the suggested approach is to categorize whether a brain tumor is present, identify its type and divide the affected area into segments.MethodsTherefore, this research introduced a novel efficient DL-based extension residual structure and adaptive channel attention mechanism (ERSACA-Net) to classify the brain tumor types as pituitary, glioma, meningioma and no tumor. Extracting features in brain tumor analysis helps in accurately characterizing tumor properties, which aids in precise diagnosis, treatment planning, and monitoring of disease progression. For this purpose, we utilized Enhanced Res2Net to extract the essential features. Using the Binary Chaotic Transient Search Optimization (BCTSO) Algorithm, the most pertinent features in terms of shape, texture, and colour are chosen to minimize complexity.ResultsFinally, a novel LWIFCM_CSA approach is introduced, which is the ensemble of Local-information weighted intuitionistic Fuzzy C-means clustering algorithm (LWIFCM) and Chameleon Swarm Algorithm (CSA). Conditional Tabular Generative Adversarial Network (CTGAN) is used to tackle class imbalance problems. While differentiating from existing approaches, the proposed approach gains a greater solution. This stable improvement in accuracy highlights the suggested classifier’s strong performance and raises the possibility of more precise and trustworthy brain tumor classification. In addition, our method’s processing time, which averaged 0.11 s, was significantly faster than that of previous approaches. Introduction Brain tumors, the primary cause of brain cancer, are the most common type of brain disease. It is crucial to diagnose brain tumors early on in this type of cancer because it is lethal. It is brought on by an unnatural and uncontrollably high brain cell count (Kumar & Kumar, 2023). There are numerous classification systems for brain tumors. Brain tumors are often classified as benign or malignant, one of the most popular classification schemes (Balamurugan & Gnanamanoharan, 2023; Akter et al., 2024). Brain tissue is not where benign brain tumors originate; instead, they form on the inside of the skull. Sometimes benign brain tumors can pose a severe threat to life. Approximately 85% of benign tumors are meningiomas. Meningiomas account for roughly 33% of cases and are slow-growing tumors. Women are diagnosed with meningiomas more often than men (Deepa et al., 2023; Ullah et al., 2024). Given their low propensity to spread to nearby brain tissue, they have a good chance of being surgically removed. Even so, meningiomas can sometimes grow into cancerous tumors. The pituitary glands, which regulate hormones and physiological functions, are the site of origin of pituitary tumors. Non-cancerous pituitary tumors do not metastasize to other organs (Nanda, Barik & Bakshi, 2023). Pituitary tumors rarely result in cancer, but they can cause long-term hormone insufficiency, which can cause blindness. Malignant tumor cells are aberrant cells that proliferate erratically and uncontrollably (Farajzadeh, Sadeghzadeh & Hashemzadeh, 2023; Ullah et al., 2022; Solanki et al., 2023). These tumors can compress, invade, or destroy normal tissues. Brain tumor detection techniques fall into two categories: DL-based and machine learning (ML)-based methods (Aamir et al., 2023; Raza et al., 2022; Özkaya & Sağiroğlu, 2023). Principal component analysis, fuzzy C-means, decision trees, support vector machines, and conditional random forests are examples of machine learning techniques that are frequently applied in diverse applications (Shahin, Aly & Aly, 2023; Sobhaninia et al., 2023). While DL is capable of learning and making decisions, ML bases its decisions on its prior knowledge (Jabbar et al., 2023). As a result, DL techniques are now widely applied. This approach, which is AI-based, allows for multi-level calculations (Kumar & Kumar, 2023; Sharif et al., 2024; Allah, Sarhan & Elshennawy, 2023). The classification of Deep Learning is based on supervised and unsupervised methods (Kishanrao & Jondhale, 2023), and DL demonstrated improved performance in many medical applications (Saurav et al., 2023; Fernando & Tsokos, 2023). This study aims to create an automated model in the medical domain that is more accurate, improved, and effective in helping healthcare experts identify brain tumors at an early stage. This is because, when contrasted to conventional methods, it is becoming increasingly essential to overcome issues of time consumption, false observations, and limited expert accessibility. In comparison to the current state-of-the-art approaches, the proposed model also avoids imbalanced classes, overfitting, the need for high computational resources, and inadequate generalization on unseen data. It also exhibits improved, efficient, and more accurate results at early phases in the medical field while providing timely treatment to patients suffering from brain tumors. Existing methods have difficulty in accurately classifying brain tumors due to variations in tumor size, shape, and location. Furthermore, many of these approaches are computationally expensive and time-consuming, resulting in diagnostic delays. Moreover, datasets with class imbalances frequently produce biased predictions. Our proposed approach, which combines an efficient ERSACA-Net, Enhanced Res2Net for feature extraction, and novel optimization and segmentation techniques, specifically addresses these limitations by enhancing classification accuracy, reducing processing time, and efficiently dealing with class imbalance, making it more suitable for reliable brain tumor classification. Motivation The classification and detection of brain tumors are motivated by the critical need for an early and precise diagnosis, which significantly improves medical outcomes and survival rates. While traditional methods rely heavily on manual analysis, which is time-consuming and prone to subjectivity, recent advances in AI and deep learning provide powerful tools for automating and improving brain tumor detection from medical images. However, current AI-based methods frequently struggle to balance accuracy, computational efficiency, and generalizability across diverse datasets. To overcome these limitations, our proposed approach incorporates the ERSACA-Net approach, which distinguishes it from previous research. This method improves detection precision while minimizing computational cost, ensuring broader applicability in real-world medical settings. By incorporating this novel methodology, we hope to significantly improve the scalability, consistency, and accessibility of brain tumor detection systems, resulting in a more robust solution that better meets the needs of modern healthcare. Contribution The major key contributions of this research are as follows: Leveraging publicly accessible datasets ensures a diverse and comprehensive collection of brain tumor images, enhancing the