Enhanced Brain Tumor Classification Using Sequential Grey Wolf Optimization and Multi-Stage Ensemble Learning with Pre-Trained Convolutional Neural Networks

Brain tumors represent one of the most challenging conditions in modern neurology, characterized by abnormal cellular proliferation within the cranial cavity that poses a severe threat to human life. Early and precise diagnosis is paramount for effective treatment planning, surgical intervention timing, and overall patient survival outcomes [1]. Magnetic Resonance Imaging (MRI) serves as the gold standard diagnostic modality, providing detailed soft tissue contrast and anatomical visualization crucial for tumor detection and characterization. However, manual interpretation of vast volumes of MRI scans by radiologists is inherently time-consuming, subjective, and prone to inter-observer variability, potentially leading to diagnostic inconsistencies and delayed treatment decisions [2]. The increasing volume of medical imaging data further compounds these challenges, necessitating the development of automated, reliable, and efficient diagnostic systems. Deep learning has significantly advanced medical image analysis, particularly in automated classification tasks, due to its capability to learn complex hierarchical features directly from raw imaging data [3]. Convolutional neural networks (CNNs) have demonstrated strong performance across diverse medical imaging applications, including brain tumor detection, classification, and segmentation [4]. Despite their individual success, single CNN architectures often suffer from limitations, including sensitivity to initialization parameters, overfitting tendencies, dataset-specific performance variations, and suboptimal generalization across diverse clinical scenarios. Ensemble learning methodologies offer a robust solution to these limitations by combining the predictions of multiple diverse models, thereby improving predictive accuracy, enhancing generalization, and reducing model variance [5]. Meta-heuristic optimization algorithms, particularly the Grey Wolf Optimization (GWO) [6], have emerged as effective tools for solving complex optimization challenges in machine learning pipelines. GWO, inspired by the hierarchical hunting behavior of grey wolves, has demonstrated good performance in feature selection, hyperparameter optimization, and ensemble configuration tasks [6]. This research introduces a brain tumor classification framework that integrates a multi-stage ensemble of pretrained CNNs with a comprehensive GWO-based optimization strategy, the architecture of which is depicted in Figure 1.

Figure 1. Architecture of the proposed sequential ensemble with GWO optimization

1.1 Main contributions

(1) Development of a stacking ensemble framework employing three pre-trained CNN architectures (EfficientNetB4, Xception, ResNet50V2) with individual accuracies of 98.53%–99.35%, demonstrating the effectiveness of transfer learning for brain tumor classification on the Figshare benchmark dataset.

(2) Implementation of sequential GWO-based optimization through a two-stage optimization pipeline utilizing GWO for intelligent feature selection (reducing features by 80% from 1,542 to 297) and hyperparameter tuning, resulting in a final ensemble accuracy of 99.84%.

(3) Advanced preprocessing and training strategies with comprehensive image preprocessing pipelines and training methodologies to maximize model performance and robustness across diverse tumor types, with architecture-specific preprocessing optimized for each CNN model.

(4) Achievement of strong per-class performance with pituitary and meningioma tumors achieving 100% accuracy, and glioma achieving 99.30% accuracy on the test set.

(5) Comprehensive experimental validation through rigorous assessment using multiple performance metrics, statistical analysis, and detailed evaluation of model behavior, training dynamics, and computational efficiency suitable for research environments.

(6) Experimental evaluation on the Figshare benchmark dataset demonstrating 99.84% overall accuracy, with improvements of 0.17% over the best reported ensemble result and 1.14%–5.10% over other methodological categories on the same dataset.

(7) Resource characterization with a total training time of approximately 2.5 hours on an NVIDIA RTX 4090, reported to support reproducibility rather than as an efficiency claim.

The proposed sequential optimization approach addresses feature selection and hyperparameter optimization in coordinated, non-overlapping stages rather than through simultaneous joint optimization. The key distinction from simply executing two consecutive steps is that each stage operates on the output of the previous one using a fixed, validated intermediate representation: GWO-selected features are locked before the hyperparameter search begins, preventing the two optimization objectives from interfering with each other. This staged design reduces conflicting gradients, facilitates convergence to stable solutions, and allows each stage to be tuned and evaluated independently. The feature selection stage identifies 297 discriminative features from the original 1,542-dimensional meta-feature space, and the subsequent hyperparameter stage optimizes the Random Forest meta-classifier configuration (n_estimators=106, max_depth=10, min_samples_split=9, min_samples_leaf=7) on this compact representation.

The remainder of this study is organized as follows: Section 2 reviews related work in brain tumor classification. Section 3 details the proposed GWO-optimized ensemble methodology. Section 4 presents the experimental setup and results. Section 5 discusses findings, limitations, and future directions. Section 6 concludes the study.

