1. Introduction

An unmanageable growth of abnormal tissues in the brain is called a brain tumor, which is also known as an intracranial tumor ^[1]. Brain tumors can be classified as primary brain tumors and secondary brain tumors. Primary brain tumors, also known as benign or non-cancerous, arise in the brain. On the other hand, secondary brain tumors, also known as malignant or cancerous, start from different parts of the body, such as the stomach, breasts, lungs, and kidneys, and relocate to the brain ^[2]. The symptoms of brain tumor-affected patients are headaches, cognitive complications, walking or balance difficulty, nausea and vomiting, muscular contractions, visual and hearing impairments, recession, dysphagia, mental disorders or confusion, etc. ^[3].

There are various forms of brain tumors in the world, notably the most recognizable brain tumors are glioma, meningioma, and pituitary tumor. Glioma tumor is situated in the glial cells, which have a higher grade or lower grade, frequently known as astrocytes, and the cerebral hemispheres. On the contrary, the gland that is placed on the front fragment of the brain is called the pituitary gland. Meningioma is formed at the outer layer of the brain. Pituitary and meningioma are identified as benign, whereas glioma is recognized as malignant ^[4]. Benign tumors can be cured by regular medications or surgery; otherwise, malignant tumors can be managed by radiotherapy/chemotherapy ^[5].

Early diagnosis and categorization of brain tumors play a vital role in the mortality rate of patients. It can be diagnosed by various types of manual techniques such as Magnetic Resonance Imaging (MRI), Positron emission tomography (PET), and computed axial tomography (CT) scans ^[6]. MRI is rarely used to diagnose various types of tumors, but it depends on human interpretation, and it may be difficult to analyze with a huge amount of data. Misclassification of brain tumors can often lead to the wrong medication for the patient, which can be life-threatening. Another diagnosis procedure is biopsy, which involves brain surgery ^{[7, 8]}. In decade ago, to prevent surgery and error-free diagnosis, many authors established automated methodologies to recognize tumors from MRI scans. Machine learning techniques are developed to detect and classify various types of brain tumors from MRI. Some related works are attached to this paper. The Authors, Thejaswini et al., suggested a procedure where Adaptive Regularized Kernel-based Fuzzy C-means clustering (ARKFCM) is used for segmenting the input images. After that, a combination of Artificial Neural Network (ANN) and Support Vector Machine (SVM) is used to detect and classify brain tumors with an accuracy of 91.4%, a Sensitivity of 98%, and a Specificity of 78%, respectively ^[9]. Ei Dahshan et al. proposed hybrid intelligent techniques for MRI brain images classification where features are extracted using Discrete wavelet transform (DWT), then features are reduced using Principal Component Analysis (PCA), and finally, features are classified using Feedforward Backpropagation Artificial Neural Network (Fp-ANN) and K-Nearest Neighbor (KNN) ^[10]. The main drawback of this procedure is that ANN cannot extract the features automatically.

To solve this problem, a decade ago deep learning network was introduced. The author, G Hemanth, deployed distinct techniques which include data collection, pre-processing, average filtering, segmentation, feature extraction, and CNN via classification and identification ^[11]. Sajjad M presented a deep-learning CNN method with augmented data. Pre-trained model VGG-19 CNN design is organized for classification with an accuracy of 87.39% ^[12]. Sultan organized an architecture with a deep learning model, CNN, with a classification accuracy of 96.13% ^[13]. Ginni Garg et al established a classification approach with stationary wavelet transform (SWT), GLCM, and a hybrid classifier ^[14].

Reza proposed a Deep Tumor Net model, which is used for adjusting the layers of GoogleNet and as an activation layer Leaky Relu layer is used ^[15]. Afshar et al. established a novel CNN architecture, which is known as the Capsule network, and achieved an accuracy of 90.89% ^[16]. Habib proposed a convolutional neural network that uses the Kaggle binary brain tumor classification dataset I, which is employed in this paper, to detect brain malignancies. This method makes use of a modified neural network architecture to obtain an accuracy of 88.7% ^[17].

The accuracy rate is high, but the pre-trained model trains the parameters within a large number of layers, so it takes a high computational time. Some Deep machine learning models like CNN can extract the features automatically within a lightweight design, but the drawback of this procedure is that it can extract unknown features randomly, so it causes overfitting problems.

To solve the recommended problems, a novel Deep Parallel Convolutional Neural Network (DPCNN) for detecting and classifying the various kinds of brain tumors from MRI scans, utilizing different classifiers to address the issues raised.

The significant contributions in this study are given below:

• A novel lightweight DPCNN model is developed to extract the features locally and globally from MRI scans. In this work, parameters are minimized by using Global average pooling instead of a fully connected layer to reduce the overfitting problem.

• To increase the size of the dataset by using geometric transformation of the original data, data augmentation is used. It increases the variability and diversity of the training dataset, which reduces the overfitting problems and leads to better performance.

• Three classifiers are introduced to compare the effect of different classifiers, such as SoftMax, KNN, and SVM.