model’s generalizability and robustness.Using CTGAN to tackle class imbalance ensures that the model is trained on a more balanced dataset, which mitigates bias and enhances the model’s ability to classify less common tumor types accurately.Utilizing the Enhanced Res2Net method for multiscale feature extraction, capturing intricate details such as shape, texture, and colour is crucial for accurate tumor classification.By selecting important features and eliminating unnecessary ones, the Binary Chaotic Transient Search Optimization (BCTSO) algorithm reduces the complexity of the experimental performance, leading to faster computation and improved model efficiency.To categorize the tumor types a novel DL based ERSACA-Net is introduced.The ensemble of Local-information Weighted Intuitionistic Fuzzy C-means (LWIFCM) and Chameleon Swarm Algorithm (CSA) enhances clustering performance by combining local information weighting with an effective optimization algorithm, improving the precision and reliability of tumor detection. Leveraging publicly accessible datasets ensures a diverse and comprehensive collection of brain tumor images, enhancing the model’s generalizability and robustness. Leveraging publicly accessible datasets ensures a diverse and comprehensive collection of brain tumor images, enhancing the model’s generalizability and robustness. Using CTGAN to tackle class imbalance ensures that the model is trained on a more balanced dataset, which mitigates bias and enhances the model’s ability to classify less common tumor types accurately. Using CTGAN to tackle class imbalance ensures that the model is trained on a more balanced dataset, which mitigates bias and enhances the model’s ability to classify less common tumor types accurately. Utilizing the Enhanced Res2Net method for multiscale feature extraction, capturing intricate details such as shape, texture, and colour is crucial for accurate tumor classification. Utilizing the Enhanced Res2Net method for multiscale feature extraction, capturing intricate details such as shape, texture, and colour is crucial for accurate tumor classification. By selecting important features and eliminating unnecessary ones, the Binary Chaotic Transient Search Optimization (BCTSO) algorithm reduces the complexity of the experimental performance, leading to faster computation and improved model efficiency. By selecting important features and eliminating unnecessary ones, the Binary Chaotic Transient Search Optimization (BCTSO) algorithm reduces the complexity of the experimental performance, leading to faster computation and improved model efficiency. To categorize the tumor types a novel DL based ERSACA-Net is introduced. To categorize the tumor types a novel DL based ERSACA-Net is introduced. The ensemble of Local-information Weighted Intuitionistic Fuzzy C-means (LWIFCM) and Chameleon Swarm Algorithm (CSA) enhances clustering performance by combining local information weighting with an effective optimization algorithm, improving the precision and reliability of tumor detection. The ensemble of Local-information Weighted Intuitionistic Fuzzy C-means (LWIFCM) and Chameleon Swarm Algorithm (CSA) enhances clustering performance by combining local information weighting with an effective optimization algorithm, improving the precision and reliability of tumor detection. Article organization: The remaining sections of the document are arranged as follows: A thorough analysis of all datasets and methodologies is provided in “Related Works”, which also covers data preprocessing, classification and segmentation techniques. The results and the additional training and validation techniques are described in detail in “Proposed Methodology”. Lastly, “Result and Discussions” contains an appendix with the conclusion. Related works This section thoroughly covers some relevant articles because of the substantial contributions that models based on deep learning can make to this field. Deep learning based classification and segmentation The enhanced fully automatic segmentation (IFAS) convolutional neural network (CNN) model is suggested by Kulshreshtha & Nagpal (2024). In IFAS, brain MRI images are segmented using a fully automatic algorithm and morphological operations. The CNN framework is employed for categorization and the U-net structure is considered for morphological segmentation in the assessment of the fully-automatic segmentation method. To precisely identify and categorize tumor cells from MRI images using the Crossover Smell Agent Optimized Multilayer Perception (CSA-MLP) was introduced by Arumugam et al. (2024). Preprocessing is done on the images to eliminate unwanted noise after they are gathered Brain tumor datasets. The images’ features are extracted to carry out the process of categorization following preprocessing. It is also possible to classify healthy and unhealthy brain cells using the CNN classifier. To reduce errors and improve the efficacy of the suggested method, the multilayer perceptron (MLP) is utilized to categorize the category. An enhanced deep learning-based framework for effective brain tumor detection is suggested by Mandle, Sahu & Gupta (2024). Preprocessing techniques are used with the compound filter, which consists of the Gaussian, mean, and median filters, to enhance the quality of brain images. The texture and intensity patterns are extracted using the GLCM based method to identify tumor areas. Utilizing the Whale Social Spider-based Optimization Algorithm (WSSOA)-based metaheuristic, the best selection of features was carried out. Lastly, the deep convolutional neural network (DCNN) was employed to precisely detect tumors. Lee, Chae & Cho (2024) presented an improved algorithm for computer-aided diagnosis tailored to classify brain tumors. Gaussian filters were used to eliminate noise from the MRI data, and Grid Mask was used to enhance the deep learning models’ generalization capabilities. Next, utilize a technique they had developed to lessen the problem of brain tumors being hidden by regular grid masks: patterned grid masks. Creating an algorithm that combines the principles of the Interval Type-II fuzzy logic system (IT2FLS) and artificial bee colony to identify the tumor region, which is surrounded by intricate brain tissues was introduced by Alagarsamy, Govindaraj & Senthilkumar (2023). The key to any successful therapeutic sequence is the oncologists’ ability to make snap decisions. The algorithm described in this article dramatically enhances decision-making by utilizing technology. Asiri et al. (2023) suggested an MRI-based brain tumor detection method that uses the SVM classifier and U-Net framework. The proposed study is predicated on enhancing and filtering MRI images to remove noise and improve contrast. To extract the area of interest, the modified U-Net architecture is utilized for MRI segmentation. The normal and tumor-affected images are classified