Contrast-enhanced MRI segmentation and classification with optimized vector machine classifier

Brain MRI image processing using Wiener filtering

Preprocessing is a vital stage in medical image analysis, serving to improve the quality of MRI scans before segmentation and classification. In this study, an adaptive Wiener filter is utilized to suppress noise, such as acquisition artifacts, scanner distortions, and thermal fluctuations, while maintaining the integrity of important anatomical structures. MRI slices from axial, sagittal, and coronal orientations are processed using a $5\times 5$ neighborhood window to capture localized intensity variations. The Wiener filter adjusts its response based on regional statistics, effectively smoothing noise in uniform areas while preserving sharper transitions at tissue interfaces. This balance is essential for maintaining boundary information, which is crucial for accurate tumor delineation in subsequent stages of the pipeline.

Figure 4 provides a visual comparison illustrating the effect of Wiener filtering. The top row shows the original noisy images across the three orientations, while the bottom row displays the corresponding denoised outputs. The filtered images exhibit enhanced uniformity, clearer structural definition, and substantial noise reduction, all of which contribute to improved reliability in later processing steps. To ensure methodological transparency, the mathematical formulation of the Wiener filter and its implementation details are presented in “Wiener Filtering Equations”. These refined images form a stable foundation for the next phase of the workflow, which involves contrast enhancement.

Coherence of brain MRI images

Low contrast in brain MRI scans can obscure important structural variations, which makes tumor localization challenging. For this reason, contrast enhancement is required. After noise removal, certain regions in the MRI remain non coherent, and these inconsistencies make it more difficult to identify tumor boundaries accurately. Multiple areas within a single scan may exhibit weak coherence, and to address this issue, oriented diffusion filtering is applied as an initial step in the coherence enhancement process. This technique requires orientation information from local image regions, known as the orientation field (OF), which guides the diffusion tensor to follow the correct direction for maintaining detail consistency in the image (Soomro et al., 2021). Identifying the most coherent region within the scan is essential because it enables a clearer visualization of fine structural details that support reliable tumor detection.

To achieve a coherent representation of brain MRI images, we adopted the optimization scheme of the Anisotropic Diffusion Filter (Gottschlich & Schönlieb, 2012; Alqahtani Saeed et al., 2023). The method involves the following steps:

• 1.

  Compute the second moment matrix for each pixel in the MRI region.

• 2.

  Assign a diffusion matrix to every pixel, ensuring region specific diffusion behavior.

• 3.

  Determine the intensity variation for each pixel using the following expression: $\mathrm{\nabla }\left(D\mathrm{\nabla }L\right)$.

• 4.

  Apply the transformation to the enhanced image using the defined difference equation, producing the CLAHE output shown in Fig. 5, with its corresponding intensity distribution illustrated in Fig. 6. (1) $$.f^{t+\mathrm{\Delta }t}=f^t+\mathrm{\Delta }t\times \mathrm{\nabla }\left(D\mathrm{\nabla }f\right).$$

Contrast enhancement of brain MRI scans

To improve the visualization of localized tumor features in brain MRI images, Contrast Limited Adaptive Histogram Equalization (CLAHE) is applied. CLAHE operates on small contextual tiles rather than on the entire image, which allows local contrast enhancement while controlling the amplification of noise (Mgbejime et al., 2022). This makes it particularly suitable for highlighting tumor regions that often appear within confined and low contrast areas of brain scans. By enhancing subtle local intensity variations, CLAHE improves tissue contrast and preserves important structural details without introducing undesirable artifacts. The method adaptively adjusts contrast within each tile and then blends the processed regions in a smooth manner to generate the final enhanced output. A representative example of CLAHE applied to a sagittal CE MRI scan is presented in Fig. 7. The mathematical formulation and implementation procedure for CLAHE are provided in ‘CLAHE Algorithm and Contrast-Limiting Function’.

Once contrast is enhanced, the image is binarized using thresholding to isolate potential tumor regions.