This section presents the comprehensive methodology for our brain tumor classification framework. The proposed approach consists of four main components: enhanced data preprocessing, three-model ensemble design, a two-stage GWO-based optimization, and performance evaluation, as outlined in Figure 1.

3.1 Dataset description and analysis

The dataset utilized in this study is the publicly available Figshare brain tumor MRI dataset [16], comprising 3,064 T1-weighted contrast-enhanced MRI images. The dataset encompasses three primary brain tumor classes representing distinct pathological entities: glioma (malignant tumors with invasive characteristics), meningioma (typically benign tumors arising from meningeal tissues), and pituitary adenomas (hormone-related tumors of the pituitary gland). This dataset has become a standard benchmark in medical imaging research due to its high-quality annotations and balanced representation of major brain tumor types encountered in clinical practice.

3.2 Evaluation methodology and data splitting

To ensure robust model training and evaluation, a stratified evaluation protocol was implemented with clear separation between training and testing phases. The dataset was divided into 80% for training (2,451 images) and 20% for testing (613 images), with stratification ensuring proportional representation of all tumor classes, as detailed in Table 1. A fixed random seed (42) was applied consistently across all stochastic operations, including data partitioning, model weight initialization, augmentation sampling, and GWO population initialization, to guarantee full reproducibility of the reported results. It should be noted that, as the publicly available Figshare dataset does not provide patient identifiers, partitioning was performed at the image level rather than the patient level. This is consistent with the majority of prior studies using the same dataset; however, it represents a limitation, since images from the same patient may appear in both subsets, which could result in optimistic performance estimates. The evaluation methodology follows a two-phase approach:

1. Training Phase: All model optimization, including individual CNN training, GWO-based feature selection, and hyperparameter tuning, is performed using only the training set (2,451 images) with 5-fold stratified CV for robust parameter selection and performance estimation.
2. Testing Phase: Final performance evaluation is conducted once on the held-out test set (613 images) to provide unbiased performance estimates. The test set is never used during any optimization or model selection process.

Table 1. Dataset statistics and stratified splitting results

3.3 Enhanced preprocessing pipeline

The enhanced preprocessing pipeline incorporates techniques specifically designed for medical imaging applications, with the parameters for each stage summarized in Table 2. The preprocessing strategy includes:

(1) A normalization and contrast enhancement step combining min-max normalization with histogram equalization approximation to improve tumor boundary visibility;

(2) Architecture-specific resizing optimized for each model (EfficientNetB4: 380 × 380, Xception: 299 × 299, ResNet50V2: 224 × 224 pixels);

(3) Data augmentation comprising geometric transformations (rotation ± 30°, translation ± 25%, shear ± 25%, zoom 0.7–1.3) and photometric transformations (brightness 0.6–1.4, channel shift ± 0.2), with horizontal flipping included as a standard spatial augmentation and vertical flipping applied as a regularization technique to increase training set diversity and reduce overfitting, rather than to model anatomically realistic variations;

(4) Noise regularization through controlled Gaussian noise injection (probability 30%, σ = 0.01) to improve model robustness.

Table 2. Preprocessing stages and parameters

3.4 Base model architecture design

Our ensemble framework incorporates three state-of-the-art pre-trained CNN architectures, each contributing unique architectural advantages for medical imaging:

(1) EfficientNetB4: Utilizes compound scaling optimization with Swish activation function and built-in regularization layers, representing efficient architecture design for medical imaging applications [17].

(2) Xception: Employs depthwise separable convolutions for efficient feature extraction with reduced computational complexity while maintaining high representational capacity, particularly effective for medical texture analysis [18].

(3) ResNet50V2: Leverages residual connections and pre-activation batch normalization to enable deep network training with enhanced gradient flow, providing robust feature extraction for complex medical patterns [19].

Each base model undergoes comprehensive fine-tuning with initialization using ImageNet [20] pre-trained weights for effective transfer learning. Minimal freezing of initial layers (5–8%) enables better fine-tuning for medical data, while enhanced multi-layer classification heads incorporate progressive dropout (0.2–0.4) for robust generalization. Advanced Adaptive Moment Estimation with Weight Decay (AdamW) [21] optimizers with weight decay (10⁻⁵ to 2 × 10⁻⁵) and adaptive learning rates ensure optimal convergence, supported by extended training epochs with sophisticated callback mechanisms.

3.5 Stacking ensemble methodology

Our ensemble framework employs a stacking methodology, illustrated in Figure 1, that combines predictions from multiple base learners through a meta-classifier. The stacking approach consists of two levels:

(1) Level-0 (Base learners): Three pre-trained CNN architectures (EfficientNetB4, Xception, ResNet50V2) serve as base learners, each trained independently on the brain tumor dataset. These models generate both class predictions and extracted features.

