A Hybrid Approach for Brain Tumor Segmentation Using Fuzzy C-Means and the Imperialist Competitive Algorithm

3.1. Overview of the Proposed Framework

This study proposes a hybrid framework that integrates the FCM clustering algorithm with ICA for segmenting brain tumor MRI images. The primary objective is to improve the accuracy and stability of the clustering process by combining the soft-clustering capability of fuzzy logic with the global optimization potential of evolutionary algorithms. Because brain tumor images are characterized by uncertainty, noise, and intensity overlap between healthy and pathological tissues, hard-clustering methods often fail to yield reliable results.

The proposed framework comprises three main stages:

Preprocessing: MRI images are preprocessed and mapped to an appropriate feature space to enhance features relevant to the tumor region.

FCM clustering: The FCM algorithm is used to partition pixels into clusters by assigning fuzzy membership degrees.

Optimization via ICA: ICA acts as an evolutionary optimization layer to determine the optimal initial cluster centers, thereby improving the convergence behavior of FCM.

This hybridization mitigates the risk of entrapment in local minima and improves final segmentation accuracy.

3.2. Data Acquisition and Preprocessing

The study used MRI sequences from the BraTS 2024 benchmark dataset (21). Because MRI scans are affected by systemic noise, intensity inhomogeneity, and inter-subject variability, preprocessing is essential to ensure data quality and clustering performance.

Initially, input images were converted to grayscale, when necessary, and pixel intensities were normalized to mitigate amplitude variation. Intensity normalization scales pixel values within a specific range, ensuring stability during the fuzzy clustering process. After normalization, each image was mapped to a set of pixels, with each pixel represented as a feature vector. In the simplest configuration, the feature vector for a pixel consists of its gray-level intensity. Thus, each pixel was treated as a sample in a one-dimensional feature space. This vector representation enables direct application of the fuzzy clustering algorithm. Although spatial or local statistical features can be appended to the vector to enhance differentiation, this study primarily focused on intensity features to demonstrate the efficacy of the proposed initialization method. The extracted feature vectors served as input for the subsequent FCM algorithm.

3.3. Fuzzy C-Means Clustering Algorithm

Fuzzy C-Means is a widely established method in medical image analysis that models inherent data uncertainty through membership degrees. In brain tumor MRI, boundaries between healthy tissue and tumor are often indistinct because of the partial volume effect; therefore, fuzzy clustering provides more flexible and accurate results than hard clustering.

Let X={x₁, x₂, …, x_N} be a set of N pixels, where the goal is to partition these pixels into C clusters. Unlike hard clustering, FCM allows each pixel to belong to multiple clusters simultaneously. The objective function of FCM is defined as follows:

$$J_m=\sum _{i=1}^N\sum _{j=1}^C{u}_{\mathrm{ij}}^m\parallel x_i-v_j{\parallel }^2$$

Where:

x_i is the feature vector of the i-th pixel.

v_j is the center of the j-th cluster.

u_ij represents the degree of membership of pixel i in cluster j.

m>1 is the fuzzification parameter (weighting exponent) that controls the softness of clustering.

To minimize the objective function, the membership matrix and cluster centers are iteratively updated using the following equations:

$$u_{\mathrm{ij}}={\left[\sum _{k=1}^C{\left(\frac{\parallel x_i-v_j\parallel }{\parallel x_i-v_k\parallel }\right)}^{\frac{2}{m-1}}\right]}^{-1}$$

$$v_j=\frac{\sum _{i=1}^N{u}_{\mathrm{ij}}^m{x}_i}{\sum _{i=1}^N{u}_{\mathrm{ij}}^m}$$

This process is repeated until a stopping criterion, such as the maximum number of iterations or a minimal change in cluster centers, is met. Despite its utility, FCM is highly sensitive to the random initialization of cluster centers because of the non-convex nature of its objective function. This sensitivity often leads to convergence at local minima, thereby reducing segmentation accuracy in complex MRI images.