Post-processing: segmentation of brain tumor region

Accurate segmentation is a crucial step in medical image analysis because it enables the precise isolation of tumor regions that support diagnosis and treatment planning. In this study, we employ the Fast Marching Method (FMM) due to its reliability, computational efficiency, and suitability for medical imaging tasks. The Fast Marching Method, introduced by James Sethian, provides a strong numerical framework for analyzing boundary evolution in images and signals. It is grounded in well established mathematical principles and has been widely used in applications such as image processing, pattern recognition, and surface extraction. The method also shares conceptual similarities with the level set approach, which has been applied extensively in various image analysis problems (Sethian, 1999).

In this work, the Fast Marching Method is utilized to address the challenging task of brain tumor segmentation in MRI scans. Its propagation based mechanism allows the algorithm to move through the image in a controlled manner that follows local intensity variations, which helps in identifying coherent boundaries that approximate the tumor region. Successful application of FMM requires an understanding of the speed function, propagation rules, and the behavior of the evolving front as it interacts with anatomical structures. The following sections describe the complete procedure used in this study, detailing how FMM is applied to MRI brain images to obtain accurate and consistent segmentation of tumor regions.

Mathematical formulation of FMM

The FMM models front propagation based on the Eikonal equation, formulated as:

(2) $$\left|{\mathrm{\nabla }}_{u(x)}\right|=\frac{1}{f(x)}for\to x\in \mathrm{\Omega }$$where $u(x)$ represents the arrival time of the expanding front at a point $x$ and $f(x)$ is the speed function controlling the propagation rate over the computational domain $\mathrm{\Omega }$. The initial condition is defined as:

(3) $$u(x)=0for\to x\in \mathrm{\partial }\mathrm{\Omega }$$where $\mathrm{\partial }\mathrm{\Omega }$ represents the seed region from which expansion begins.

FMM iteratively calculates the arrival times at each point in $\mathrm{\Omega }$, ensuring the front propagates smoothly until it converges at the desired boundary. This makes it highly effective in brain tumor segmentation, where the evolving interface accurately delineates the tumor based on intensity variations.

The segmentation approach integrates FMM with the Level Set Method, where a deformable contour is represented by a static level-set function. The governing equation for this evolution is:

(4) $$\frac{\mathrm{\partial }\mathrm{\Phi }}{\mathrm{\partial }t}=g_1\left(1-{\mathrm{E}}_k\right)\left|\mathrm{\nabla }\mathrm{\Phi }\right|-\beta <\mathrm{\nabla }P-\mathrm{\nabla }\mathrm{\Phi }>=0.$$

This formulation consists of three primary computational steps:

• 1.

  Speed function calculation

  • The velocity of the evolving contour is computed based on the image gradient, which determines the strength of movement:

    (5) $$g_1(x)=\frac{1}{1+\left|(G\sigma \ast I)(x)\right|p}p>1.$$

• 2.

  Smoothing for boundary enhancement

  • A Gaussian filter is applied to refine the image and enhance boundary localization. The term $g_1\left(1-{\mathrm{E}}_k\right)$ accounts for curvature effects, while $-\beta <\mathrm{\nabla }P-\mathrm{\nabla }\mathrm{\Phi }>$ directs the front toward the object boundary.

• 3.

  Contour evolution for tumor segmentation: The tumor segmentation is conducted in two steps:

  • FMM is used to evolve the contour:

    (6) $$g_1\left(1-{\mathrm{E}}_k\right)\left|\mathrm{\nabla }\mathrm{\Phi }\right|.$$

  The process iterates until the final boundary is detected:

  (7) $$\left|\mathrm{\nabla }\mathrm{\Phi }\right|=\frac{1}{g_1}.$$

  Once segmentation is complete, Eq. (4) is solved using the computed boundary conditions, ensuring precise detection of the tumor. The results are depicted in Fig. 8, illustrating the effectiveness of the proposed approach.

The following pseudo-codes outline the algorithmic steps for brain tumor segmentation using the FMM and Level Set-Based Contour Evolution.

Algorithmic implementation for brain tumor segmentation: The following pseudo-codes outline the algorithmic steps for brain tumor segmentation using the FMM (please refer to Algorithm 1) and Level Set-Based Contour Evolution (please refer to Algorithm 2).