(2) Level-1 (Meta-classifier): A Random Forest classifier serves as the meta-learner, trained on the concatenated outputs from base learners. The meta-feature vector construction follows:

$\vec{F}_{\text {meta }}=\left[\vec{P}_1, \vec{P}_2, \vec{P}_3, \vec{H}_1, \vec{H}_2, \vec{H}_3\right]$         (1)

where, $\overrightarrow{P_l}$ represents the prediction probabilities from model $i$, and $\overrightarrow{H_l}$ represents the high-level features extracted from the penultimate layer of model $i$. The total meta-feature dimension is 1,542, comprising 3 class probability outputs and 511 high-level features from each of the three models (3 × (3 + 511) = 1,542). The 511-dimensional feature vector per model is extracted from a custom bottleneck dense layer (Dense(512), ReLU activation) appended to each base model’s classification head before the final Softmax layer. One unit exhibiting consistently near-zero activation across the training set was excluded during feature construction, yielding 511 discriminative features per model. This bottleneck design standardizes the feature space across architectures with different native output sizes (1,792 for EfficientNetB4; 2,048 for Xception and ResNet50V2), ensuring balanced representation from each model in the meta-feature vector. To prevent overfitting in the stacking process, stratified 5-fold CV was applied to the training set only, where base learners were trained on four folds and generated predictions for the remaining fold, creating unbiased meta-features for meta-classifier training, as outlined in Algorithm 1.

3.6 Grey Wolf Optimization framework

Our sequential GWO-based optimization strategy addresses multiple aspects of ensemble configuration through a comprehensive two-stage optimization pipeline.

(1) Algorithm foundations: The GWO algorithm mimics the leadership hierarchy and hunting mechanism of grey wolves in nature [6], employing a social hierarchy consisting of alpha (α), beta (β), and delta (δ) wolves representing the best, second-best, and third-best solutions, respectively.

• Core position update mechanism: The algorithm updates each wolf’s position based on the guidance of the three best solutions:

$\vec{X}(t+1)=\frac{\vec{X}_1+\vec{X}_2+\vec{X}_3}{3}$        (2) 

where, $\overrightarrow{X_1}$, $\overrightarrow{X_2}$, and $\overrightarrow{X_3}$ represent position updates influenced by alpha, beta, and delta wolves, respectively.

• Exploration and exploitation balance: The algorithm balances exploration and exploitation through coefficient vectors $\vec{A}$ and $\vec{C}$, where $\vec{A}$ decreases linearly from 2 to 0 during iterations, enabling a smooth transition from exploration (global search) to exploitation (local search) phases.

(2) Feature selection optimization: The first optimization stage focuses on intelligent feature selection from the high-dimensional meta-feature space generated by concatenating outputs from all three base models (1,542 total features). Binary GWO Implementation: Each wolf position represents a binary feature selection mask $\vec{X} \in\{0,1\}^D$ where each element indicates whether a feature is selected (1) or excluded (0) [22]. The population initialization ensures balanced feature selection with approximately 50% of features initially selected.

• Enhanced transfer function: We employ an optimized sigmoid-based transfer function for continuous-to-binary conversion:

 $T_x=\frac{1}{1+e^{-10(x-0.5)}}$        (3)

The binary position update rule becomes:

$X_i^{(t+1)}=\left\{\begin{array}{cc}1 & \text { if } \operatorname{rand}()<T\left(X_i^{\text {continuous }}\right) \\ 0 & \text { otherwise }\end{array}\right.$        (4)

• Fitness function: The fitness function balances classification accuracy and feature efficiency using a weighted combination approach:

 $f(X)=0.9 \times \operatorname{Accuracy}(X)+0.1 \times \frac{D-|X|}{D}$        (5)

The weights 0.9 and 0.1 were chosen to assign overwhelming priority (90%) to classification accuracy, ensuring the optimization algorithm’s primary goal is to maximize predictive performance. The smaller weight (10%) for feature reduction acts as a secondary objective or tie-breaker. It encourages the selection of simpler, more computationally efficient models (those with fewer features) only when their classification accuracy is nearly identical, thus striking a balance between high performance and model efficiency. where $|X|$ is the number of selected features and $D$ is the total number of features. This approach encourages compact, efficient feature subsets while maintaining high predictive performance.

(3) Hyperparameter optimization: Following feature selection, the second stage focuses on optimizing the Random Forest meta-classifier hyperparameters using the selected feature subset.

• Random forest meta-classifier: The meta-classifier utilizes Random Forest [23], which consists of multiple decision trees with optimized parameters.
• Parameter space: The optimization targets key Random Forest parameters with appropriate ranges for medical imaging applications, as detailed in Table 3.

