Brain Tumor Classification in MRI Using Hybrid ASA-Based Deep Learning and Masi-Entropy Multilayer Thresholding Segmentation with Sunflower Optimization

Brain tumors arise from excessive cell mutations or proliferation, leading to abnormal cell clusters that interfere with brain function and damage healthy cells [1]. Common symptoms of brain tumors include memory issues, fatigue, personality shifts, nausea, speech difficulties, and vision problems [2, 3]. Radiologists used to detect and classify malignancies in brain images [4]. Magnetic resonance imaging (MRI) and computed tomography (CT) are commonly used to obtain data on various regions of the human body [5, 6].

Physicians use MRI of the brain to perform two types of tasks: (i) determining whether an MRI image is healthy or tumorous [7-9] and (ii) classifying an MRI image into distinct types [10-12]. Moreover, manual identification is time-consuming, inefficient, and unreliable when dealing with a vast amount. An inaccurate diagnosis can have serious consequences, possibly leading to patient death. Furthermore, classifying brain tumors into multiclass classifications presents more challenges than binary classifications. There is an urgent need for a reliable CAD system for tumor categorization to assist [13].

Although medical examinations such as MRI, CT, positron emission tomography (PET), and magnetic resonance spectroscopy (MRS) assist healthcare professionals in detecting brain tumors and are regarded as essential tools, these techniques have some limitations, such as failure to diagnose infected tumors, particularly in the early stages, which can result in false negative results. Some medical examinations, such as biopsies and intrusive imaging, can be invasive and unpredictable. Furthermore, these techniques are extremely expensive, and radiation exposure during CT and MRI scan procedures can gradually increase the possibility of developing a brain tumor. Other medical treatments, including angiography, can cause pain. As a result, all the implemented approaches are entirely dependent on the analysis of experts who perform these evaluations according to their expertise.

However, the algorithms suggested by various researchers have produced excellent outcomes and made significant contributions to the medical profession by assisting researchers in disease identification and classification; however, the possibility of obtaining more accurate, efficient, and improved results when recognizing brain tumors. Furthermore, the majority of existing models face difficulties, such as underfitting, poor generalization, an imbalanced dataset, overfitting, the need for expensive computational resources, and complexity.

Recent developments greatly improved the ability to identify and classify patterns in medical images, with a focus on MRI data for tumor detection. Although prior studies have demonstrated the potential of ML and DL algorithms in this area, they often have significant drawbacks, including high computational demands. Moreover, biases in training data and the complex, often opaque nature of DL model decision-making can further hinder their effectiveness [14]. To overcome these issues, it is essential to use robust DL models trained on diverse datasets and apply methods like cross-validation to ensure the model generalizes well. The approach proposed here by delivering precise and prompt evaluations, utilizing a straightforward, end-to-end DL algorithm that is practical for real-time use.

To address these issues, the proposed enhanced AlexNet classifier with the Anopheles Search Algorithm (ASA) was combined with a DL algorithm in the suggested research to achieve more accurate, efficient, and improved results in the medical field. The enhanced AlexNet classifier, with its deep convolutional layers, excels at extracting high-level features from MRI images that are critical for distinguishing between tumor and nontumor regions. It detects complex patterns and textures that are indicative of brain tumors. ASA integration allows the hybrid system to fine-tune the feature-representation process. The ASA efficiently tunes the AlexNet architecture's hyperparameters, such as the learning rate, batch size, and regularization parameters. This optimization process ensures that the classifier is well suited to the unique features of brain tumor MRI datasets. The hybrid classifier aims to detect and classify brain tumors with high accuracy, sensitivity, and specificity by combining DL capabilities with optimization techniques. The hybrid approach, which takes advantage of both enhanced AlexNet's classifier capabilities and the ASA's optimization, has the potential to achieve cutting-edge results in brain tumor classification tasks. It seeks to outperform traditional methods by incorporating advanced techniques for classification and parameter optimization.

The novelty of the proposed method is summarized as follows.

·The enhanced AlexNet algorithm was optimized using the Anopheles Search Algorithm (ASA) for detecting brain tumors.

·The proposed technique is flexible and can be adapted in variation of size and location of the tumors.