FMM is employed in this study because of its strong capability to segment complex anatomical structures while maintaining a low computational workload. The method produces accurate contour progression by efficiently solving the Eikonal equation, which provides a well structured approach for boundary tracking without the need for reinitialization or iterative energy minimization. Traditional level set methods often encounter computational challenges, particularly when handling topological changes or weak boundary gradients. In contrast, FMM offers improved numerical stability and faster performance, making it highly suitable for images that contain heterogeneous intensity patterns, such as brain tumor regions in MRI scans. Level set based techniques may drift or overfit in areas with limited contrast, whereas FMM adapts its propagation behavior to local spatial variations, ensuring boundary consistency throughout the segmentation process. This degree of reliability, combined with its reduced processing cost, makes FMM a practical and scalable choice for clinical workflows where precision, robustness, and efficiency are essential.

SVM-based tumor classification framework

Following segmentation, the identified tumor regions must be classified to distinguish abnormal tissue from healthy areas. In this stage, SVM is used for accurate categorization. The final component of the proposed brain tumor detection pipeline employs a SVM classifier to reliably separate tumor regions from non tumor regions based on the extracted features. SVM is a well established supervised learning technique widely applied in medical image analysis because of its ability to construct an optimal separating hyperplane that maximizes class distinction and supports precise decision making (Vankdothu & Hameed, 2022).

Dataset representation and classification objective: The annotated dataset used for training is represented as:

(8) $$D=\{(x_i,y_i)\mid x_i\in X,y_i\in Y\},$$where $x_i$ denotes the feature vector derived from the preprocessed MRI scan, and $y_i$ is its corresponding binary label: $+1$ for tumor and $-1$ for non-tumor.

The goal of the SVM is to learn a classification function $f(x)$ such that:

(9) $$f(x)=y,$$with maximum margin and minimal misclassification.

Hinge loss and decision boundary: The optimization process in SVM relies on the hinge loss function:

(10) $$L(y,f(x))=max(0,1-y\cdot f(x)),$$which imposes zero loss for correctly classified points lying outside the margin and a linear penalty for those misclassified or within the margin. This loss formulation encourages maximum margin separation while minimizing classification error.

The decision function used for classification is:

(11) $$D(m)=\sum _{i=1}^N{\alpha }_i{y}_{i}K(d_i,m_i)+t,$$where ${\alpha }_i$ are Lagrange multipliers, $y_i$ are labels, $d_i$ are support vectors, $m_i$ is the test sample, $K(\cdot )$ is the kernel function, and $t$ is the bias.

Kernel Selection: This study employs the Radial Basis Function (RBF) kernel due to its capacity to model nonlinear decision boundaries, which are often present in brain tumor data. The RBF kernel is defined as:

(12) $$K(x_i,x_j)=\exp (-\gamma ||x_i-x_j|{|}^2),$$where $\gamma$ is a hyperparameter controlling the width of the Gaussian kernel. The RBF kernel demonstrated better classification performance than linear and polynomial kernels based on empirical evaluation.

Classification pipeline: The classification process consists of the following stages:

• Feature extraction: Key features that describe texture, intensity, shape, and topological characteristics are computed from the segmented and skeletonized images. A total of 100 features are extracted and organized into four major groups. Intensity features such as mean, variance, and skewness characterize pixel value distributions within the tumor region. Texture features derived from the Gray Level Co occurrence Matrix (GLCM) and Local Binary Patterns (LBP) capture spatial relationships and local variations in tissue structure. Shape descriptors, including area, eccentricity, and solidity, help summarize the geometry of the tumor. Topological features such as the Euler number and skeleton length provide information about structural complexity. These features are selected due to their established relevance in brain tumor analysis and are refined using correlation filtering to retain the most informative and non redundant variables.

• Model training: The SVM classifier is trained using the extracted features along with labeled data so that it can learn optimal decision boundaries that separate tumor from non tumor regions.

• Prediction: Once the model is trained, it is applied to new MRI scans to classify regions of interest as tumor or non tumor based on the learned decision function.

Figure 8 presents a visual example of the SVM based classification workflow, illustrating the final stage of the detection pipeline. This produces a clinically interpretable output with clearly delineated tumor regions, which supports accurate diagnostic decision making.

Algorithmic implementation The complete classification process is summarized in Algorithm 3.