using the support vector machine (SVM) algorithm following segmentation. Machine learning-based based classification and segmentation Classification and detection of brain tumor using Kernel based SVM was proposed by Rao & Karunakara (2022). The next step was to extract features using a combination of GLCM and SGLDM strategies. The Harris Hawks Optimization (HHO) algorithms were employed for selecting the features. Subsequently, KSVM-SSD was utilized to perform categorization. Here, the brain tumor was categorized as benign or malignant employing KSVM, and the malignant tumor was further categorized as low, medium, or high employing the social ski driver (SSD) optimization algorithm. Sekhar et al. (2021) developed a model based on TL to categorize the tumors into three types. The characteristics of the brain MRI images were extracted employing a pre-trained CNN, i.e., Google LeNet. The characteristics are subsequently categorized using classifiers like softmax, SVM, and K-Nearest Neighbor. Amin et al. (2024) suggested an unsupervised clustering strategy for tumor segmentation. Furthermore, a fused characteristic vector is employed, consisting of Gabor wavelet features (GWF), histograms of oriented gradient (HOG), local binary pattern (LBP), and segmentation-based fractal texture analysis (SFTA) characteristics. The random forest (RF) classifier was used to distinguish among three subtumoral regions: complete, enhancing, and non-enhancing tumor. Major challenges Among the problems with standard BT classification models are the following: Although machine learning has made significant progress in classifying and segmenting brain tumors, large research gaps still present opportunities for new deep learning-based methods. Managing tumor heterogeneity and the lack of high-quality annotated data are two obstacles. Integrating multimodal imaging data and creating interpretable, broadly applicable models are still challenging. In particular, when it comes to small or asymmetrical tumors, current models frequently have trouble correctly drawing the borders of the tumors. There are still issues with seamlessly integrating AI models into clinical workflows, prospectively validating them in real-world scenarios, and connecting imaging features to genomic data. It’s also crucial to manage the high computational demands of advanced models, address model biases, and ensure patient data privacy. By utilizing cutting-edge methods for data augmentation, multimodal learning, interpretability, and efficient computation, novel deep-learning approaches can close these gaps and create reliable, accurate, and clinically valuable tools for brain tumor analysis. Proposed methodology Early BT detection is essential for treatment planning and patient care. For BT to be manually classified using MRI with similar structures or appearances, the radiologist’s skill and experience in identifying and classifying BT is necessary. The main objective is to create and refine an efficient method. Initially, the raw data is fed into the preprocessing phase. The raw is noise-reduced, skull-stripped and enhanced. Then, the significant features are extracted and selected with the help of the Binary Chaotic Transient Search Optimization (BCTSO) Algorithm and Enhanced Res2Net. We introduced a novel DL-based ERSACA-Net to categorize the tumor types. Then, to segment the tumor categories, we utilized an efficient ensemble Local-information weighted intuitionistic Fuzzy C-means clustering algorithm (LWIFCM) and Chameleon Swarm Algorithm (CSA). Figure 1 shows the overall proposed framework. Overall framework of the proposed methodology. Preprocessing The following are the main actions that will be taken on the MRI images in this step to ensure that the system can read the correct input and create a better environment for image analysis: Skull removal: Since the focus of neuroimaging analysis is on brain tissue, the skull and surrounding structures are removed. Advanced image processing techniques are employed to accurately segment and extract the brain tissue from MRI scans, eliminating irrelevant background noise and artifacts. This process ensures that subsequent analysis focuses solely on the brain, improving the precision of classification and segmentation tasks. Image filtration: A median filter is applied to the MRI images to reduce noise while preserving essential features. This filter is particularly effective in enhancing the quality of MRI images by suppressing unwanted noise and highlighting important anatomical structures. As a result, this step ensures cleaner input data, which improves the reliability of the machine-learning algorithms used for brain tumor analysis. Contrast enhancement: To further improve image quality, we utilize Adaptively Clipped Contrast Limited Adaptive Histogram Equalization (ACCLAHE). This technique enhances local contrast by adaptively adjusting the contrast limits for different regions of the image. By doing so, ACCLAHE significantly increases the visibility of tumor boundaries and tissue differentiation, making the features more distinguishable for accurate classification and segmentation. This contrast enhancement is crucial for improving the overall performance of the model, especially in detecting fine details in brain tumor images. The sample preprocessed image is shown in Fig. 2. Sample preprocessed image with their histogram. Data augmentation To address the imbalanced data problem and lessen network fitting, CTGAN-based data augmentation is carried out following the preprocessing. It subtracts a few samples from majority classes and adds more samples to minority classes. It is an architecture based on GANs and intended for tabular data synthesis. Reducing the difficulties associated with using GAN structure for tabular data modelling is the main objective of CTGAN’s features. More precisely, the CTGAN architecture handles non-Gaussian and multimodal distribution by using the mode-specific normalization (MSN) that transforms constants of any distribution into a bounded vector. This is an appropriate representation for neural networks. The two neural network stages of CTGAN are a generator G and a critic C, which resembles the discriminator in a traditional GAN architecture. To manage the non-Gaussian and multimodal variation of successive fields in tabular data, CTGAN employs mode-specific normalization. An approach to the issue of imbalanced collections in continuous columns is to use sample-based training in conjunction with a conditional generator. Additionally, CTGAN uses some of the most recent developments in GAN training, such as the loss function of WGAN-GP and the critical framework of PacGAN, to improve the quality of generated data and training stability. The following equation displays the CTGAN loss function. (1) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$L = {E_{G\left( z \right)\sim{P_g}}}\left[ {D(G\left( z \right)} \right] - {E_{x\sim{P_r}}}\left[ {D\left( x \right)} \right] + \lambda {E_{y\sim{P_y}}}\left[ {{{({\nabla _y}D\left( y \right) - 1)}^2}} \right].