3.4. Optimization via the Imperialist Competitive Algorithm

The Imperialist Competitive Algorithm is a population-based optimization method inspired by sociopolitical processes such as imperialistic competition, assimilation, and the collapse of empires. ICA uses evolutionary mechanisms to search the solution space effectively and has demonstrated robustness in solving nonlinear, complex, and multidimensional problems.

In ICA, each candidate solution is modeled as a country. The population is divided into two groups: imperialists, which are the countries with the lowest cost or best fitness, and colonies. Empires are formed by assigning colonies to imperialists according to the power of the imperialists. The main evolutionary steps include the following:

Assimilation: Colonies move toward their corresponding imperialist, improving their position in the search space.

Competition: Imperialistic competition occurs as weaker empires gradually lose their colonies to stronger empires.

Elimination: Empires that lose all colonies eventually collapse.

This continuous competition drives the population toward promising regions of the search space, substantially increasing the probability of finding the global optimum. In this study, ICA was used to overcome the initialization sensitivity of FCM. By intelligently searching for the optimal initial cluster centers that minimize the fuzzy objective function, ICA ensures robust convergence and stability.

3.5. Proposed Hybrid FCM-ICA Framework

To address the limitations of conventional FCM, particularly its dependence on initial values, this study introduces a hybrid FCM-ICA framework. The core concept is to use the global search capability of ICA to determine the optimal initial cluster centers (v_j) before executing the standard FCM iterations.

Encoding strategy:

In the proposed framework, each country in the ICA population represents a candidate solution for the cluster centers. Mathematically, a country is defined as a vector containing all cluster centers:

Country =[v₁,v₂, …, v_C]

where each element represents a cluster center in the feature space, and the number of elements depends on the number of clusters.

Fitness function:

The quality of each country is evaluated using the FCM objective function (J_m). The goal of the ICA is to minimize this value. Countries that yield a lower J_m are considered more powerful and are more likely to be selected as imperialists.

Execution flow:

After the population is initialized, the ICA process, including assimilation, competition, and elimination, iterates until a stopping criterion is met. The best country found by ICA, or the global best, is then extracted as the set of optimal initial cluster centers. Finally, the standard FCM algorithm is initialized with these optimized centers and run to fine-tune membership degrees and produce the final segmentation. This hybrid approach ensures that FCM starts from a near-optimal point, avoids local minima, and improves segmentation stability, particularly in noisy MRI data.

The procedural steps of the proposed method are detailed in Algorithm 1, and the overall process flow is illustrated in Figure 1.

Figure 1.

Overall process flow of the hybrid FCM-ICA clustering method for brain tumor segmentation.

3.1. Algorithm 1. Hybrid FCM-ICA Clustering for Brain Tumor Segmentation

Input:

MRI image dataset

Number of clusters C

Fuzzification parameter m

Maximum number of ICA iterations T_ICA

Maximum number of FCM iterations T_FCM

Output:

Final segmented image

1. Preprocess the MRI image and extract feature vectors X_i.

2. Initialize a population of countries in ICA.

3. Represent each country as a set of cluster centers.

4. Country =[v₁,v₂, …, v_C]

5. For each country, compute the fitness value using the FCM objective function J_m.

6. Form initial empires based on country fitness values.

7. While the ICA stopping criterion is not satisfied:

7.1. Assimilate colonies toward their imperialist.

7.2. Perform imperialistic competition.

7.3. Eliminate weak empires.

7.4. Update the fitness values of all countries.

8. end while

9. Select the best country as the optimal initial cluster centers.

8. Initialize FCM using the optimized cluster centers.

9.11. for t=1t = 1t=1 to TFCMT_{FCM}TFCM do

9.1. Update membership values U_ij.

9.2. Update cluster centers V_j.

10. Generate the final segmented MRI image.

3.6. Implementation and Computational Environment

The proposed algorithms were implemented in Python 3.9 using the NumPy and Scikit-learn libraries. All experiments were conducted on a workstation equipped with an Intel Core i7 processor and 16 GB of RAM. Statistical analyses and performance evaluations were performed in the same computational environment to ensure consistency and reproducibility.