SVM optimization and training strategy

To ensure reproducible and reliable classification results, a structured optimization strategy was applied to the Support Vector Machine (SVM) model. The classification stage was built around the RBF kernel, which was selected because of its strong capacity to handle non linear decision boundaries that commonly arise in feature spaces derived from brain MRI scans. Initial experimentation showed that the RBF kernel provided more stable and accurate performance compared with linear and polynomial kernels, particularly when dealing with heterogeneous feature distributions. This consistency made the RBF kernel the most suitable choice for the classification framework used in this study.

Hyperparameter Selection: Two principal hyperparameters, the regularization parameter C and kernel coefficient $\gamma$, were tuned using a grid search technique. This involved evaluating model performance over a parameter grid:

$C\in \{0.1,1,10,100\},\gamma \in \{0.001,0.01,0.1,1\}$.

Each combination of C and $\gamma$ was assessed to identify the optimal configuration that balances margin maximization and generalization.

Cross-Validation Framework: A 10-fold stratified cross-validation approach was adopted to ensure statistical robustness and class balance. The dataset was partitioned into ten equally sized subsets, maintaining consistent tumor and non-tumor ratios across folds. For each fold, the model was trained on nine partitions and validated on the remaining one. This procedure was repeated iteratively, and the final model performance was derived from the averaged results across To ensure robust evaluation and minimize bias due to dataset partitioning, we employed stratified $k$-fold cross-validation with $k=5$. This technique maintains the proportional distribution of tumor classes (glioma, meningioma, and pituitary) across all folds, ensuring balanced representation in both training and testing subsets. Each fold involved training on four partitions and testing on the remaining one, cycling through all combinations to compute average performance metrics. This stratification is particularly important given the class imbalance within the CE-MRI dataset. Performance metrics, including Dice Similarity Coefficient, Jaccard Index, accuracy, sensitivity, and specificity, were averaged over the five folds, with standard deviation reported to capture variability. This approach enhances the statistical reliability and generalizability of the reported results.

Evaluation Metrics: The grid search optimization was guided by multiple performance indicators, including classification accuracy, Dice Similarity Coefficient (DSC), sensitivity, and specificity. The parameter combination yielding the highest average performance across folds was selected as the final model configuration. This comprehensive training and tuning strategy enhances model reliability, promotes generalizability across unseen data, and ensures methodological transparency for future replication.

To ensure transparent and reproducible optimization, the hyperparameters of the SVM classifier were tuned through a structured grid-search procedure. The regularization constant C was varied across a logarithmic range from ${10}^{-3}$ to ${10}^3$, while the kernel width $\gamma$ was explored over a comparable interval to evaluate the non-linear separability of the feature space. Training and validation curves were generated to identify the region where the validation accuracy stabilizes without signs of overfitting. The optimal configuration was obtained at $C=10$ and $\gamma =0.1$, providing an effective balance between margin flexibility and generalization capability. The validation trend shown in Fig. 9 illustrates this behavior, where performance improves up to the optimal range and declines thereafter due to overfitting. This optimization strategy ensures that the final SVM model remains stable, well-calibrated, and suitable for accurate tumor classification.

Composed algorithm

The proposed algorithm introduces a comprehensive and robust strategy for detecting brain tumors by addressing core challenges in MRI analysis, including noise suppression, contrast enhancement, and accurate classification. The full workflow is presented in Fig. 10.

A major difficulty in MRI based tumor detection is the presence of noise and uneven contrast across brain regions, which can obscure important anatomical details and reduce segmentation accuracy. To overcome these issues, the algorithm incorporates a multi stage enhancement process that includes Wiener filtering for noise reduction, followed by oriented diffusion filtering and CLAHE for contrast improvement. This preprocessing sequence refines the visual quality of MRI scans and prepares them for more reliable downstream analysis.

After enhancement, the FMM is applied to perform precise tumor segmentation. This step is essential for obtaining accurate boundary delineation while preserving critical structural information, which improves the reliability of the segmentation output. The extracted tumor regions are then used to compute descriptive features that form the basis of the classification stage.

For classification, the algorithm employs the Support Vector Machine (SVM), a model widely recognized in medical imaging for its strong performance in binary decision tasks. With optimized kernel selection and parameter tuning, the SVM effectively distinguishes tumor regions from non tumor regions and reduces the occurrence of false detections.