$$\end{document}L=EG(z)∼Pg[D(G(z)]−Ex∼Pr[D(x)]+λEy∼Py[(∇yD(y)−1)2]. Here, the gradient penalty factor is represented by the symbol λ, and the sample y is continuously interpolated to the actual data x. Pr and Pg show how the generated and accurate data are distributed. The generator in CTGAN receives conditioning information (y) and random noise (z) from the tabular data (T). In addition to other characteristics of the final data, the conditioning information may define the range or type of samples in each column. After that, the discriminator (D) receives the fake samples the generator produced to reach a final judgment. Once the discriminator has established via backpropagation if the instance is genuine or fraudulent, it sends a sign to the generator. The generator uses this indication to adjust its weights and enhance its capacity to generate accurate data. Figures 3 and 4 show the distribution of categories after and before augmentation. Label distribution of proposed dataset before data augmentation. Label distribution of proposed dataset after data augmentation. Feature extraction The process of feature extraction is essential to categorization. To accurately represent images of brain tumors, we retrieve the colors, textures, and shapes that are essential in them. It is difficult to extract the best features from brain images. The feature extraction process converts unprocessed data into numerical data while preserving the original content. Features can be extracted using manual models or automated ones. While all significant features are extracted by manual feature extraction, only issue-related significant features are extracted by automated feature extraction. We extracted the features using the Enhanced Res2Net model. For brain tumor analysis, the Enhanced Res2Net approach to feature extraction provides superior multiscale representation, enhancing the accuracy and robustness of tasks related to classification and segmentation. With little computational overhead, it effectively captures both fine and coarse details. More accurate treatment planning and improved diagnostic support result from this. However, in the context of brain tumor categorization, standard deep learning models can struggle to detect the intricate details of MRI images, such as subtle differences in tumor shape, texture, and intensity, which are critical for accurate diagnosis. To capture multi-scale features and rich contextual information, we used an improved feature extraction approach called Enhanced Res2Net. This allowed us to focus on the most important aspects of the images, improving the classification process’s accuracy and robustness. While deep and transfer learning algorithms are well-known for their ability to extract characteristics from images, we chose to use Enhanced Res2Net for extracting features in our approach to address the unique challenges of brain tumor classification. Brain tumors vary significantly in size, shape, texture, and location, making it difficult for generic deep-learning models to capture all critical features effectively. Using Enhanced Res2Net, we can extract multi-scale features that provide a more complete and detailed representation of tumor characteristics. This multi-scale feature extraction captures not only the tumor’s overall structure but also the fine details required for accurate classification. Res2Net approach Res2Net performs better in terms of generalization than ResNet. The residual structure of cells of the model incorporates hierarchical small residual blocks, thereby augmenting the effective sensory field of every layer and enhancing the overall network’s extraction of features effectiveness. We chose the Res2Net model as the foundation network for the antler slice categorization task because of its ability to extract meaningful features from such images, considering the small size and difficulty of identifying deer antler slices. Enhanced Res2Net approach The feature extraction module employed the Res2Net convolutional neural network as its foundational network. The three-by-three convolution group of the multi-level residual architecture replaced the bottleneck structure’s three-by-three convolution. This expands the network’s receptive field, allowing it to gather various degrees of fine-grained scale characteristic data regarding objects. This multiscale refers to the conjunction of various receptive fields at a finer granularity rather than the combination of levels. Initially, the input map of features underwent a 1 × 1 convolution process. Then, they were split equally into s-map subsets in the channel dimension. The number of channels in these subsets was decreased to 1/s of the input channels, but they still have the same scale size. The receptive field is then increased, and all the characteristic data collected in the previous stage is contained in every 3 × 3 convolution operation. While the hierarchical residual connections within distinct residual blocks can capture fine-grained changes globally and for details, Res2Net can obtain combined data on features of different numbers and receptive field sizes. To obtain the last feature information, the output characteristics of every phase were parallelized using the concat function. The extracted features histogram is shown in Fig. 5. Histogram equalization of extracted features. Feature selection In optimization problems, feature selection algorithms built on bio-inspired metaheuristic have proven successful and significant. Numerous wrapping-based algorithms that employ binary metaheuristic algorithms have been used to solve the selecting features issue. High-accuracy feature-obtaining algorithms are essential. We employed a Binary Chaotic Transient Optimization Algorithm (BCTSO) to select the features. There are several advantages to using the BCTSO for brain tumor analysis feature selection, including improved accuracy, robustness, efficiency, and interpretability. This strategy improves model performance and guarantees scalability and adaptability to large-scale and complex medical imaging datasets by concentrating on the most pertinent features. In the end, this helps to improve the accuracy and dependability of brain tumor diagnosis and treatment planning. Chaos theory Chaotic systems unpredictably behave in deterministic systems because of their seemingly random and irregular motions. Chaotic systems are defined by their inherent, ubiquitous nature. Chaotic behaviour is not completely disordered; instead, it is deterministic. Research has revealed that their initial values strongly influence the behaviour of chaotic systems. Chaotic models are additionally stochastic, ergodic, and sensitive to initial values. Transient search optimization To optimize outcomes