·The proposed model compared to existing methods, extensive simulations were performed using publicly available datasets from Kaggle, such as the Brain Tumor Classification (BTC) and CE-MRI Figshare datasets.

The paper is structures as follows. Section II discusses the deep analysis of existing model and techniques. Section III tailored the methodology. Section IV shows the experimental findings and Section V discusses conclusions and future implications.

The proposed approach is divided into five stages, as illustrated in Figure 1. Brain tumor images were initially collected from the Kaggle dataset. The image was converted to grayscale, and the noise was removed using a filter. An FNLM filtering approach was employed. In addition to this segmentation, MasiEMT-SFO was used. The image was segmented based on histogram values, which increased segmentation accuracy. Subsequently, the features were obtained using VGG-19. The Minimum Redundancy Maximum Relevance (mRMR) method was used to select the best features from the feature extraction process. Finally, the selected features were allocated to a hybrid classifier called the AlexNet classifier method, which was improved using the ASA for classification. This will also provide information regarding tumor malignancy.

1.jpg

Figure 1. Proposed methodology block diagram

3.1 Dataset

3.1.1 BTC dataset

The brain tumor classification (BTC) dataset [40] in this study finds appropriate DL classifiers by training and testing multiple learning-based algorithms on MRI datasets. The dataset contained two MRI brain images: test and training. This study only utilized MRIs for meningiomas, pituitary gland tumors, and gliomas. Table 2 presents a description of the BTC dataset. Figure 2 shows the sample images.

Table 2. BTC dataset description

2.jpg

Figure 2. Sample MRI dataset

3.1.2 CE-MRI Figshare dataset

The publicly available CE-MRI Figshare dataset [41] was the second dataset used to classify tumors. The T1 modality highlights particular features of each multiclass brain tumor. The most recent version of this dataset contains 708 images of meningiomas, 1426 gliomas, and 930 images of pituitary tumors. We randomly divided the dataset by allocating 80% for training and 20% for testing.

3.2 Preprocessing

Preprocessing is a critical step in brain tumor detection, as it enhances detection accuracy by removing unnecessary elements from MRI scans, such as surrounding tissues or blood vessels. The images are first converted to grayscale to eliminate extraneous data, simplifying the processing and reducing data demands. Following this, noise is removed using the Fast Non-Local Means (FNLM) filter [42], which is particularly effective in preserving image structure while reducing noise.

3.2.1 Grayscale conversion

Converting to grayscale minimizes storage and processing requirements, making the subsequent analysis more efficient.

3.2.2 Filtering

Unlike conventional filters that may blur the entire image, the FNLM filter retains crucial edge information by weighting pixel values based on their Euclidean distance. This approach, now feasible due to advances in computing power, has proven effective in enhancing image clarity, as demonstrated by qualitative and quantitative analyses.

$N_L[i](M)=\sum_{N \in i} W(M, N) i(N)$    (1)

where, W(M, N): weight of the pixels M&N, ranges between 0≤W(M,N)≤1 and $\sum_{N \in i} W(M, N)=1$.

Here, i(N) and i(M) denote the luminance of pixels N and M, respectively, and Eq. (2) establishes weight W(M, N) regarding the similarity between pixels M and N.

$W(M, N)=\frac{1}{Z(M)} e^{-\frac{\left\|P\left(x_M\right)-P\left(x_N\right)\right\|}{D^2}}$    (2)

where, X_M and X_N are pixel vectors for M and N. Z (M) is the normalization coefficient and determined as in Eq. (3).

$Z(M)=\sum_N e^{-\frac{\left[P\left(X_M\right)-P\left(x_N\right)\right]^2}{D^2}}$    (3)

The coefficient D in Eqs. (2) and (3) control how much noise is eliminated, and this value is typically a constant multiplied by the standard deviation. A larger D value improves the efficacy of noise elimination while increasing image smoothing.

3.3 Segmentation

Segmenting images is an essential component of medical diagnosis. Segmentation based on images can be challenging. A later step is used to segment infected brain areas on MRI: The preprocessed MRI brain tumor image is initially turned into a binary image using an FNLM filter. The binary image is then segmented into the next stage. The MasiEMT-SFO approach was employed in this study to segment MRI tumor images efficiently.