4.1. Experimental Setup and Dataset

To evaluate the performance of the proposed hybrid framework, experiments were conducted using the BraTS 2024 benchmark dataset (21), a reference standard for brain tumor analysis. This dataset comprises MRI scans annotated by expert radiologists, which served as the ground truth for validating segmentation algorithms. The study specifically focused on images containing active tumor regions, as accurate delineation of these areas is critical for clinical diagnosis and treatment planning. The selected images exhibited substantial heterogeneity in intensity and tissue structure and contained noise and ambiguous boundaries, which are common challenges in medical image segmentation. Before clustering, preprocessing was applied to improve algorithmic stability. This included intensity normalization to scale pixel values to the required range and the application of smoothing filters to suppress noise while preserving structural details. Representative images from the BraTS 2024 dataset illustrating these challenges are presented in Figure 2.

Figure 2.

Sample images from the BraTS 2024 dataset.

4.2. Performance Metrics

For quantitative evaluation, the segmentation results were compared pixel by pixel with the ground truth. Performance was assessed using five standard metrics: Accuracy, Sensitivity (Recall), Specificity, Precision, and the Dice Similarity Coefficient (DSC). These metrics are defined as follows:

Accuracy: Measures the overall correctness of the segmentation.

$$\mathrm{Accuracy}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}$$

Sensitivity (Recall): Measures the proportion of actual tumor pixels correctly identified.

$$\mathrm{Sensitivity}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$

Specificity: Measures the proportion of non-tumor pixels correctly identified.

$$\mathrm{Specificity}=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}$$

Precision: Measures the proportion of identified tumor pixels that are actually tumor.

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

Dice Coefficient (DSC): Measures the overlap between the predicted segmentation and the ground truth.

$$\mathrm{DSC}=\frac{2\mathrm{TP}}{2\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}$$

Where true positive, true negative, false positive, and false negative represent the pixel-level classification counts.

4.3. Parameter Sensitivity Analysis

The performance of the proposed FCM-ICA framework is influenced by two key parameters: the number of clusters in the FCM component and the number of countries (N _country) in the ICA optimization layer. Extensive experiments were conducted to determine the optimal configuration.

Effect of cluster number:

As shown in Table 1, setting the number of clusters to two yielded suboptimal results, with accuracy values of approximately 25% to 32%. This poor performance is attributed to substantial overlap in gray-level intensities between tumor and healthy tissues. A binary clustering approach fails to distinguish tumor regions from other high-intensity brain structures. Increasing the number of clusters to three resulted in a substantial improvement in accuracy, up to 88.1%, because it enabled better separation of the tumor, healthy tissue, and background.

Effect of country count (N _country):

The number of countries in ICA affects convergence behavior.

$N_{\mathrm{country}}=30$: Resulted in premature convergence, indicating an insufficient population to effectively explore the search space. $N_{\mathrm{country}}=70$: Led to increased computational complexity and divergence, reducing stability. $N_{\mathrm{country}}=50$: Provided the optimal balance, achieving the highest segmentation accuracy. Consequently, $C=3$ and $N_{\mathrm{country}}=50$ were selected as the optimal parameters for the proposed framework.

4.4. Comparative Analysis

To validate the efficacy of the proposed method, its performance was compared with three established algorithms: K-means, standard FCM, and Kernel-based Intuitive Fuzzy C-Means (KIFCM). For fairness, all comparative clustering methods, including K-means, FCM, and KIFCM, were executed using the same number of clusters (C = 3). The comparison was based on Precision, Recall, and Accuracy, as summarized in Table 2 and visualized in Figure 3.

Figure 3.

Bar chart comparing the performance metrics of K-means, FCM, KIFCM, and FCM-ICA.