1. Introduction
bnormal cell growth in any part of the body is called a tumor. In general, brain tumors are divided into two types: benign and malignant. This most widely used tumor grading system has been proposed by the World Health Organization (WHO) [1]. Neurologists play an important role in the evaluation and treatment of brain tumors. When a brain tumor is clinically diagnosed, a neurological evaluation should determine the location of the tumor and its relationship to surrounding structures. This information is very vital for choosing the best treatment method, including surgery and radiation therapy, but doctors always have difficulty in accurately diagnosing the location of the tumor. This problem is usually due to excessive fatigue of doctors, high image artifact, and so on. Given this, a computer-based intelligent diagnostic system can help neurologists make the correct diagnosis. The use of intelligent medical diagnostic systems as an assistant to physicians and radiologists, besides helping them, can pave the way for accurate and error-free identification and distinguishing these diseases from other similar diseases [2].
In recent years, the use of diagnostic systems based on the deep learning method has been widely used because of its high efficiency, so many studies have been conducted in this field. For example, Gupta and Khanna [3] conducted a study entitled “A non-invasive and adaptive CAD system to detect brain tumor from T2-weighted MRIs using customized Otsu’s thresholding with prominent features and supervised learning.” They used non-homogeneous techniques for preprocessing brain MR images segmented by Otsu’s thresholding technique. In the next step, several feature extraction methods such as Tamura, LBP, Gabor filters, GLCM, and Zernike were applied to segmented images. Then, from the used methods, two prominent samples were selected using entropy measures. Finally, a support vector machine (SVM) was used for the classification. They have achieved 98% accuracy and 100% sensitivity. 
Zikic et al. [4] developed an interpretive method for converting 4D data, such that standard 2D convolutional neural network (CNN) architectures can also be used to perform brain tumor segmentation. The results reported on the BRATS dataset show a dice score of 83.7% for the whole tumor region, 73.6% for the core tumor region, and 69% for the active tumor region. To classify brain lesions in MRI images of the breast, Amit et al. [5] used a VGG network and the extracted features were classified using SVM. The reported accuracy on a database of 123 MRI images, without the use of color channels, is 73% with a sensitivity of 77% and a specificity of 68%. Using a simple SegNet network, Korfiatis et al. [6] performed segmentation on the BRATS data and reported an average dice accuracy of 87.6%. Sajid et al. [7] used deep learning networks to diagnose brain tumors using BRAST database images and different MRI images. They reported that their method was successful with very good simulation results. 
One of CNN’s new approaches is to evaluate the performance of brain tumor diagnosis using deeper architectures [8] by implementing 3 × 3 filters in convolutional layers. Using this method, more convolutional layers can be added to the architecture, without reducing the impact of the acceptance field of previous larger filters. Besides, deeper architectures use more nonlinearity and have less filter weight because they use smaller filters which reduce the likelihood of preprocessing. A modified version of the rectified linear unit (ReLU) called leaky ReLU (LReLU) has been used as a nonlinear activation function. CNN’s proposal, which has 11 deep layers on the BRATS suite, achieves a dice score of 88%, 83%, and 77% for the overall, core, and active tumor regions, respectively. Pereira et al. [9] used CNN for the segmentation of MRI images. Their proposed network was based on a U-net which has two upsampling and expansion paths and used LReLU with α = 0.3 as the activation function. After each layer of convolution with 3*3 kernel size and LReLU activator, a dropout = 0.2 technique was used. The proposed architecture was used for both segmentation of the whole binary tumor and segmentation within the multiclass tumor. 
The only difference is the number of kernels of the last layer, which is proportional to the number of classes. The cross-entropy function is introduced as a network error function and is used to optimize SGD information with a learning rate of 0.01. Havaei et al. [10] presented a fully automated brain tumor segmentation method based on deep learning. First, they discussed different types of architectures based on CNN. Their network examines both local and global contextual features simultaneously. It used the last layer, which is a convolutional simulation of the fully-connected (FC) layer, and allowed a 40-fold increase in speed. They also introduced a two-phase learning process that allowed us to address the problems associated with unbalanced tumor segmentation labels. Finally, they examined a cascaded architecture in which the base CNN output was considered as a source of additional information for a subsequent CNN. The results reported on the 2013 BRATS dataset showed that their architecture was 30 times faster than existing optimized CNNs. 
Pereira et al. [11] performed brain tumor segmentation using full CNNs. The network architecture for HGG consisted of 11 layers with a max pooling layer followed by three convolutional layers and finally three FC layers, while the network architecture for LGG consisted of 9 layers with one max pooling layer after two convolutional layers and finally three FC layers. The HGG architecture is deeper than that of LGG; hence, a dropout of 0.5 was used in LGG, while in HGG it was 0.1. This technique was used only in FC layers. LReLU activator function was used in all layers, while the softmax function was applied for the last FC layer and padding technique for convolutional layers. During learning, the categorical cross-entropy error function was used with the SGD optimizer. Amin et al. [12] proposed a new CNN architecture based on DDN for the diagnosis of brain tumors. Seven layers were used for classification: 3 convolutional layers, 3 ReLU layers, and one softmax layer. The input MR images were first divided into several segments and then the center pixel value of each segment was presented to the DNN. 
Extensive experiments were performed using 8 large-scale benchmark datasets, including BRATS 2012, 2013, 2014, 2015, and ISLES 2015, 2017. Dong et al. [13] used a U-Net based fully convolutional network for brain tumor detection and segmentation. They used a 28-layer CCN architecture and could achieve the desired accuracy.
As mentioned above, various factors affect the identification and classification of tumors such as imaging format and noise, physician fatigue, etc., so it is necessary to develop an intelligent diagnostic system for the classification of tumors. In this paper, we used a deep learning method based on a 2D CNN with high accuracy to detect brain tumors from MRI images. The second section introduces the proposed computer-aided detection (CAD) system. The results and discussion are presented in the third and fourth sections, and finally, in section 5 we present the conclusions.
2. Materials and Methods
Studies on MRI images of 3 types of brain tumors were collected. Figure 1 shows a block diagram of the study method.
 