of searches based on the transient behaviours of switched electrical systems containing capacitors and inductances as storage elements, transient search optimization (TSO) is utilized. Equation (2) illustrates the transient and steady-state responses (final responses) of a circuit with fully reacting resistive components (R) and energy-storing components like capacitors (C), inductors (L), or both (LC). (2) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$Complete\; response = Transient\; response + Final\; response$$\end{document}Completeresponse=Transientresponse+Finalresponse A circuit with just one storing component (either RL or RC) is referred to as first-order. There are two storage components (RLCs) in a second-order circuit. No amount of switching these circuits will quickly shift the procedure toward a subsequent steady state. This allows one to compute the first-order circuit’s transient response using Eq. (3). (3) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$\displaystyle{d \over {dt}}x\left( t \right) + \displaystyle{{x\left( t \right)} \over \tau } = K$$\end{document}ddtx(t)+x(t)τ=K The solution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $x\left( t \right)$\end{document}x(t) can be found by solving the equation for differentials, as shown in Eq. (4). (4) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$x\left( t \right) = x\left( \infty \right) + \left( {x\left( 0 \right) - x\left( \infty \right)} \right){e^{\textstyle{{ - t} \over \tau }}}$$\end{document}x(t)=x(∞)+(x(0)−x(∞))e−tτ An RL circuit needs an inductor current (i(t)) to function. For an RL circuit, it is L/R, and for an RC circuit, it is RC. τ is the circuit’s time constant. The response value at the end is indicated by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $x\left( \infty \right)$\end{document}x(∞). The transient effect of the second-order circuit can be found by solving the Formula (5). (5) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$\displaystyle{{{d^2}} \over {d{t^2}}}x\left( t \right) + 2\alpha \displaystyle{d \over {dt}}x\left( t \right) + \omega _0^2x\left( t \right) = f\left( t \right)$$\end{document}d2dt2x(t)+2αddtx(t)+ω02x(t)=f(t) Equation (6) illustrates how to get the solution to the second-order differential equation. (6) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$x\left( t \right) = {e^{ - \alpha t}}\left( {{B_1}\cos \left( {2\pi {f_d}t} \right)} \right) + x\left( \infty \right)$$\end{document}x(t)=e−αt(B1cos(2πfdt))+x(∞) The damping coefficient, denoted by α, the resonant frequency, denoted by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\omega _0^2$\end{document}ω02 the damped resonant frequency, and the constants \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${B_1}$\end{document}B1 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${B_2}$\end{document}B2 are all included in this formula. There are three primary steps in TSO: (1) Assign the search agents’ upper and lower boundaries for the search area; (2) Consider options (3) Determining the best course of action. (7) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$Y = lb + rand \times \left( {ub - lb} \right)$$\end{document}Y=lb+rand×(ub−lb) TSO is exploitable since the initial order discharge decays exponentially. The random number, r1, is chosen so that the exploitation and exploration of TSO algorithms are balanced (r1 ≥ 0.5). The TSO algorithm is utilized and mathematically investigated, as shown in Eq. (7), obtained from Eqs. (5) and (7). (8) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $${Y_{l + 1}} = \left\{ {\matrix{ {Y_l^* + \left( {{Y_l} - {C_l} \cdot Y_l^*} \right){e^{ - T}}{\rm \; \; \; \; }{r_1} \; < \; 0.5} \cr {Y_l^* + {e^{ - T}}\left[ {\cos (2\pi T) + \sin (2\pi T)} \right]{Y_1} - {C_l} \cdot Y_l^*{\rm \; \; }{r_1} \ge 0.5} \cr } } \right..$$\end{document}Yl+1={Yl∗+(Yl−Cl⋅Yl∗)e−Tr1<0.5Yl∗+e−T[cos(2πT)+sin(2πT)]Y1−Cl⋅Yl∗r1≥0.5. The optimal solution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $Y_l^*$\end{document}Yl∗ of the TSO algorithm approximates the state of equilibrium \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $x\left( \infty \right)$\end{document}x(∞) of an electrical circuit (9) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$T = 2 \times z \times {r_2} - z$$\end{document}T=2×z×r2−z (10) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $${C_1} = k \times z \times {r_2} + l$$\end{document}C1=k×z×r2+l (11) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$z = 2 - 2\left( {l/{L_{max}}()} \right).$$\end{document}z=2−2(l/Lmax()). A search area’s lower and upper bounds are denoted by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $lb$\end{document}lb and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $ub$\end{document}ub, accordingly. Equation (6) involves a uniformly distributed random number, rand, and a change in z from 2 to 0. T and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${C_1}$\end{document}C1 are the random coefficients, and the random numbers \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${r_1}$\end{document}r1, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${r_2}$\end{document}r2, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${r_3}$\end{document}r3 are uniformly distributed among 0 and 1. The position of a search agent is denoted by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${Y_l}$\end{document}Yl optimal position by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $Y_l^*$\end{document}Yl∗, and iteration number by l. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $k$\end{document}k is a constant and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${L_{max}}$\end{document}Lmax has the greatest number of iterations. Additionally, the process’s exploration and exploitation balance is determined by the coefficient T, which varies between (−2, 2). While the TSO algorithm is explored when T < 0, it is exploited when T > 0. Population encoding Each individual was encoded using binary encoding. In this method, an individual is represented as a string of binary ‘0s’ and ‘1s’. The feature is said to be enabled when its value is 1, and disabled when its value is 0. The sigmoid function was used to convert continuous numbers produced by ITSO operators to binary values, as demonstrated in Eq. (13). (12) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$sigmoid\left( {x_i^d} \right) = \displaystyle{1 \over {1 + exp \left( { - x_i^d} \right)}}$$\end{document}sigmoid(xid)=11+exp(−xid) (13) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$binary\left( {x_i^d} \right) = \left\{ {\matrix{ {1\; if\; sigmoid\left( {x_i^d} \right) \;> \; {r_d}} \cr {0\; else} \cr } } \right..