Empirical evaluations show that the proposed framework achieves superior performance compared with conventional approaches in terms of accuracy and computational efficiency. The integrated pipeline, which combines enhancement, segmentation, and classification, not only improves diagnostic precision but also maintains fast processing times, making the method suitable for real time clinical workflows. Furthermore, the algorithm demonstrates consistent and reliable performance across different MRI scans, reinforcing its relevance in automated tumor detection.

In conclusion, the proposed system successfully addresses the limitations of existing methods by integrating advanced preprocessing, precise segmentation, and optimized classification. Its ability to improve MRI quality, maintain computational efficiency, and enhance classification accuracy highlights its value in medical image analysis and supports its potential for assisting clinical decision making.

Measuring parameters: segmentation of brain tumor

The segmentation quality is quantitatively evaluated using six metrics: skewness, kurtosis, entropy, contrast, standard deviation, and mean. These descriptors capture the statistical and spatial complexity of the segmented tumor regions, offering insights into textural heterogeneity and contrast variations. Detailed definitions and roles of these metrics are referenced from Almalki et al. (2022).

Measuring parameters: classification of brain tumor

Classification performance was measured using both training and testing datasets. To ensure robust and unbiased evaluation, k-fold cross-validation (k = 10, stratified) was applied. The classification metrics include accuracy, specificity, and sensitivity, which were computed based on confusion matrix outputs. In addition, segmentation-informed classification was evaluated using Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and Jaccard Similarity Index (JSI). All mathematical formulations of the above evaluation parameters, Confusion Matrix, Sensitivity, Specificity, Accuracy, DSC, HD, and JSI—are comprehensively explained in ‘C: Evaluation Metrics Formulations’.

Quantitative evaluation of preprocessing impact

To evaluate the effectiveness of preprocessing, various quality metrics were computed, demonstrating its substantial impact on image enhancement. CLAHE significantly increases contrast, improving CNR from 2.4 to 4.1, which helps distinguish tumor regions more effectively. The entropy of processed images also increases from 5.21 to 6.15, indicating improved information retention and enhanced visualization of fine structures. Diffusion filtering enhances SSIM from 0.68 to 0.88, ensuring smoother intensity transitions while maintaining the integrity of anatomical features. The combination of CLAHE and diffusion filtering achieves the best overall enhancement, with PSNR reaching 34.1 dB and SSIM improving to 0.91, reinforcing its effectiveness in preprocessing. Table 1 summarizes the quantitative impact of preprocessing techniques.

The results confirm that applying CLAHE and diffusion filtering significantly enhances image contrast, reduces noise, and improves structural clarity, leading to better segmentation accuracy and improved tumor detection.

Comparative analysis of Wiener filtering vs. other denoising methods

Denoising is essential to maintain the integrity of tumor structures while reducing noise-induced artifacts in MRI images. Various denoising techniques were compared, highlighting the superiority of Wiener filtering over Gaussian, median, and bilateral filtering. Gaussian filtering exhibits moderate performance in edge preservation but leads to blurring effects that reduce segmentation accuracy. Median filtering effectively retains edges but offers limited noise suppression. Bilateral filtering balances noise reduction and structure preservation, yet Wiener filtering achieves the highest PSNR (33.5 dB) and SSIM (0.89), demonstrating optimal noise suppression with minimal structural distortion. Table 2 presents the comparative performance of different denoising techniques.

The findings indicate that Wiener filtering effectively reduces noise while preserving essential details, unlike Gaussian filtering, which smooths the image but compromises critical features. Median and bilateral filtering offer moderate improvements but lack the robustness of Wiener filtering, making the latter the preferred denoising method for medical imaging applications.

Empirical justification for preprocessing sequence

To justify the specific order of preprocessing operations—namely, Wiener filtering followed by oriented diffusion filtering and finally CLAHE a comparative analysis was performed by altering the sequence of these steps. The impact of each configuration was assessed in terms of segmentation accuracy using key evaluation metrics such as DSC, (JSI, and HD). Results are summarized in Table 3.

The proposed sequence, beginning with Wiener filtering, provided the highest segmentation accuracy. Initiating preprocessing with Wiener filtering effectively reduced baseline noise while preserving anatomical structures. Subsequent application of diffusion filtering improved edge continuity, and final contrast enhancement via CLAHE helped delineate tumor boundaries more distinctly. In contrast, reversing the order (e.g., applying CLAHE first) amplified noise artifacts, which degraded segmentation quality.