3.3.1 MasiEMT-SFO

A new multi-level threshold-based image segmentation approach was developed using Masi entropy as the objective function. While bi-level thresholding is effective for segmenting images with a single object against a background, it often falls short when dealing with complex images containing multiple objects. In such cases, multi-level thresholding provides better segmentation results. To optimize the threshold search process and reduce computational costs, metaheuristic algorithms are commonly used. This study proposes the combination of the Sailfish Optimization (SFO) algorithm with Masi entropy for multi-level threshold-based image segmentation, significantly lowering computational complexity. The SFO algorithm optimizes the search space to find ideal thresholds, making the segmentation process more efficient even as the number of thresholds increases. For the segmentation of an image, we assume an array of m thresholds {T₁ ... T_N}. These thresholds were applied for image segmentation in the N+1 class. Gray levels [0, T₁], [T₁+1, T₂] ... [T_N+1, L−1] are covered by pixels G₀ and G₁.... G_N, in the range mentioned above. To perform multilayer thresholding, Masi entropy is expressed as follows:

$\begin{gathered}\mathrm{M}_{\mathrm{R}}\left(\mathrm{I} \mid\left\{\mathrm{T}_1, \ldots, \mathrm{~T}_{\mathrm{N}}\right\}\right) =\mathrm{M}_{\mathrm{R}}\left(\mathrm{G}_0\right)+\mathrm{M}_{\mathrm{R}}\left(\mathrm{G}_1\right)+\ldots+\mathrm{M}_{\mathrm{R}}\left(\mathrm{G}_{\mathrm{N}}\right)\end{gathered}$    (4)

where,

$\mathrm{M}_{\mathrm{R}}\left(\mathrm{G}_0\right)=\frac{\left(\log \left(1-(1-R) \sum_{i=1}^{T_1} \frac{H_i}{W_0} \log \left(\frac{H_i}{W_0}\right)\right)\right)}{1-R}$    (5)

$\mathrm{M}_{\mathrm{R}}\left(\mathrm{G}_1\right)=\frac{\left(\log \left(1-(1-R) \sum_{i=T_1+1}^{T_2} \frac{H_i}{W_1} \log \left(\frac{H_i}{W_0}\right)\right)\right)}{1-R}$    (6)

$\mathrm{M}_{\mathrm{R}}\left(\mathrm{G}_{\mathrm{N}}\right)=\frac{\left(\log \left(1-(1-R) \sum_{i=T_N+1}^{L-1} \frac{H_i}{W_N} \log \left(\frac{H_i}{W_1}\right)\right)\right)}{1-R}$    (7)

where, $\mathrm{W}_0=\sum_{i=1}^{T_1} H_{\mathrm{i}}, \mathrm{W}_1=\sum_{i=T_1+1}^{T_2} H_{\mathrm{i}}, \ldots, \mathrm{W}_{\mathrm{N}}=\sum_{i=T_N+1}^{L-1} H_{\mathrm{i}}$.

The pollination process of the two nearest sunflower ports in the direction of the sun was simulated using the SFO algorithm. Sunflowers have a consistent life cycle: they emerge each day with the sun, much like the hands of the clock. At night, they faced in the reverse direction, patiently waiting for the sun to rise in the morning. The inverse square law of radiation is another important nature-based optimization method. In this context, the inverse square-law principle is a key component of natural optimization. The amount of radiation received varies directly with the plant's distance from the sun but remains relatively constant at closer distances. In contrast, longer distances resulted in less solar radiation absorption. Hence, similar considerations apply to this study: moving closer to solar sources maximizes the potential for global optimization as shown in Figure 3. The calorie Q consumed by each plant was calculated using Eq. (8):

$\mathrm{Q}=\frac{P}{4 \pi d^2}$    (8)

where, Q is the quantity of heat obtained, P is the power of the sun, and c is the separation between the sun Y^* (the best solution) and the sunflower Y_i.