$$\end{document}binary(xid)={1ifsigmoid(xid)>rd0else. The range of the random number, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${r_d}$\end{document}rd, is 0 to 1. T Using our dataset, we first apply cross-validation, choosing one fold for testing and the other four for training. During the training phase, the TSO algorithm generates a feature subset. The fitness function then assesses the quality of the generated subset of features. This iterative process illustrates the systematic approach used to train and evaluate our model since it is repeated until a preset amount of iterations is reached. Objective function The suggested ITSO algorithm removes the superfluous attributes before choosing the best set of features. The algorithm uses fitness functions to complete this task. The fitness function unifies two competing goals: selecting features ratio \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $n/N$\end{document}n/N and classifier effectiveness \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $Acc$\end{document}Acc. Fitness is calculated for every individual vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${Y_l}$\end{document}Yl using the Eq. (14). (14) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$Fit = {w_1} \times Acc + {w_2} \times \displaystyle{n \over N}.$$\end{document}Fit=w1×Acc+w2×nN. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $N$\end{document}N is the overall number of features, n is the number of characteristics that have been chosen, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $Acc$\end{document}Acc denotes the classification accuracy. Accordingly, the weight coefficients allocated to the accuracy and selecting features ratio components in Eq. (14) are \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${w_1}$\end{document}w1 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${w_2}$\end{document}w2, respectively. The algorithm emphasizes accuracy while taking the significance of minimizing the set of features into account, as shown by the values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${w_1}$\end{document}w1 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${w_2}$\end{document}w2, which are set at 0.6 and 0.4. Improved transient search optimization (ITSO) This section introduces an Improved transient search optimization (ITSO) algorithm. Without a doubt, the TSO method provides a quasi-optimal answer to an optimization issue. Chaotic maps were utilized in TSO to improve results and speed up convergence. The exploration space might not be thoroughly examined because most intelligent algorithms employ a random initialization population. Chaos theory has proven useful in many areas of mathematics, including algorithmic initialization. Nonlinear deterministic bounded systems lack periodicity and convergence and are called chaotic systems. A chaotic system also depends significantly on its initial conditions and parameters. Because chaos is unpredictable, ergodic, regular, and random, it offers trustworthy randomness. Compared to random sequences produced from uniform distributions, chaotic sequences offer heuristic optimization algorithms a more efficient search strategy. In this article, a logistic map is employed to create solution sequences. (15) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $${x_{t + 1}} = r \times {x_t} \times \left( {1 - {x_t}} \right),\; t = 1,2,3,....,{t_{max}}.$$\end{document}xt+1=r×xt×(1−xt),t=1,2,3,....,tmax. A random coefficient slightly impacts the effectiveness of TSO. This problem might cause local optima to become stuck. You can go past this bottleneck by thoroughly exploring the search space. Because chaotic sequences are non-periodic, they have lower repetition values than conventional random data, which allows them to cover the entire search area. Our investigations show that chaotic maps produce this feature, and the TSO algorithm’s coefficient C1 aids in searching the search space. Classification Our suggested approach and its variations are thoroughly explained in this section. There are various advantages to categorizing brain tumors using the ERSACA-Net. By fusing adaptive attention with residual learning, ERSACA-Net improves feature extraction by directing the model’s attention to the most pertinent channels and regions within the brain tumor images. This methodology enhances the precision of differentiating between tumor types by efficiently capturing minute details and tumor morphological variations. Furthermore, the network’s capacity to adaptively weigh features guarantees strong performance in various imaging scenarios and patient demographics, which eventually contributes to more accurate and dependable tumor classification and improved diagnosis and individualized treatment planning. Firstly, the CER-Block is introduced, combining channel expansion concepts and residuals to extract image information accurately. The ER-Net, the network’s backbone for the categorization of brain tumor diseases, is built based on the CER-Block. Secondly, the design of the ACA-Block directs the network’s backbone to concentrate on tumor disease data to minimize redundant data interference. CER‑Block and ER‑Net Convolutional operations are typically used by traditional image classification networks to scale their channels, which can increase the number of parameters. Gradient information will disappear as the network gets deeper because many training settings lead to a significant computational aspects load. By triple-expanding the amount of channels without adding more parameters this aids in broadening the perceptual field. Subsequently, the data aggregation layer receives the features acquired through max-pooling, enabling the network to concentrate on leaf disease data from various angles. CER-Block is the foundation upon which ER-Net, the leading network, is built. The CER-Net comprises three CER-Blocks and two downsampling layers, as illustrated in Fig. 6. The CER-Block consists of a residual link layer and data extracted from the characteristics of images. Three maximum pooling layers with varying window sizes and a data aggregation layer comprise the image characteristic data extraction layer. A basic convolutional layer finds it challenging to convert composite characteristic data from simple to abstract because tumor diseases frequently appear as combinations of attributes like color, texture, and shape. ER-Net network structure. ACA‑Block and ERC‑Net The CBAM integrates spatial and channel attention. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $x$\end{document}x the kernel’s size in the convolution and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $y$\end{document}y be the feature map’s channel count. Next, the following represents the original Gaussian likelihood density function: (16) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$y = \displaystyle{1 \over {\sigma \sqrt {2\pi } }}{e^{ - \textstyle{{{{(x - \mu )}^2}} \over {2{\sigma ^2}}}}}.