A Novel DenseNet Framework for Brain Tumor Classification Enhanced by XgBoost and Fire Hawk Optimization

The brain is the most important part of the human organ which comprises all the nervous system responsibilities of every activity. Based on the World Health Organization (WHO), Brain Tumour is the most severe disease of the human brain that have an irregular brain cell pattern. This tumour has several types, namely secondary and primary metastatic brain tumour [1]. The primary is a brain tumour that originates from human brain cells, which is a non-cancerous disease. But in secondary metastatic tumours, it spreads to other body parts, to the brain, through the blood.

Based on WHO, Brain Tumours is into four different categories, such as Grade I to Grade IV [2]. These Grades are categorised based on their malignant or benign. The malignant tumour can affect other parts of the human body. Also, benign tumour never attacks nearby tissue or other organs of the human body. To predict the brain tumour, magnetic resonance imaging (MRI) and computer tomography (CT) are used as standard methods so far. The malignant tumours, Grade III and Grade IV, are very aggressive brain tumours that also affect other parts.

In some cases, there are three major primary brain tumours, namely Glioma, pituitary and meningioma [3]. The glioma tumours develop from the brain’s glial cells. The Pituitary tumour is a type of benign tumor that develops in the pituitary gland. This gland provides a few essential hormones in the body and also forms a base layer of the brain. The Meningioma tumours develop in a protective membrane of the spinal cord and brain. These types of tumours can be diagnosed if it is predicted earlier. But unfortunately, the MRI and CT cannot provide that much accuracy in prediction.

In recent times, the medical field has emerged with a machine learning (ML) and deep learning (DL) models of its higher training rate. The ML and DL models are very supportive of earlier prediction by training huge datasets of tumour images [4]. Some of the popular DL and ML methods that are applied for medical applications are Convolutional Neural Networks (CNN), Support Vector Machine, AdaBoost, Naive Bayes, Decision Tree, DenseNet, GoogLeNet, VGGNet, ResNet, AlexNet, Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM), etc. These learnings are methods used for classification and feature extraction models for an early diagnosis.

In brain tumour classifications, the current approaches struggle to identify and use the most relevant features from medical images. Existing feature extraction techniques, such as traditional CNNs, may fail to capture complex patterns in MRI images due to their limited ability to model spatial dependencies and multi-scale features. Furthermore, feature selection methods in existing systems are often either too simplistic or computationally expensive. It leads to either a loss of crucial information or the inclusion of irrelevant features.

Nowadays, to enhance the prediction ability, metaheuristic methods are used as problem-solving models. These methods are inspired by and developed from the nature and biological behaviour of birds and animals. The metaheuristics methods are used to find the optimal solution among all possible solutions. The metaheuristics methods are applied to a medical application.

This work proposes a superior approach by combining SC-SE DenseNet for advanced feature extraction and FHO for optimized feature selection. This hybrid method not only improves the relevance and quality of the selected features but also enhances classification accuracy by curtailing the impact of redundant or noisy features. The remaining part of this work is contributed in the following: in Section 2, the related works based on brain Tumours are described. Section 3 discusses the proposed work. Section 4 discusses the experimental results, and the paper concludes with a conclusion and references.

The proposed system outlines a method for efficient brain tumor categorization using MRI images. The steps involved in the proposed system are given in Figure 1. Initially, feature extraction is performed using SC-SE DenseNet. It is used to capture detailed characteristics from the MRI images. Then, feature selection is applied to increase the relevant features while eliminating redundant ones. The optimized feature vectors are fed into the XGBoost classifier for robust classification. Finally, FHO is used again to fine-tune the feature set and optimize the classification model’s performance.

1.png

Figure 1. Proposed workflow

3.1 SC-SE DenseNet

The SC-SE DenseNet is a modified version of DenseNet that integrates Self-Calibrated Convolutions (SC-Conv) and Squeeze-and-Excitation (SE) Blocks to improve the feature extraction and representation. The SC-SE DenseNet consists of an initial convolutional layer, four densely connected blocks (Dense Blocks), and a global average pooling (GAP) layer to extract more relevant features of the input image. The architecture of SC-SE DenseNet is given in Figure 2.