Segmentation module performance assessment

The quantitative analysis of the segmentation model evaluates its capability to enhance contrast, reduce distortions, and improve tumor boundary definition. The computed statistical values, including skewness, kurtosis, entropy, contrast enhancement, and intensity variations, provide a comprehensive insight into segmentation performance.

Table 4 presents the segmentation results, showing a substantial improvement in contrast across tumor types. Notably, for pituitary tumors, the IC value of 0.030 indicates enhanced differentiation between tumor and non-tumor regions. The increased entropy values across all tumor types confirm improved information retention, essential for identifying subtle variations in tumor structures. The skewness and kurtosis values demonstrate a more uniform intensity distribution, ensuring balanced segmentation results without excessive brightness or dark regions.

The results confirm that the proposed segmentation method enhances tumor visibility by increasing contrast and entropy while maintaining stable intensity variations. This balance ensures that the model successfully captures fine details without introducing distortions, making segmentation more reliable and interpretable for clinical diagnosis.

Comparative evaluation of FMM with deep learning based segmentation models

To provide a rigorous assessment of the proposed FMM for brain tumor segmentation, a comparative evaluation was performed against two widely used deep learning architectures, U Net and a conventional Convolutional Neural Network (CNN) based segmentation model. All methods were tested on the CE MRI dataset using identical preprocessing steps, which included Wiener filtering and CLAHE based contrast enhancement, to ensure fair and consistent comparison. The quantitative results summarized in Table 5 present the performance across several evaluation metrics, including DSC, JSI, HD, SE, SP, and AC.

As anticipated, U Net achieved the highest DSC and JSI values, demonstrating its strong ability to capture complex spatial patterns and detailed tumor structures. However, this level of accuracy required extensive training, high memory consumption, and noticeably longer inference times. The CNN based model showed reduced segmentation performance, which reflects its limitations in modeling diverse tumor shapes and intensity variations without more advanced architectural components or access to large annotated datasets.

The proposed FMM based approach, although slightly lower than U Net in absolute accuracy, delivered competitive performance while maintaining minimal computational cost and requiring no training. This makes FMM particularly suitable for clinical environments where rapid inference, lower hardware requirements, and high reliability are essential. The balance between accuracy and efficiency confirms the practicality of FMM for real time or resource constrained medical imaging applications.

These findings demonstrate that the proposed FMM-based segmentation offers a compelling trade-off between accuracy and computational efficiency. While deep learning models such as U-Net may be preferable when ample labeled data and high-performance hardware are available, the deterministic nature of FMM renders it advantageous in clinical scenarios that demand real-time processing, interpretability, and scalability without reliance on large-scale model training.

Figure 11 illustrates a direct visual comparison between the proposed FMM-based segmentation output and those generated by widely used deep learning models, U-Net and ResNet, across three CE-MRI examples. Each row presents the original scan, ground truth annotation, and segmentation results produced by the three methods. The FMM approach demonstrates high alignment with the annotated masks, preserving tumor contours and volume with strong consistency. Although U-Net and ResNet perform well, some discrepancies can be observed, particularly in border precision and tumor completeness, where slight over- or under-segmentation is evident. The proposed method, despite not relying on extensive training or deep architectures, delivers localization accuracy comparable to or superior to that of reference masks while maintaining minimal deviation from them. This figure reinforces the earlier metric-based findings and visually confirms the reliability of the FMM-based pipeline in capturing diverse tumor shapes and sizes, supporting its viability for use in clinical environments with limited computational resources.

Qualitative validation of segmentation performance

To further validate the segmentation results, a visual comparison between the original MRI images, expert-annotated ground truth masks, and the segmentation outputs is presented in Fig. 12. The first column represents the original MRI scans, the second column contains expert-annotated ground truth masks, and the third column showcases the automated segmentation output generated by the proposed method.

The segmentation results align closely with ground truth annotations, confirming that the model accurately delineates tumor boundaries with minimal discrepancies. The smooth contours of segmented tumor regions further demonstrate that the method efficiently mitigates segmentation inconsistencies, such as excessive fragmentation or over-segmentation.

The qualitative results demonstrate high segmentation precision, reinforcing that the proposed method effectively isolates tumor structures while preserving important tissue details. The consistency across different tumor types highlights the model’s robustness in handling various MRI intensity distributions and complex anatomical structures.