$$\end{document}y=1σ2πe−(x−μ)22σ2. The following is a comparison of the formulas used to determine the map of the featured data: (17) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$F_{BAM}^{\prime} = BN\left( {MLP\left( {Avgpool\left( F \right)} \right)} \right) + {M_s}\left( F \right)$$\end{document}FBAM′=BN(MLP(Avgpool(F)))+Ms(F) (18) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$F_{CBAM}^{\prime} = \sigma \left( {MLP\left( {Avgpool\left( F \right)} \right) + MLP\left( {Maxpool\left( F \right)} \right)} \right) + {M_s}\left( F \right)$$\end{document}FCBAM′=σ(MLP(Avgpool(F))+MLP(Maxpool(F)))+Ms(F) (19) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$F_{ACA - Block}^{\prime} = F + \sigma \left( {IGPDF(Avgpool\left( F \right) + IGPDF(Maxpool\left( F \right)} \right)) + {M_s}\left( F \right)$$\end{document}FACA−Block′=F+σ(IGPDF(Avgpool(F)+IGPDF(Maxpool(F)))+Ms(F)where the outcome of the CBAM, BAM, and ACA-Block are, in that order \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\ F^\prime_{BAM}$\end{document} FBAM′, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $F_{CBAM}^{\prime}$\end{document}FCBAM′, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $F_{ACA - Block}^{\prime}$\end{document}FACA−Block′. Batch-normalization is known as BN. Two two-dimensional convolutions combine to form the multilayer perceptron known as MLP. SPACE is represented by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${M_s}\left( F \right)$\end{document}Ms(F). F is the input data. Inverse Gaussian likelihood density function, or IGPDF for short. Average pooling is known as AvgPool. MaxPool is the maximum pooling. To increase the accuracy of the classification of brain tumors, as depicted in Fig. 7, the ACA-Block can focus on tumor features, remove unnecessary data, and further boost the information connections among the feature maps. ACA-Block framework. ERCP‑Net The output layer of the conventional image categorization network receives the prediction results. However, the semantic information obtained from the network’s lower layers is all that can be sent to this output layer. It is challenging to concentrate on the image’s pixel-level details in the interim. The enhanced output layer can produce a reliable prediction result by concentrating on semantic and pixel-level data. The ERCP-Net network’s structure is shown in Fig. 8. The framework of the ERCP-Net network. The semantics is applied to the shallow data, and more profound information about pixels is added, resulting in a combined map of features with more channels. Further semantic and pixel information refinement is achieved by downsampling and recombining the merged feature data with the characteristics data from the last CER-Block+ACA-Block structure. Feedback from the final set of information is sent to the output layer for accurate classification results. Table 1 shows all of the output tensor dimensions of the ERCP-Net layers. 10.7717/peerj-cs.2496/table-1Table 1The ERCP-Net specifications and the tensor sizes of every result layer.LayerSize of the tensorLayerSize of the tensorInput(3, 416, 416)CER_Block_3(432, 13, 13)Conv(16, 208, 208)ACA_Block_3(432, 13, 13)MaxPool(16, 104, 104)BIF_Block(1,296, 13, 13)CER_Block_1(48, 52, 52)Global average pool(1,296, 1, 1)ACA_Block_1(48, 52, 52)Fully connected layer(1,296)CER_Block_2(144, 26, 26)Softmax(38)ACA_Block_2(144, 26, 26) The ERCP-Net specifications and the tensor sizes of every result layer. Segmentation based on LWIFCM_CSA approach We introduced LWIFCM_CSA to segment the tumors. By combining the advantages of both methods, the ensemble of the Chameleon Swarm Algorithm (CSA) and the LWIFCM approach provides improved segmentation accuracy. While CSA maximizes cluster centres for better boundary delineation, LWIFCM efficiently manages image noise and intensity inhomogeneity. In complex medical imaging scenarios such as brain tumor delineation, this fusion approach improves the robustness and precision of segmentation tasks. Intuitionistic FCM algorithm By including two additional parameters, the degree of non-membership \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}γ and the degree of uncertainty are denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\pi$\end{document}π the intuitionistic fuzzy sets (IFS), a significant expansion of fuzzy sets, delicately capture the fuzzy nature of the objective world. On the set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $X$\end{document}X, the IFS \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $A$\end{document}A is specified as (20) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$A = \left\{ {{u_A}\left( x \right),{\gamma _A}\left( x \right),{\pi _A}\left( x \right)|\forall x \in X} \right\}.$$\end{document}A={uA(x),γA(x),πA(x)|∀x∈X}. In the case when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\left( x \right) \to \left[ {0,1} \right]$\end{document}(x)→[0,1], where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0 \le u\left( x \right) + \gamma A\left( x \right) \le 1$\end{document}0≤u(x)+γA(x)≤1. Furthermore, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\left( x \right) = 1 - \left( x \right) - \gamma \left( x \right)$\end{document}(x)=1−(x)−γ(x) can be used to define the degree of uncertainty. The following is a definition of intuitionistic fuzzy entropy (IFE): (21) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$IFE\left( A \right) = \mathop \sum \limits_{i = 1}^n {\pi _A}\left( {{x_i}} \right) exp (1 - {\pi _A}\left( {{x_i}} \right)).$$\end{document}IFE(A)=∑i=1nπA(xi)exp(1−πA(xi)). The degree of non-membership is written as (22) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $${\gamma _A}\left( {{x_i}} \right) = {(1 - {u_A}{({x_i})^\alpha })^{1/\alpha }},x \in X.$$\end{document}γA(xi)=(1−uA(xi)α)1/α,x∈X. While the uncertainty parameter is called \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\alpha$\end{document}α. The degree of uncertainty can then be stated as (23) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $${\pi _A}\left( {{x_i}} \right) = 1 - {u_A}\left( {{x_i}} \right) - {(1 - {u_A}{({x_i})^\alpha })^{1/\alpha }},x \in X.$$\end{document}πA(xi)=1−uA(xi)−(1−uA(xi)α)1/α,x∈X. Consequently, the IF set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $A$\end{document}A is written as follows: (24) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $${A^{IFS}} = \left[ {{u_A}\left( {{x_i}} \right),{{(1 - {u_A}{{({x_i})}^\alpha })}^{1/\alpha }},1 - {u_A}\left( {{x_i}} \right) - {{(1 - {u_A}{{({x_i})}^\alpha })}^{1/\alpha }}|x \in X} \right].$$\end{document}AIFS=[uA(xi),(1−uA(xi)α)1/α,1−uA(xi)−(1−uA(xi)α)1/α|x∈X]. Moreover, the fuzzy membership degree in the original FCM is changed to the intuitionistic fuzzy membership degree, which forms the IFCM clustering algorithm. The IFCM’s objective function, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${J^*}\left( {U,V} \right)$\end{document}J∗(U,V) is described as (25) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$J^*\left(U,V\right)=\sum \limits_i^c \sum \limits_j^n {u_{i,j}^{*m}}d^2 \left(x_j,v_i\right)+\sum \limits_{i=1}^c {\pi_i^*}exp \left(1-\pi_i^*\right).