2.png

Figure 2. Proposed SC-SE DenseNet architecture

3.1.1 SC-Conv

The SC-Conv improves feature extraction by dynamically calibrating the input features based on their spatial and contextual information. This is achieved through the combination of two parallel paths: the Standard Convolution Path and Self-Calibration Path. The Standard Convolution Path extracts spatial features using conventional convolution operations. The Self-Calibration Path dynamically adjusts feature representations using global context information. Let the input feature map be $\mathrm{X} \in \cdot \mathbb{R}^{H \times W \times C}$, where H denotes the height of the feature map, W denotes the width of the feature map and C denotes the number of channels. The SC-Conv operation is defined as:

$Y=F_{ {conv }}(X)+F_{ {calibrate }}(X)$                 (1)

where, $Y=F_{ {conv }}(X)$ is the output of the standard convolution path and $Y=F_{ {calibrate}}(X)$ is the output of the self-calibration path. This Standard Convolution Path applies a conventional convolution operation. It can be defined as;

$F_{{conv}}(X)=W * X+b$                   (2)

where, * denotes the Convolution operator, W denotes the weights of the convolutional kernel and b is the bias term. The self-calibration path involves three key steps: Global Context Extraction, Channel-Wise Calibration and Feature Recalibration. For global context extraction, it computes a spatial summary of X using global average pooling as follows:

$g_c=\frac{1}{H \times W} \sum_{h=1}^H \sum_{w=1}^W X(h, w, c)$                  (3)

where, $g_c$ represents the aggregated information for channel c. The result $\mathrm{g} \in \mathbb{R}^C$ Captures the global context. For channel-wise calibration, the global context vector g is calibrated through a learnable transformation:

$s=\sigma\left(W_s g\right)$                  (4)

where, $W_s$ is the learnable weight matrix, σ is the non-linear activation function, $\mathrm{g} \in \mathbb{R}^C$ is the calibrated scaling vector for each channel. For feature recalibration, the calibrated vector s is used to reweight the original input as follows:

$F_{ {calibrate }}(X)=X \odot s$                (5)

where, ⊙ denotes element-wise multiplication. The final output Y of the SC-Conv operation integrates both paths:

$Y=F_{ {conv }}(X)+F_{ {calibrate }}(X)$                (6)

3.1.2 SE

The SE block is a powerful mechanism in neural networks designed to enhance the representational power of CNNs [22]. This block is inserted to concentrate on important features and suppress the irrelevant features. The SE block works by learning channel-wise feature recalibration weights and allowing the network to adaptively attend to essential features. This process involves two fundamental operations: squeezing and exciting as shown in Figure 3. The squeezing step aggregates channel-wise information to capture global statistics and the exciting step performs feature re-weighting based on learned parameters.

3.png

Figure 3. SE block

In Squeeze operation, it compresses the spatial dimensions of X into a channel descriptor by performing global average pooling. In an excitation operation, the squeezed vector passes through a fully connected (FC) bottleneck network with two layers to capture channel dependencies. Overall, the SE block focuses purely on recalibrating the channel importance of feature maps using global context. The SC-Conv improves feature extraction by combining spatial and channel recalibration. It includes both standard convolutions and a dynamic self-calibration path.

3.2 Fire Hawk Optimizer (FHO)

The metaheuristic model [23] of the FHO method is inspired by the hunting strategy of black kites, whistling kites, and brown falcons that are known as Fire Hawks. These birds are often used to set a fire around the prey to catch them [24].

Initially, the fire hawk is positioned as the prey in its territory. The fire hawk is considered to catch in its territory and protect that not affected by other fire hawks. The distance limit between the number of the prey and the fire hawk is estimated. Meanwhile, position updation is performed by fire hawks so that the prey from its territory is not affected by other territories' fire hawks. The Gaussian distribution is used to provide a search loop in random where the number of prey and total fire hawks are estimated.