3.jpg

Figure 3. Flowchart of the sunflower optimization algorithm

The direction vector of sunflowers is defined as follows:

$\vec{S}=\frac{\mathrm{Y}^*-\mathrm{Y}_{\mathrm{i}}}{\left\|\mathrm{Y}^*-\mathrm{Y}_{\mathrm{i}}\right\|} \mathrm{i}=1,2, \ldots \ldots, \mathrm{~m}$    (9)

where, Y^* represents the global best solution, Y_i represents the solution i, and m represents the population size. Eq. (10) shows the calculation of each sunflower Y_i’s step size (D).

$D=\mu \times p_i\left(| | \mathrm{Y}_{\mathrm{i}}-\mathrm{Y}_{\mathrm{i}-1} \| \mid\right) \times\left\|\mathrm{Y}_{\mathrm{i}}-\mathrm{Y}_{\mathrm{i}-1}\right\|$    (10)

where, µ is a constant value indicating a plant's "inertial" displacement, and $p_i\left(\left\|\mathrm{Y}_{\mathrm{i}}-\mathrm{Y}_{\mathrm{i}-1}\right\|\right)$ is the probability of pollination.

All sunflower steps were limited to the maximum step size of the DMAX to avoid skipping each solution boundary. Eq. (11) shows the computation of the maximum step size:

$D_{M A X}=\frac{\left\|Y_{M A X}-Y_{M I N}\right\|}{2 \times \mathrm{m}}$    (11)

where, Y_MAX is the upper boundary and Y_MIN is the lower boundary. The most current planting method in the population is:

$\vec{Y}_{i+1}=\vec{Y}_i+D \times \vec{S}$    (12)

This approach begins with creating a uniform or random population of entities. Evaluating each individual helps to determine which one will evolve into the "Sun." Although future versions may manage multiple suns, the current study focused on a single individual. Subsequently, all new entities resembling sunflowers align themselves with the sun and migrate in a controlled random manner. The SFOA was used to identify high-quality solutions in problem-solving scenarios by reducing the fitness function.

An optimization algorithm that finds the most optimal solution by analogizing sunflower pollination is shown in Figure 3. This algorithm begins by initializing a population of random agents. A predetermined percentage of flowers that are the furthest from the sun is then eliminated by the algorithm as part of a selection procedure. By taking this step, resources can be directed towards the more promising applicants. Each bloom then goes through a procedural upgrade to improve its pollination behavior or position. The best-performing flowers are chosen to pollinate close to the sun, producing new individuals close to the optimal solution. By concentrating on regions with significant potential, this local exploration seeks to enhance the best solution.

A new population is created once the new individuals have been created. If a better option is found, the algorithm updates the sun or the best solution. Otherwise, it checks to see if this population constitutes a global minimum. The procedure is repeated until an ideal solution is found or the maximum number of iterations (called days) is reached. At this stage, the algorithm comes to an end after repeatedly selecting, orienting, and pollinating the population to get the greatest potential result.

3.4 Feature extraction

Feature extraction is an essential component for classification. The significant elements of the images have an enormous impact on the classification performance. Object characteristics are classified as global or local, based on color, size, or structure. Texture and color are local characteristics; however, their structures are global. Deep and artisan features were acquired for image categorization in this study. Feature extraction techniques are critical for tumor detection. The VGG-19 algorithm obtains various features from the segmented images. The process described above extracts specific features from an image for brain tumor imaging.

3.4.1 VGG-19

VGG-19 enables it to capture distinct features in images, which is critical for tasks such as medical image analysis, in which subtle details can have a diagnostic impact. Medical imaging studies, including MRI analysis, have shown that VGG-19 excels in feature extraction tasks because of its ability to effectively capture low and high-level features [43]. This performance reliability is critical for accurate detection of tumors using MRI. VGG-19 achieves a balance between depth and computational feasibility as shown in Figure 4.

4.jpg

Figure 4. VGG-19 Architecture model

The feature extraction capabilities of VGG-19 provide a rich set of features for the enhanced AlexNet brain tumor classifier for processing. These features are more abstract and representative than earlier layers, which improves the classifier's ability to detect complex tumor-related patterns. Using VGG-19 as the feature extractor, the enhanced AlexNet classifier can focus on refining these features through additional training. This fine-tuning process improves classification accuracy and robustness.