Visual representation of classification performance

To further analyze the performance, Fig. 14 presents a bar chart illustrating the mean DSC values for each tumor type, with error bars representing standard deviation. The overall DSC of $0.963\pm 0.012$ highlights the strong segmentation accuracy of the model.

Additionally, Fig. 15 displays a boxplot of DSC scores, showing the distribution, interquartile range, and presence of any outliers. The narrow range and absence of significant outliers confirm the stability of the model’s segmentation performance.

To further validate the performance of the proposed FMM-SVM framework and address concerns regarding the high reported metrics (DSC = 0.963 and accuracy = 98.6%), a detailed evaluation using classwise confusion matrices and additional performance indicators was conducted. The confusion matrix for the three tumor classes: meningioma, glioma, and pituitary adenoma is presented in Fig. 16. This matrix provides a clear distribution of true positives, false positives, true negatives, and false negatives, allowing a transparent examination of the model’s decision outcomes.

The confusion matrix results indicate that the proposed method achieves consistently low misclassification rates across all tumor categories. Meningioma and pituitary tumors show particularly strong classification stability, while glioma cases demonstrate slightly higher confusion with neighboring classes an expected challenge due to their heterogeneous intensity profiles. To complement this visual assessment, additional performance metrics including precision, recall, F1-score, and per class DSC were also computed. These metrics further confirm that the overall accuracy and DSC are not dominated by any single class and remain consistent across tumor types.

In addition, statistical tests were performed to evaluate whether the improvements observed with the FMM-SVM model were significant compared to baseline classifiers such as KNN, SOM, and FCM. Both Wilcoxon signed-rank tests and paired t-tests (where normality conditions were satisfied) showed statistically significant improvements ( $p<0.05$) in DSC, accuracy, and specificity, confirming that the performance gains are meaningful and not the result of random variations.

Figure 16 illustrates the confusion matrix used in this evaluation.

Together, these results demonstrate that the strong performance metrics reported for the FMM–SVM framework are well-supported by detailed classification analysis, statistical validation, and consistent behavior across all tumor classes.

Comparative analysis of classification methods

To evaluate the effectiveness of the proposed FMM-SVM framework, a comparative analysis is conducted against widely used classification techniques, including Kernel-Based SVM, GCNN, GA, SOM, and KNN. The performance metrics highlight improvements in accuracy, sensitivity, and segmentation quality, reinforcing the superiority of the proposed approach. Table 7 summarizes the classification performance of these methods.

The results confirm that the FMM-SVM method outperforms traditional classifiers, achieving higher accuracy and superior segmentation quality, making it an efficient choice for brain tumor classification. To improve the robustness and interpretability of the performance evaluation, we conducted statistical analysis on key metrics across five independent runs using stratified cross-validation. The proposed FMM-SVM model achieved a Dice Similarity Coefficient (DSC) of $0.963\pm 0.007$ and a Jaccard Similarity Index (JSI) of $0.901\pm 0.009$, indicating consistent segmentation accuracy. We also evaluated standard deviation and 95% confidence intervals for classification accuracy, sensitivity, specificity, and DSC. Furthermore, the Wilcoxon signed-rank test was applied to compare the proposed method against U-Net and CNN-based models. The observed differences in DSC and accuracy were statistically significant with $p<0.05$, suggesting that the performance gains were not due to random variation. These statistical insights confirm the reliability of the proposed approach under different cross-validation splits and reinforce its applicability in clinical scenarios.

Performance comparison of classifier-based methods

A wide range of classifier based approaches have been employed for brain tumor detection, with performance typically assessed using accuracy, sensitivity, specificity, and Dice Similarity Coefficient (DSC). Table 8 summarizes the reported results of several well-known classifiers including EL-FCM, Deep LSTM, CNN, GANs, DWA-DNN, BP-NN, GCNN, GA, SOM, and KNN—together with the proposed FMM-SVM framework evaluated in this study. The comparative results indicate that the FMM-SVM model achieves superior DSC, accuracy, specificity, and sensitivity while offering a considerably lower computation time. This efficiency is particularly relevant for real-time clinical diagnostics where rapid decision-making is essential. The findings show that the FMM–SVM method achieves competitive or superior performance across all classification metrics while offering the fastest execution time among all compared models. Its lightweight computational footprint and high accuracy make it highly suitable for deployment in clinical workflows where both performance and scalability are critical. To validate whether the observed improvements were statistically meaningful, non-parametric Wilcoxon signed-rank tests were conducted against baseline classical models (KNN, SOM, and FCM), complemented by paired t-tests where normality conditions were satisfied. These statistical tests confirmed that the improvements in accuracy and DSC were significant ( $p<0.05$), reinforcing the strength of the proposed approach.