• The proposed system is verified against various types of pre-trained models like Googlenet, VGG16, VGG19, ResNet18, and ResNet50 during the training of the dataset.

• The proposed system is optimized, which is computationally effective and yields the best result.

Following the introduction, the model is trained with performance metrics represented in the Methods and Materials section, which are reported in Section 2. The outcome of the proposed method is analyzed and compared with alternative approaches are presented in the Results and Discussion section, which is reported in Section 3. Finally, the paper is concluded in Section 4.

2. Methods and Materials

The flowchart of the proposed method for detecting and classifying the various types of brain tumors from MRI scans is shown in Fig. 1, which includes three crucial steps. First data pre-processing step, second feature extraction step, and finally the classifier step. In the pre-processing step data are resized without removing any content, then grayscale conversion is performed, and finally data are augmented. The second step is the feature extraction step, in this step hyperparameter selection is performed, and a lightweight PDCNN is used to extract the features locally and globally. The third and final step is the classifier step, where different types of brain tumors are classified by various types of classifiers such as SoftMax, KNN, and SVM.

Fig. 1. Flowchart of the proposed method.

2.1. Data Acquisition

A dataset is the assemblage of images that are used to train or analyze the model. In this model, datasets are collected from open-source “Kaggle” ^[18]. The MRI images of this dataset contain four types of brain images, such as glioma (826 images), meningioma (822 images), no tumor (395 images), and pituitary (827 images). Original images of the Kaggle dataset are at a size of $256 \times 256$. Fig. 2 presents the four forms of brain tumor images of the Kaggle dataset. Accordingly, 80% data are loaded for training and 20% data are loaded for testing.

Fig. 2. Four types of brain tumor images from the Kaggle dataset.

2.3. Feature Extraction Step

The technique of extracting features from data is called feature extraction. The hyperparameter selection part and the lightweight DPCNN part are the two components included in this step.

2.3.1 Hyperparameter selection

2.3.1.1 Activation function An activation function in machine learning is a mathematical operation that works with the non-linearity of the model. It works in the output of each neuron. It grips the weighted sum from the first layer, produces the output, and finally proceeds to the next layer. There are different forms of activation functions like Rectifier Linear Unit (ReLU), Leaky ReLU (LReLU), and Clipped ReLU (CReLU).

The activation function ReLU is commonly incorporated after the convolutional layer to reduce gradient diffusion and generate better performance ^[20]. The equation of the ReLU function is ^[21]:

$f(g) = \max(0, g).$

In this equation, “$g$” indicates the input value. When the value of $g$ is negative or zero, it represents that the output is zero. When the value of $g$ transcends zero, it represents 1. But the drawback of this function is that it always works with positive input; if it gets a negative value, the learning process will be stopped.

Clipped ReLU presents as a threshold operator that involves upper and lower thresholds for the output. The equation of clipped ReLU is

$f(g) = \min(\max(g, 0), c).$

The problem with Clipped ReLU is that it introduces smoother learning. This function needs tuning, but it is not possible to do so exactly; it affects the learning performance.

Leaky Relu (LReLU) is an enhanced variant of ReLU and Clipped ReLU that can deal with the issues and provide the best performance. The equation of Leaky ReLU is

$f(g) = \begin{cases} g, & g > 0, \\ \alpha g, & g \le 0. \end{cases}$

2.3.1.2 Pooling operation A pooling layer is utilized to downsample and reduce the spatial dimensions of the feature map of the convolutional neural network. The two most common pooling operations in the CNN are Max pooling and Average pooling.

Max pooling considers the maximum value from the subset. Some important information is ignored using Max pooling since only the strongest feature elements are considered.

To eliminate this problem, Average pooling is used. Average pooling takes the average value from the subset. Although Average pooling selects all the elements, they are always treated equally and provide contributions from all essential elements. It extracts features more smoothly than Max pooling ^[22].

2.3.1.3 Optimization algorithms Optimizers are algorithms that are used to change the strategies of weights and learning rates and reduce losses. There are various types of optimizers; among them, ‘sgdm‘, ‘adam‘, and ‘rmsprop‘ are used in this paper. RMSprop provides a slow learning rate, and Adam has poorly tuned hyperparameters. To overcome this problem, sgdm optimizer is used, which has a high variance of the parameters and reduces losses ^[23].

2.3.1.4 Normalization selection The normalization layer is used for scaling the features of the input data in a range between the mean of 0 and a standard 1. Cross-channel normalization works with different channels to normalize local reactions for each activation. It introduces computational complexity when the relationship between channels is complex. Batch normalization is used to enhance the problem by normalizing the features in intermediate layers of deep neural networks. It reduces overfitting problems, fast training, and more stability, and improves accuracy.

The formula of batch normalization:

$\hat{x} = \frac{x - \mu}{\sqrt{\sigma^2 + \epsilon}}.$

2.3.1.5 Output layer selection A fully connected layer is also recognized as a dense layer that connects every node or neuron from the previous layer to the final layer. The features from the final convolution layer are flattened to one vector. It leads to a large number of parameters, so the overfitting problem and computational complexity arise. Global average pooling is used instead of a fully connected layer. Without flattening global average pooling, calculate the average value of each feature map and produce a single value per channel. It reduces the number of parameters, prevents overfitting, and provides the best performance ^[24].