3.5 Feature selection (FS)

The FS is a critical task that reduces the number of features by removing redundant, irrelevant, and noisy data. In this study, the mRMR method is used to select the best features. mRMR-based feature selection is especially effective in detecting brain tumors from MRI images because of its ability to select highly relevant and non-redundant features, optimize subset sizes, robustly handle noise, ensure interpretability, and accommodate complex data structures.

3.5.1 mRMR

mRMR is a type of feature selection method that acts as a filter to identify the optimal feature set by maximizing the correlation with the target variables while minimizing redundancy among features. This classic approach effectively gathers relevant and nonredundant features. mRMR is efficient for selecting feature subsets by emphasizing relevant features and excluding irrelevant features. Mutual information plays a crucial role in assessing the similarities between features. Specifically, for features f_i within set S, maximizing mutual information enhances the relevance between features and the target class c:

$\max D(S, c), D=\frac{1}{|S|} \sum_{f_i \in S} I_M\left(f_i ; c\right)$    (13)

where, I_M represents the mutual information between feature f_i and class c.

$\min R(s), R=\frac{1}{\left|S^2\right|} \sum_{f_i, f_j \in S} I_M\left(f_i, f_j\right)$    (14)

Feature selection operates on the entire feature vector by using leave-one-out cross-validation with a voting scheme. In this process, each case selected the best features through cross-validation. Each feature receives a vote if selected by case, and those with the highest accumulated votes are chosen as the final features. These selected features are subsequently utilized in the classification stage to categorize each super pixel as tumor or non-tumor. The mRMR-based feature selection is highly beneficial for enhancing the performance of an enhanced AlexNet classifier in brain tumor detection from MRI images. It optimizes feature selection by maximizing relevance and minimizing redundancy, thereby enhancing the classification accuracy and robustness in medical imaging. The integration of mRMR with enhanced AlexNet combines their strengths, providing a potent approach for precise and dependable brain tumor diagnosis using MRIs.

3.6 Classification

After the FS, the selected features are fed into a hybrid AlexNet classifier with the ASA as an input. Conventional algorithms frequently require handwritten functions, which are time consuming and may not fully capture the complexities of medical imagery. DL models such as AlexNet can automatically discover important features and decrease the need for manual feature engineering. The enhanced AlexNet output layer uses a SoftMax classifier to detect brain tumors [44]. Ultimately, the proposed classifier efficiently detects brain tumors while improving the classification accuracy. Figure 5 shows a flowchart of the proposed method.

5.jpg

Figure 5. Process flow of the proposed system

AlexNet as shown in Figure 6 consists of five convolutional layers, three max-pooling layers, two normalization layers, two fully connected layers, and a softmax layer, with ReLU activation applied after each convolution. The input size is typically 224 × 224 × 3, but with padding, the actual output is 227 × 227 × 3, and the model contains over 60 million parameters.

6.jpg

Figure 6. AlexNet architecture

3.6.1 Enhanced AlexNet

The architecture includes five convolutional layers, three fully connected (FC) layers, and various pooling and normalization layers, processing input images of size 227 × 227 × 3 [45]. The ReLU activation function is used throughout to address issues like gradient vanishing, allowing for faster convergence during training. Dropout is applied in the FC layers to prevent overfitting by training only a subset of neurons in each iteration, thus improving the network's generalization capability.

The enhancements described in this study differ from the existing AlexNet classification framework in the following ways.

·Convolutional layers were added to the existing AlexNet architecture to maintain responsive local areas and employ max-average pooling methods to improve image classification.

·A global average pooling layer that can significantly reduce overfitting while retaining the final features. The absence of many network variable computations does not affect the final result, which improves the network performance.

·Finally, a local response normalization (LRN) layer was added to the convolutional layer to avoid specific additional numerical issues, thus eliminating neuron saturation. Following each convolution layer, the BN layer was included and directed to the next network layer as shown in the Figure 7.