MRI images were first entered as input to the proposed network. Previously, traditional methods were used to extract features and classify them. In this study, feature extraction and classification were performed by deep learning methods in an integrated way based on a 2D deep convolutional network. Its properties are presented in the next sections.
Input data
Standard MRI images are used to diagnose brain tumors. MRI imaging has different protocols, of which T1, T2, and FLAIR protocols are commonly used for research. In this regard, we used 1.5 Tesla MRI images with the T1 protocol. For these images, the database presented in reference [14] was used. This database contains images of 3 different types of brain classes in a dimension of 512 × 512 related to meningiomas (708 images), gliomas (1426 images), and pituitary tumors (930 images). Figure 2 illustrates three images from three classes of this database.
The dimension of these images was changed from 512 × 512 to 128 × 128 for use in the proposed method.
Proposed convolutional network
Training of a network minimizes the error function based on real network outputs compared to optimal network outputs. This procedure is done by modifying the free parameters of the network, including weights and biases. The training method used in the proposed network is supervisor training. In this method, a supervisor monitors the learner’s behavior and reminds it how to function properly. In other words, the learning system is a set of data pairs consisting of network inputs and favorable outputs. After applying the network input, its output is compared with the favorable output and the learning error is calculated and used to modify the network parameters so that if the same inputs are reapplied to the network, the network output will be closer to the favorable output. In this method, images are applied to the network in batches of 64; although it requires more memory, it will have higher stability. The proposed network architecture is shown in Figure 3.
The most important layers of this architecture are briefly described in the next section.
Convolutional layers
Convolution operation in the image is such that the filter with the desired size convolves over the original image. In this way, at any moment, the filter matrix arrays are placed on the pixel arrays of the image and multiplied. Finally, the results of all parts are summed up and one number is obtained for each filtered part. In the next step, the filter of one unit goes to the right and the operation is repeated. This operation continues until all pixels of the image are swept. Finally, a new matrix is produced. In this study, 3 convolutional layers with the ReLU activation function are used; there are 32 filters with a 3×3 kernel size in each layer [15].
Pooling layers
Pooling layers are placed at regular intervals between successive convolutional layers. This technique is very common in a convolutional architecture and can be used to reduce the size of the network feature map and its parameters. In other words, the function of pooling layers is to reduce the spatial size of the image to reduce the number of parameters and calculations within the network and, thus, control overfitting. The most common way is to use them with 2×2 filters and max pooling [16]. Table 1 shows the pooling layers of the proposed network. 



Flattening layer
The flattening layer is an important layer that is placed between the layers that perform convolutional feature extraction and output classification. This layer converts the data that intends to enter the classification stage into a vector.
FC layers
After convolutional layers, there are the last FC layers. These layers act like their counterparts in traditional artificial neural networks and comprise almost 90% of the parameters of a CNN. The FC layer allows us to present the network result in the form of a vector with a specified size. This vector can be used to categorize images or continue further processing [17]. In this regard, different classification algorithms are used, one of which is the softmax activation function. In mathematics, the softmax function or normalized exponential function is a generalization of the logistic function. This function takes a vector z of K real numbers as input and gives a vector σ (Z) of K real numbers (0, 1) as output whose sum of the components is 1 (see Formula 1). 