Comparison with mostly used classifiers for brain tumor detection

To further demonstrate the performance of the proposed framework, a focused comparison with recent state-of-the-art classifiers including VGG19, ResNet50, DenseNet121, and CNN–LSTM hybrids—was performed. The evaluation relied on consistent performance criteria, namely accuracy, sensitivity, specificity, and processing time. Table 9 summarizes the findings and shows that the proposed FMM–SVM method achieves the highest accuracy (98.6%) and specificity (99%), while reducing inference time to just 0.43 s per image.

The enhanced performance of the proposed method is attributed to four factors: refined preprocessing, precise FMM-based segmentation, optimized SVM classification, and a computationally lightweight architecture. Together, these components yield a robust and clinically viable framework capable of supporting real-time brain tumor detection and classification.

Comparison of processing time across different hardware

To further validate the efficiency of the FMM-SVM model, a processing time analysis was performed against deep learning-based methods across different hardware configurations. Table 10 presents the processing time comparison on both CPU and GPU platforms.

Comparison with state-of-the-art models

To validate the effectiveness of the FMM-SVM approach, a comparative assessment was conducted against MRI segmentation techniques introduced from 2019 onward. The evaluation, as presented in Table 11, highlights the notable advantages of the proposed method over existing classifiers. Table 11 presents a comparison of the proposed framework with existing approaches. The comparative results presented in Table 11 include both values reported in the literature and methods that we re-implemented and tested on the CE-MRI dataset. Reported values were taken from studies that used the same database and the same number of images, ensuring consistency in comparison. The results indicate that while CNN-based and deep learning approaches have been widely adopted for brain tumor classification, their performance remains suboptimal due to the lack of innovative processing techniques and high computational demands.

Despite advancements in deep learning architectures, models such as CapsNet, ELM, and GANs have exhibited variable performance in terms of classification accuracy and segmentation precision. One of the main limitations of these methods is their dependence on large-scale labeled datasets and computationally expensive training phases, making them impractical for real-time clinical deployment. In contrast, the FMM-SVM approach improves tumor classification by integrating enhanced preprocessing, precise segmentation, and robust classification techniques, leading to better performance while maintaining computational efficiency.

The proposed method achieves an accuracy of 98.6%, surpassing many deep learning-based classifiers, including CNN, GANs, and autoencoder models. This improvement can be attributed to contrast normalization techniques, adaptive feature extraction, and the efficient segmentation strategy of FMM, which enables precise tumor boundary delineation while reducing the risk of false classifications. Unlike traditional CNN models, which rely on extensive training and parameter tuning, the FMM-SVM framework ensures high performance even with smaller datasets, making it adaptable for real-world clinical applications.

The comparative evaluation underscores that the FMM-SVM model consistently outperforms traditional and deep learning classifiers in segmentation accuracy, classification robustness, and computational efficiency. Deep learning models often require substantial computational resources, high-end GPUs, and large labeled datasets, making them less practical for real-time applications. The proposed method achieves comparable or superior classification accuracy while maintaining low computational overhead, making it a scalable and efficient alternative for clinical implementation.

While the FMM-SVM framework demonstrates excellent classification performance, future work will focus on expanding the training dataset with multi-modal MRI scans to further enhance the model’s generalization capability. Additionally, integrating ensemble learning techniques could further improve classification robustness across a wider range of tumor types. The proposed model effectively overcomes the computational limitations of CNN-based methods, providing a lightweight, scalable, and efficient solution for automated brain tumor classification. By enhancing segmentation precision and optimizing classification boundaries, the FMM-SVM model emerges as a highly reliable alternative to conventional deep learning-based techniques, paving the way for real-time clinical deployment in brain tumor diagnosis.