7.jpg

Figure 7. Architecture of the enhanced AlexNet model

The output from the convolutional layer is given as:

$X_j^1=f\left(\sum_{a=1}^N W_j^{I-1} * Y_a^{I-1}+b_j^i\right)$    (15)

where, X_j¹ s the j-th feature map on layer l, Y_a^I-1 is the feature map on layer I–1, W_j^I-1 is the j-th kernel on layer I-1, N is the total number of features on layer I–1, b_jⁱ is the bias for the j-th feature map on layer l, and (*) denotes convolution. Pooling layers reduce the feature map size, and the final layer utilizes the SoftMax function for classification:

$\operatorname{ReLU}(\mathrm{X})= \begin{cases}0, & X<0 \\ X & X \geq 0,\end{cases}$    (16)

$\operatorname{SoftMax}\left(X_i\right)=\frac{e^{x_i}}{\sum_{y=1}^m e^{x_y}}$    (17)

where, X_i and m are the input data and number of classes, respectively. To normalize features, AlexNet uses parameters to adjust the mean and variance.

$P_i=\varphi \tilde{t}_i+\beta$    (18)

$\tilde{t}_i=\frac{t_i-\mu}{\sqrt{\sigma^2+\gamma}}$    (19)

where, $\mu$ and $\sigma$ are the mean and variance, respectively. $\gamma$ is a constant, and is the parameter for feature extraction. To optimize the parameters and reduce error rates during classification, AlexNet was tuned using the Adaptive Simulated Annealing (ASA) algorithm. This optimization adjusts $\beta$ to increase accuracy and $\phi$ to minimize error rates.

3.6.2 ASA

Recent studies have uncovered the molecular mechanism of female Anopheles mosquitoes, which enables them to detect their prey several miles away. Human sweat contains a compound called methyl phenol 4, which is detected by insect olfactory receptors (Igor-1), helping it locate humans. It has been observed that Igor-1 cells are exclusively present in female mosquitoes. When malaria parasites are ready to move from one host to the next, they prompt the infected body to release odors that attract Anopheles mosquitoes. Furthermore, the search agents within the solution space act like mosquitoes. These agents evaluate the optimality of each point by randomly navigating through the solution space. As the iterations progress, more agents converge towards the optimal points.

$\mathrm{D}=\mathrm{a} \log (\mathrm{C})+\mathrm{b}$    (20)

where, D denotes the odor density, b signifies the tracking constant, and C represents the Weber-Fechner chemical concentration coefficient. In Eq. (1), if a denotes the inverse distance between point X_i mosquito A_i.

$Odor _{A_i X_i=\frac{1}{{Dist}\left(A_i X_i\right)} \times log \left( { fitness }\left(X_i\right)\right)+b, \quad 0 \leq b \leq 0.5}$    (21)

where, ${Dist}\left(A_i X_i\right)=\sqrt{\sum_{j=1}^n\left(\left(x_j\left(A_i\right)-x_j\left(X_i\right)\right)^2\right)}$.

The metaheuristic ASA consists of the following steps.

·The optimization problem can be described as follows:

${{}_{x}^{max}{f(X)}},~~X \in x_{i}~~~i = 1,2,\ldots.N$    (22)

where, f(x) denotes the objective function, X represents the set of decision variables, x_i denotes a variable with a set of possible values for each decision, and N signifies the number of decision variables. The definition of continuous decision variables includes the upper and lower bounds, as specified in Eq. (23).

$l_{x_i} \leq x_i \leq u_{x_i}$    (23)

$x_i \in\left\{x_1, x_2, \ldots . x_n\right\}$    (24)

·The number of Anopheles mosquitoes, number of decision variables, and range of each decision variable were the variables of the search algorithm.

·In this step, mosquitoes are dispersed randomly throughout the search space, and the location of each mosquito's optimal point is determined.

·Each mosquito senses the odor density according to Eq. (23), and based on the value obtained, it travels in the direction of the optimal point. In a two-dimensional space, for instance, we have:

$Odor _{A_i X_i=\frac{1}{\sqrt{\left(x_{A_i}-x_i\right)^2+\left(y_{A_i}-y_i\right)^2}}} \times log \left(\right. fitmess \left.\left(X_i\right)\right)+b, 0 \leq b \leq 0.5$,    (25)

$x_{ {new }}=x_{ {old }} \pm Odor _{A_i X_i} \times\left|x_{A_i}-x_i\right|$    (26)

$y_{ {new }}=y_{ {old }} \pm Odor _{A_i x_i} \times\left|y_{A_i}-y_i\right|$    (27)

·The stopping conditions are as follows:

For a fixed number of iterations, based on the conducted simulations, a range between 15 and 30 is recommended. This approach is proposed as a decision parameter for problems characterized by intricate objective functions and a substantial number of decision variables.