It is mainly used in the field of mathematics, especially in the theory of probability and related fields. Classification by softmax function has a unique advantage over N-dimensional vectors. To classify extracted vectors In deep learning, it determines the probability of the extracted vectors and then classifies them. The softmax function is a form of logistic regression. The main idea of logistic regression is to use the logical regression method in classification, which judges the input data and gives a single discrete output [18]. The softmax method has only one drawback i.e., its high computational complexity. However, this problem has been solved with an improvement in the GPU. Owing to the mentioned advantages, we used the softmax function in the FC layers.
Dropout layer
The dropout layer is used in FC layers and between them to prevent the network from overfitting and avoid the complexity of FC layers (Figure 4). 
 
In this study, two dropout layers were used and the rate of each layer was set to be 0.5.
Performance evaluation
The performance of a classification system is measured by various parameters such as accuracy, precision, sensitivity, and F1 score. Using these parameters allows users to understand how well a model performs in analyzing textual data. To evaluate the performance of a classification system, a fixed test dataset (a set of textual data with predefined size whose labels are specified) can be used. Such a process in the evaluation phase divides the training data into two subsets: the first subset is used to train the model and the second subset is used to test the model. 
Classification accuracy defines the number of correct predictions made by a classifier divided by the total number of predictions made by that classifier (Formula 2): 



In the above Formula, TP is true positives, TN stands for true negatives, FP is false positives, and FN stands for false negatives.
Classification precision defines the ratio of the number of correct predictions made for samples of a particular class to the total number of predictions for samples of the same class (including all correct and incorrect predictions) (Formula 3): 



Classification sensitivity refers to the number of image data correctly classified in a particular class to the total number of data that should be classified in that class (Formula 4): 



The F1 score combines precision and sensitivity parameters to determine how well a classification model performs (Formula 5): 



3. Results
In the proposed method, 80% of the database images were selected for training the model and 20% for testing the model, while there is no interference between training and test data. Training data with a batch size of 64 were sent to the network; this network was trained in 100 iterations. The proposed model performed better than other methods by using hierarchical learning and extracting high-level features. A computer with a Corei7-6700HQ processor, NVIDIA 1060-6G byte graphics card, 8G byte DDR4 RAM, and 1TB hard drive was used to implement the proposed deep learning network codes. To run the program in Windows, Anaconda software version 3.7, Spyder software, and TensorFlow machine learning platform were used. Table 2 presents the mean overall accuracy and other parameters of the proposed method per 5 times of program execution.



The proposed method was compared with several similar studies presented in Table 3. 



4. Discussion
This study presents a multi-class classification method for MRI images of brain tumors using a deep learning method. The used images were T1 type and have three types of brain tumor classes: 708 images of meningioma tumors, 1426 images of glioma tumors, and 920 images of pituitary tumors. The proposed method includes a main classification stage using a 12-layer CNN consisting of three convolutional layers, three max pooling layers, one flattening layer, two dropout layers, three FC layers, and the softmax activation function was used to classify the three classes. The average accuracy of the proposed method after 5 times running was 98.68%. The proposed system can be used for the detection of brain tumors. It was compared with several similar studies that mostly used the image zoning technique and the accuracy of our method was found much higher than theirs.
One of the strengths of this study was the use of a 12-layer deep CNN, which increases the network accuracy to 98.68%, while one of its drawbacks was the lack of using clinical MRI images with different classes to diagnose multiple brain tumors simultaneously.
5. Conclusion
The brain is the most important organ in the body that controls all other parts of the body. If there is a disease in this part, disorders can occur in other parts which sometimes can lead to death. In this study, an automated CAD system based on deep learning was presented to diagnose and classify brain tumors. Since in the proposed method, high-level features were extracted by deep learning, the accuracy of classification and detection was very high and the size of the feature vector also decreased. Because of the large number of images in the used database, the network was designed to provide the highest execution speed and accuracy.
Ethical Considerations
Compliance with ethical guidelines
This study was extracted from a thesis. No ethical approval was needed.
Funding
This study received no financial support from any organization
Authors' contributions
All authors had an equal contribution in preparing this paper.
Conflicts of interest
The authors declare no conflict of interest.
Acknowledgements
The authors would like to thank the Deputy for Research of the Islamic Azad University of Ferdows Branch for their cooperation.