$$\end{document}J∗(U,V)=∑ic∑jnui,j∗md2(xj,vi)+∑i=1cπi∗exp(1−πi∗). The membership degree \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $U$\end{document}U updating mode in the IFCM remains unchanged, while the clustering centre \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $V$\end{document}V modifying mode is modified to (26) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$v_i^* = \left(\sum \limits_{j=1}^N u_{i,j}^{*m}x_k\right)/\sum \limits_{j=1}^N u_{i,j}^{*m}.$$\end{document}vi∗=(∑j=1Nui,j∗mxk)/∑j=1Nui,j∗m. Nevertheless, there are some disadvantages to the IFCM clustering algorithm, including its sensitivity and noise susceptibility. LWIFCM approach This section proposes the LWIFCM clustering algorithm to solve the aforementioned algorithms’ drawbacks. To avoid over-reliance on local data in low-noise regions and undervaluing the impact of the degree of membership, the local weight of information is denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${k_{ij}}$\end{document}kij is incorporated to modify the extent of the adaptive influence of local information on clustering outcomes. Conversely, the suggested clustering algorithm can fully utilize the local information factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $G$\end{document}G in high noise regions. The local information weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${k_{ij}}$\end{document}kij and the optimization function with the objective in the LWIFCM are represented as (27) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$\hat J\left( {U,V} \right) = \mathop \sum \limits_i^c \mathop \sum \limits_j^n {\hat u_{ij}}{d^2}\left( {{x_j},{v_i}} \right) + \mathop \sum \limits_{i = 1}^c \pi _i^* exp (1 - \pi _i^*) + {k_{ij}}{G_{ij}}$$\end{document}J^(U,V)=∑ic∑jnu^ijd2(xj,vi)+∑i=1cπi∗exp(1−πi∗)+kijGij (28) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $${k_{ij}} = \displaystyle{{\sigma _j^2 + \rho } \over {{{\mathop {\overline\sigma} \limits}^2} + \rho }}$$\end{document}kij=σj2+ρσ¯2+ρwhereas \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\overline {\sigma}^2}$\end{document}σ¯2 is represented as mean squared error of the sample, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\sigma _j^2$\end{document}σj2 is the variance. The terms Nj and k have definitions that align with Eq. (28). Boosted LWIFCM approach with Chameleon Swarm algorithm Two factors primarily influence the segmentation approach based on fuzzy clustering: the initial clustering centers selection and the key parameter setting. The uncertainty parameter ( \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\alpha$\end{document}α) and fuzzifier constant ( \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $m$\end{document}m) are essential parameters in the LWIFCM and are typically found through trial and error investigations. The choice of starting clustering centers significantly impacts the clustering outcomes, which are correlated with the accuracy of image segmentation. To maximize the effectiveness of clustering techniques and boost adaptability, algorithms based on swarm intelligence are typically added. This article proposes the Chameleon Swarm algorithm (CSA), a novel swarm intelligence algorithm that mimics the social behaviours of a swarm and is inspired by those behaviours. This section introduces the CSA implementation and the clustering algorithm based on the CSA (CSA-LWIFCM). The CSA initializes the population as a metaheuristic algorithm to facilitate the optimization process. Assume that there are C people in the population overall and they are in the D search space. The way the initial population is produced in the dimension and the search space is randomly initialized can be expressed as follows: (29) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $${a^i} = {L_j} + rand \times \left( {{U_j} - {L_j}} \right).$$\end{document}ai=Lj+rand×(Uj−Lj). The representation of an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $i$\end{document}i represents the ith chameleon’s initial vector. In the jth dimension, the search area’s lower and upper bounds are denoted, respectively, as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${L_j}$\end{document}Lj and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${U_j}$\end{document}Uj. The number generated at random and falling between 0 and 1 is called a rand. The improved chameleons’ search capabilities in the field of search can be stated as follows: (30) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$\rho = \delta {\ exp ^{\left( { - \alpha t/R} \right)}}.$$\end{document}ρ=δ exp(−αt/R). In this case, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\rho$\end{document}ρ is the iteration-specific parameter that decreases as the number of iterations increases. The preset parameters that are used to control the exploration and exploitation phases are δ, α, and β. The following are the revolving centred coordinates that are used to update the chameleons’ positions in the search space: (31) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$arand_r^i = m \times ac_r^i.$$\end{document}arandri=m×acri. The chameleon’s rotating centred coordinates are called \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $rand_r^i$\end{document}randri. The rotation matrix is represented by m, and the centring coordinates at the rth iteration are represented by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $ac_r^i$\end{document}acri. The iteration inertia weights are provided as follows: (32) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$W = {\left( {1 - r/R} \right)^{\left( {\lambda \sqrt {r/R} } \right)}}.$$\end{document}W=(1−r/R)(λr/R). In this case, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $W$\end{document}W it represents the inertia’s weight, and λ is the random number regulating exploitation capacity. A value of one is associated with λ. The chameleon’s rate of acceleration is expressed as (33) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$y = 2590 \times (1 - { exp ^{ - \log (r)}})$$\end{document}y=2590×(1−exp−log(r))where the chameleon’s acceleration is calculated \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $y$\end{document}y. We see that the CSA starts the optimization process and uses the equations to update the locations of the chameleons. (34) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $$a_{r + 1}^{i,j} = \left\{ {\matrix{ {a_r^{i,j} + {p_1}\left( {P_r^{i,j} - G_r^j} \right)ran{d_1} + {p_2}\left( {G_r^j - a_r^{i,j}} \right)ran{d_2}{\rm \; \; \; }ran{d_i} \ge P} \cr {a_r^{ij} + \rho \left( {{U^j} - {L^j}} \right)ran{d_3} + L_b^j{\mathop{\rm sgn}} \left( {rand - 0.5} \right){\rm \; \; \; \; \; }ran{d_i} \;< \; P} \cr } } \right.$$\end{document}ar+1i,j={ari,j+p1(Pri,j−Grj)rand1+p2(Grj−ari,j)rand2randi≥Parij+ρ(Uj−Lj)rand3+Lbjsgn(rand−0.5)randi