2.3.2 Lightweight PDCNN architecture

In present times, CNN has a great impact on deep learning networks. The CNN can extract features and classify them from MRI scans automatically, and provide a great performance. Notably, the drawback of CNN is that it cannot extract unknown local and global features randomly. To eliminate the problem, this research proposes a lightweight DPCNN where two pathway CNNs are used to extract the known features locally and globally to detect and classify the brain tumor images. Two pathways, a local pathway and the global pathway, within 23 layers.

The local pathway consists of three convolutional layers. The kernel size of the first convolutional layers in the local pathway is $3 \times 3$. The kernel size of the second and third convolutional layers is $5 \times 5$ and $7 \times 7$, respectively. The global pathway consists of three convolutional layers. The kernel size of the first convolutional layers in the global pathway is $12 \times 12$. The kernel size of the second and third convolutional layers is $15 \times 15$ and $17 \times 17$, respectively. The stride size of this architecture is (1,1), and the same padding is used. Although the Batch normalization (BN) and the activation function Leaky ReLU (LReLU) layer are used to remove overfitting.

Down-sampling the dimension of the output feature map, an average pooling layer is used with a kernel size of ($3 \times 3$). Both pathways are connected by a concatenation layer. Global average pooling is used instead of a fully connected layer. Various types of classifiers, such as SoftMax, KNN, and SVM, are used to classify the brain tumor images, and finally, DPCNN-SVM provides the best accuracy. Fig. 3 shows the diagram of proposed model.

Fig. 3. Diagram of the proposed model.

3. Experimental Results and Discussion

The key point of this proposed framework is to detect and classify the brain tumors from MRI scans of the dataset, which carries 3459 samples of glioma, meningioma, pituitary, and no tumor. Primarily, the size of the dataset is changed from ($256 \times 256$) to ($60 \times 60$) by using resize, then grayscale conversion is used, and finally, the data is augmented to enhance the number of images to improve the model’s generalization. Features are extracted by using a lightweight DPCNN and classified the features by various types of classifiers like SoftMax, KNN, and SVM, respectively. During the dataset, 80% is loaded for training and 20% is loaded for testing. The model is trained according to the epoch (15, 20, 25) and the learning rate (.001, .002, .001). Finally provides the best accuracy at 97.7% with the epoch 20, learning rate of .002, and the proposed model DPCNN-SVM. All the experiment is carried out in MATLAB 2020. The system specification includes AMD Ryzen 5, 8GB of Memory, AMD Radeon Graphics, and Windows 10 Home 64-bit.

3.3. Pre-trained Model

A pre-trained model is a machine learning model that can be enhanced for a particular task and has been trained on a large dataset. Machine learning models are advanced using pre-trained models as a foundation since they offer a collection of baseline weights and biases that may be modified for a particular task. Several pre-trained models are used in this research, such as Google Net, VGG16, VGG19, ResNet18, and ResNet50, to compare the performance of the proposed model by changing the last four layers, including input layer, the fully connected layer, the SoftMax layer, and the output layer. The $60 \times 60$ image size is used as input image for all the pre-trained models to train. Fig. 4 shows the comparison proposed model with the pre-trained model.

Several pre-trained models such as Google Net, VGG16, VGG19, ResNet18, and ResNet50 provide performance accuracy of 93.90% 94.80%, 95.30%, 96%, and 96.8% respectively. Those models have a large number of layers, so faces are computationally complex and require a long time due to the large number of parameters. To address these issues, the suggested lightweight DPCNN architecture provides the best accuracy of 97.7%.

Fig. 4. Comparison proposed model with the pre-trained models.

4. Conclusion

The early and precise identification and classification of different kinds of brain tumors can prevent fatal diseases. Considering these issues, an effective lightweight Deep Parallel Convolutional Neural Network (PDCNN) is developed, which is capable of precisely detecting and classifying a wide range of tumor forms, including glioma, meningioma, no tumor, and pituitary tumors. The proposed model is designed based on the examination of the results of the experiments by tuning the hyperparameters like epochs, learning rates, pooling layers, activation layers, optimizing algorithms, and lastly output layers. Initially, resized data is augmented by using geometric transformation operators, then augmented data is fed to the DPCNN to extract the known features both locally and globally, and lastly, different kinds of classifiers like SoftMax, KNN, and SVM are used to classify them. The accuracy of DPCNN-SoftMax, DPCNN-KNN, and DPCNN-SVM is 95.5%, 93.4%, and 97.7% accordingly, and among them, lightweight DPCNN-SVM provides the highest accuracy of 97.7%. The suggested model, compared with the pre-trained model, has the least amount of computing parameters and time. Future studies can attempt to develop an attention mechanism and expand studies on the other types of cancers, like stomach cancer, liver cancer, breast cancer, etc.