Table 3. Random Forest hyperparameter space

• Fitness evaluation: Hyperparameter configurations are evaluated using stratified 5-fold CV on the training set, ensuring robust and unbiased performance estimation. The average classification accuracy across folds is used as the fitness function:

$f(H)=\frac{1}{K} \sum_{k=1}^K \operatorname{Accuracy}_k(H)$        (6) 

where, $K=5$ is the number of folds and $H$ represents the hyperparameter configuration being evaluated.

3.7 Training and evaluation protocol

The training protocol follows a systematic approach optimized for competitive accuracy:

1. Base model training: Individual training of each CNN architecture using stratified 5-fold CV on the training set with adaptive epochs and callbacks for optimal convergence.
2. Meta-feature generation: Extraction of both prediction probabilities and high-level features from trained base models to create comprehensive meta-feature vectors.
3. Sequential GWO optimization: Two-stage application of GWO for feature selection followed by hyperparameter tuning, using CV performance on the training set as the objective function.
4. Final ensemble training: Training of the optimized Random Forest meta-classifier on the complete training set using selected features and optimal hyperparameters.
5. Performance evaluation: Assessment on the held-out test set using multiple metrics, including accuracy, precision, recall, F1-score, per-class accuracy, and detailed confusion matrix analysis.

3.8 Performance evaluation metrics

Our evaluation framework employs a comprehensive set of performance metrics specifically designed for medical classification tasks [24].

Primary classification metrics:

Accuracy $=\frac{T P+T N}{T P+T N+F P+F N}$         (7)

Precision $=\frac{T P}{T P+F P}$          (8)

$\operatorname{Recall}($ Sensitivity $)=\frac{T P}{T P+F N}$         (9)

Specificity $=\frac{T N}{T N+F P}$         (10)

F1-Score and weighted metrics:

$F 1-$ Score $=2 \times \frac{\text { Precision × Recall }}{\text { Precision }+ \text { Recall }}$        (11)

Weighted $F 1=\sum_{i=1}^C w_i \times F 1_i$         (12)

where, $w_i=\frac{n_i}{N}$ is the weight for class $i$, $n_i$ is the number of samples in class $i$, and $N$ is the total number of samples.

Reported runtimes for each component are as follows: base model training times include the full 5-fold CV loops; GWO optimization times cover all fitness evaluations (30 iterations × 20 wolves × 5-fold CV per evaluation); and meta-classifier training refers to the final training on the complete selected feature set. The total (150:16 min) is the sum of all these components.

The proposed framework achieved 99.84% overall accuracy on the Figshare test set. The improvement of 0.17 percentage points over the result reported by Aurna et al. [14] is modest in absolute terms and corresponds to a single additional correctly classified image on the current test partition; it should therefore not be interpreted as a definitive superiority claim, particularly since the two studies employed different preprocessing pipelines, data splits, and base architectures. More meaningful accuracy gains are observed relative to other methodological categories, with improvements of 0.98%–5.10% over custom CNN architectures and 1.14%–5.64% over single-model optimization approaches on the same benchmark. It is worth noting that near-99% classification accuracies have been reported by several prior methods on this dataset (e.g., Aurna et al. [14] at 99.67%, Nassar et al. [15] at 99.31%), and the proposed framework contributes an incremental methodological improvement rather than a uniquely high performance figure.

The observed performance stems from combining three architecturally diverse pretrained CNNs (EfficientNetB4, Xception, ResNet50V2) with a two-stage GWO pipeline that addresses feature selection and hyperparameter tuning in separate, coordinated stages. The GWO feature selection step reduced the meta-feature space by 80.6% while increasing CV accuracy by 0.87%, indicating that the original 1,542-dimensional feature space contained substantial redundancy. Per-class results show that pituitary and meningioma tumors were classified without error on the test set, while glioma achieved 99.30% accuracy; the single misclassified image belonged to the glioma class, which is consistent with its known morphological variability and partial overlap with other tumor types in MRI appearance. Regarding the choice of Random Forest as the meta-classifier, it was selected for its robustness in high-dimensional feature spaces and its well-defined hyperparameter space amenable to GWO optimization. A systematic comparison with alternative meta-learners such as logistic regression, linear SVM, or gradient boosting was not performed in this study and represents a direction for future work. The total training time of approximately 2.5 hours on an NVIDIA RTX 4090 reflects the cumulative cost of training three large CNNs with CV alongside two GWO optimization stages; this figure is reported as a resource characterization for reproducibility purposes rather than an efficiency claim. Several important limitations should be acknowledged. All reported results are obtained from a single publicly available benchmark dataset with image-level partitioning, as the Figshare dataset does not provide patient identifiers. The framework is evaluated on three tumor types using a single MRI modality (T1-weighted contrast-enhanced) and has not been validated on external data or in a clinical setting. These findings should therefore be interpreted strictly within the context of this benchmark study; generalization to other datasets or clinical workflows cannot be assumed without further investigation.