References

1. Davis FG, Freels S, Grutsch J, Barlas S, Brem S. Survival rates in patients with primary malignant brain tumors stratified by patient age and tumor histological type: An analysis based on Surveillance, Epidemiology, and End Results (SEER) data, 1973-1991. Journal of Neurosurgery. 1998; 88(1):1-10. [DOI:10.3171/jns.1998.88.1.0001] [PMID]
2. El-Dahshan ESA, Mohsen HM, Revett K, Salem ABM. Computer-aided diagnosis of human brain tumor through MRI: A survey and a new algorithm. Expert Systems with Applications. 2014; 41(11):5526-45. [DOI:10.1016/j.eswa.2014.01.021]
3. Gupta N, Khanna P. A non-invasive and adaptive CAD system to detect brain tumor from T2-weighted MRIs using customized Otsu’s thresholding with prominent features and supervised learning. Signal Processing: Image Communication. 2017; 59:18-26. [DOI:10.1016/j.image.2017.05.013]
4. Zikic D, Ioannou Y, Brown M, Criminisi A. Segmentation of brain tumor tissues with convolutional neural networks. Paper presented at: MICCAI Workshop on Multimodal Brain Tumor Segmentation Challenge (BRATS). September 2014; Boston, MA, USA. https://www.researchgate.net/publication/303703706
5. Li BN, Chui CK, S. Chang S, Ong SH. Integrating special fuzzy clustering with level set methods for automated medical image segmentation. Computers in Biology and Medicine. 2014; 41(1):1-10. [DOI:10.1016/j.compbiomed.2010.10.007]
6. Korfiatis P, Kline TL, Erickson BJ. Automated segmentation of hyperintense regions in FLAIR MRI using deep learning. Tomography. 2016; 2(4):334-40. [DOI:10.18383/j.tom.2016.00166] [PMID] [PMCID]
7. Iqbal S, Ghani MU, Saba T, Rehman A. Brain tumor segmentation in multi-spectral MRI using Convolutional Neural Networks (CNN). Microscopy Research and Technique. 2018; 81(4):419-27. [DOI:10.1002/jemt.22994]
8. Dvořák P, Menze B. Local structure prediction with convolutional neural networks for multimodal brain tumor segmentation. In: Menze B, Langs G, Montillo A, Kelm M, Müller H, Zhang S, et al, editors. Medical Computer Vision: Algorithms for Big Data. MCV 2015. Lecture Notes in Computer Science. Vol. 9601. Cham: Springer; 2016. pp. 59-71. [DOI:10.1007/978-3-319-42016-5_6]
9. Pereira S, Pinto A, Alves V, Silva CA. Brain tumor segmentation using convolutional neural networks in MRI images. IEEE Transactions on Medical Imaging. 2016; 35(5):1240-51. [DOI:10.1109/TMI.2016.2538465] [PMID]
10. Havaei M, Davy A, Warde-Farley D, Biard A, Courville C, Bengio Y, et al. Brain tumor segmentation with deep neural networks. Medical Image Analysis. 2017; 35:18-31. [DOI:10.1016/j.media.2016.05.004] [PMID]
11. Pereira S, Oliveira A, Alves A, Silva CA. On hierarchical brain tumor segmentation in MRI using fully convolutional neural networks: A preliminary study. Paper presented at: 2017 IEEE 5^th Portuguese Meeting on Bioengineering (ENBENG). 16-18 Feb 2017; Coimbra, Portugal. [DOI:10.1109/ENBENG.2017.7889452]
12. Amin J, Sharif M, Yasmin M, Fernandes SL. Big data analysis for brain tumor detection: Deep convolutional neural networks. Future Generation Computer Systems. 2018; 87:290-7. [DOI:10.1016/j.future.2018.04.065]
13. Dong H, Yang G, Liu F, Mo Y, Guo Y. Automatic brain tumor detection and segmentation using U-Net based fully convolutional networks. Paper presented at: Annual Conference on Medical Image Understanding and Analysis. 11-13 July 2017; Edinburgh, United Kingdom. [DOI:10.1007/978-3-319-60964-5_44]
14. Cancer Net. Brain tumor [Internet]. 2020 [Updated 2020]. Available from: https://www.cancer.net/cancer-types/brain-tumor/statictics
15. Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Communications of the ACM. 2012; 60(6):84-90. [DOI:10.1145/3065386]
16. Giusti A, Cireşan DC, Masci G, Gambardella LM, Schmidhuber J. Fast image scanning with deep max-pooling convolutional neural networks. Paper presented at: IEEE International Conference on Image Processing. 15-18 Sep 2013; Melbourne, VIC, Australia. [DOI:10.1109/ICIP.2013.6738831]
17. Thakkar V, Tewary S, Chakraborty C. Batch normalization in convolutional neural networks-A comparative study with CIFAR-10 data. Paper presented at: 5^th International Conference on Emerging Applications of Information Technology (EAIT). 12-13 Jan 2018; Kolkata, India. [DOI:10.1109/EAIT.2018.8470438]
18. Sainath TN, Vinyals O, Senior A, Sak H. Convolutional, long short-term memory, fully connected deep neural networks. Paper presented at: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 19-24 Apr 2015; Brisbane, QLD, Australia. [DOI:10.1109/ICASSP.2015.7178838]