Core Principles of FHO:

(1) Initialization

A population of potential solutions (denoted X) is initialized randomly within the bounds of the problem's search space. Each solution vector represents a candidate position in the d-dimensional problem space. It can be mathematically defined as:

$x_j^i(0)=x_{j, \text { min }}^i+r_1\left(x_{j, \text { max }}^i-x_{j, \text { min }}^i\right)$                   (7)

where, $x_{j, \min }^i$ and $x_{j, \max }^i$ are the bounds, $r_1$  is a uniform random number, i=1,2,…,N, and j=1,2,…,d.

(2) Classification into Fire Hawks and Prey

In this stage, the candidates with better objective function values become fire hawks and others represent prey. Fire hawks spread "fires" strategically to herd prey into more favourable search regions. The total distance between each fire hawk and its nearest prey is computed to define territories as follows:

$D_l^k=\sqrt{\left(x_2-x_1\right)^2+\left(y_2-y_1\right)^2}$                   (8)

This calculation ensures each fire hawk's territory is distinguished for optimal search coverage.

(3) Position Updates

Fire Hawks (FH) adjust their positions based on proximity to the global best solution (GB) and influence from other hawks. It can be mathematically defined as:

$F H_l^{{new }}=F H_l+r_1\left(G B-F H_l\right)-r_2\left(F H_{ {near }}-F H_l\right)$                (9)

where, $r_1$ and $r_2$  are random coefficients, l is the index for FH. l=1,2,…,n. n is the total number of fire hawks in the search space. Likewise, the prey reacts by moving within or outside fire hawk territories.

$P R_q^{\text {new }}=P R_q+r_3\left(F H_l-P R_q\right)-r_4\left(S P_l-P R_q\right)$                  (10)

where, $S P_l$ represents a safe zone or position within the fire hawk’s territory, q is the index of Prey Reactions (PR).

(4) Iterative Search and Optimization

Fire hawks simulate dynamic behaviors like dropping burning sticks and expanding their territories, while prey constantly adapts to escape threats. These steps refine solution quality iteratively.

3.3 FHO-based feature optimizations

Feature selection is an essential step of DL models which identifies the most relevant input features and eliminates irrelevant or redundant ones. In this work, feature selection is combined into the DenseNet model using the FHO to improve the tumor classification accuracy. For feature optimization, the FHO is initialized with a population of candidate solutions. Each candidate denotes a subset of the features extracted by the proposed DenseNet. The population is randomly set, where each solution relates to a different combination of features.

The pseudocode for the proposed feature selection is given below:

Each solution is calculated using a fitness function that measures the classification loss of a classifier trained on the selected features. The algorithm refines the feature subsets by identifying better-performing solutions (Fire Hawks) and selecting random solutions from this set. The current feature subset is updated based on these better solutions. The solution is accepted when the update improves the classifier’s performance. The process continues until convergence, with the algorithm returning the best feature subset corresponding to the minimum classification error. This subset represents the most relevant features for the classification task.

3.4 FHO-based parameter tuning

XGBoost is a robust ensemble learning method used for classification tasks [25]. The performance of the XGBoost model is greatly influenced by the hyperparameters. It includes the learning rate, maximum tree depth, and the number of estimators. The proper tuning of these hyperparameters improves the model’s accuracy and efficiency. In this work, FHO is used to identify the optimal hyperparameter configuration for XGBoost. This technique accelerates the optimization process and increases the overall model performance. The pseudocode for the proposed parameter tuning is given below:

The population is initialized randomly. The solutions are classified based on their fitness values whereas Fire Hawks represent better solutions. Fire Hawks adjust their positions based on proximity to the global best solution (X_best) and prey. The prey reacts to the influence of Fire Hawks. The fitness of each solution guides the optimization process in both the exploration and exploitation phases.