Attaining a sufficiently good solution: this criterion is suitable for problems with a defined optimality threshold.

Number of iterations without alteration in the optimal function; this is applicable to problems where achieving solution convergence is crucial.

3.6.3 Hybrid Enhanced AlexNet with ASA

The optimization was based on Anopheles mosquitoes with approximately three tents. Certain Anopheles mosquitoes were discovered near three tents but were not carriers. Initially, an infant infected with malaria and parasites was ready to spread. Second, malaria-infected children and parasites are unprepared for spreading the disease, and third, there is the existence of a healthy child. The results indicated that mosquitoes showed twice the attraction to the scent of children in the first tent compared to the others. Figure 8 illustrates the flowchart for the ASA-AlexNet.

8.jpg

Figure 8. Flowchart of proposed hybrid enhanced AlexNet with ASA

Step 1: Initialization.

where ASA denotes the population, B denotes the best value, and I denote the number of iterations. The decision variable was bounded by lb (lower bound) and ub (upper bound). The ASA algorithm was operationalized using the following equations:

$\begin{aligned} P_i=(1, j) & =l b_j+X\left(u b_j-l b_j\right) ; j =1,2,3 \ldots . N\end{aligned}$    (28)

$P_i=P_i+X\left(P_i\right)$    (29)

In Eq. (30), the inverse distance between X_i and point Y_i is calculated, where C denotes the optimal point in the solution space based on the location X_i, as follows:

$O_{X_i Y_i}=\frac{1}{{Dist}\left(X_i Y_i\right)} \times log \left(\right. fitness \left.\left(Y_i\right)\right)+b, 0 \leq b \leq 0.5$    (30)

Step 2: Random generation.

Eq. (31) illustrates the calculation of distance.

${Dist}\left(X_i Y_i\right)=\sqrt{\sum_{j=1}^n\left(X_j\left(X_i\right)-X_j\left(Y_i\right)\right)^2}$    (31)

If b=0.5, O increases. The value of b toward the apex of the Anopheles motion was the maximum. If b=0, distance and optimization are used to compute the motion of the Anopheles mosquito.

Step 3: Fitness function.

The fitness function is represented in Eq. (32):

Fitness function $\left(X_i Y_i\right)=Maximixe (Accuracy), Minimize (Error rate)$    (32)

where, $\beta$ represents the accuracy and $\varphi$ represents the error rate.

Step 4: ASA updates are aimed at enhancing the accuracy and reducing the error rates. This optimization process calculates parameters by assessing the scent density detected by each mosquito and advancing toward the optimal point based on these values. The focus was on improving the accuracy while minimizing the error rates, as depicted in the following maximization equation:

Maximixe (Accuracy) $\left(X_i\right), X_i \in x_i, i=1,2, \ldots . N$    (33)

where, Maximixe (Accuracy) (X_i), represents the objective function. If the decision variable is continuous, it is divided by an upper bound ($U_{X_i}$) and a lower bound ($L_{X_i}$), as shown in Eqs. (34) and (35).

$L_{X i} \leq X_i \leq U_{X i}$    (34)

$X_i \in\left\{X_1, X_2 \ldots \ldots X_n\right\}$    (35)

The minimized error rate is given as:

Minimize $($ Error rate $)\left(Y_i\right)=X_1^2+\left(X_2-1\right)^2$    (36)

Eqs. (31) and (34) are the optimization equations for achieving the desired function. The smells of anopheles were used to optimize the parameters.

Step 5: Termination.

Finally, the fitness function is employed to achieve optimum accuracy while reducing the computational time. Finally, the ASA mechanism selects the optimal value of the AlexNet classifier to effectively classify input MR images as normal or abnormal.