Brain tumor magnetic resonance image classification: a deep learning approach

From the past decade, many researchers are focused on the brain tumor detection mechanism using magnetic resonance images. The traditional approaches follow the feature extraction process from bottom layer in the network. This scenario is not suitable to the medical images. To address this issue, the proposed model employed Inception-v3 convolution neural network model which is a deep learning mechanism. This model extracts the multi-level features and classiﬁes them to ﬁnd the early detection of brain tumor. The proposed model uses the deep learning approach and hyper parameters. These parameters are optimized using the Adam Optimizer and loss function. The loss function helps the machines to model the algorithm with input data. The softmax classiﬁer is used in the proposed model to classify the images in to multiple classes. It is observed that the accuracy of the Inception-v3 algorithm is recorded as 99.34% in training data and 89% accuracy at validation data.


Introduction
The developments in the medical field support the medical practitioners to facilitate the patients effectively. With the utilization of artificial intelligence (AI) in the health care helps the medical domain to serve more to the patients (Siegel et al. 2015). According to the statistics of 2019, most of the deaths in the world are happened due to the cardiovascular diseases and in the next place the cancer diseases are occupied (Siegel et al. 2017). Brain tumor disease is one of the life threating diseases in the world (Association 2020). Magnetic resonance imaging (MRI) is one of the safest imaging techniques that extracts the good images and helps in the process of medical diagnosis. Many researchers are focused on improving the quality of the MR images and also to develop the new methods for quicker and easy medical diagnosis from the MR images.
This study concentrated on the brain tumor detection from MR images (with multiscale Two-Pathway-Group conventional neural networks'' 2019; Naseer et al. 2019).
The major issues related to the manual diagnosis process are time consuming, interpretation is difficult and costly. To overcome these issues, there is a necessity for computer aided diagnosis (CAD) tool. In the recent years, many automated CAD approaches for brain tumor detection were developed (Zuo et al. 2017). Approximately, they are classified into two categories: supervised and unsupervised mechanisms. In most of the cases, supervised mechanisms are performed better when compared against the unsupervised mechanisms. Major brain tumor detection techniques are depend on the segmentation and less importance are given to the feature extraction and classification. This will improve the performance of the CAD process in the MR images. Deep learning mechanism is one of the efficient approaches to classify the tasks. Recently, deep learning approach is employed in the medical diagnosis for classifying the abnormal brain images (Charron et al. 2018). Deep learning approaches are important in identifying many diseases like lung cancer and breast cancer (Cheng et al. 2020).
The traditional machine learning techniques are considered as the basics for the proposed deep learning approach. The drawbacks of the prediction models motivate the researchers to focus on the new machine learning mechanisms to improve the accuracy of the image classification. Deep learning is considered as the subset of the machine learning which became more popular with respect to effective prediction from extensive data of text or image. At another side, deep learning is efficient in finding the results from the larger dataset (Gu et al. 2018). In medical image processing, the deep learning approaches are considered for identifying the disease effected part in the images. The applications of deep learning approach are not limited to image processing and it can be applied to much object recognition, activity monitoring and many more. The accuracy of the prediction models in deep learning approach depends on the sample data set and its training mechanism which leads to better results. Transfer learning is applied to the deep learning models to reduce the limitations of training the data samples. Transfer learning is a mechanism where the larger dataset features are deployed to the smaller dataset to train them with extracted features (Youse et al. 2018).
This paper concentrated on developing the deep learning approach for brain tumor detection in MR images. Section 2 deals with the related work regarding the deep learning mechanisms in brain tumor detection. Section 3 explains the convolution neural network and Inception v3 models. Section 4 deals the brain tumor dataset and their properties. Section 5 explains about the result analysis with respect to accuracy. Finally, Sect. 6 concludes the research work.

Literature survey
Machine learning mechanisms play a crucial role in different fields like preventive medicine and medical diagnosis. Only small part of work has been carried in brain tumor detection using the MR images. Most of the researchers concentrated only on the traditional machine learning approaches to deal with MR images. Some of the algorithms used the deep learning mechanisms for brain tumor detection.
In (Rehman et al. 2019), the authors developed the framework called as convolution neural network (CNN) to deal with classification of different type of brain tumor datasets such as VGGNet, AlexNet and GoogLeNET. The proposed algorithm identifies the region of interests in the brain MR images. It also performs the fine tuning process to the image data set for further classification. The authors achieved 98.6% accuracy in classification of brain tumor images by applying the VGGNet architecture.
In (Deepak and Ameer 2019), the authors proposed the classification mechanism by using the deep learning approach. The image features are extracted and used in the aggregation and classification process. They used fivefold mechanism for classification at the patient's side. The proposed mechanism achieved 98% accuracy in the image classification. This work stated that the automated classification models achieved better accuracy compared against the manual classification mechanisms. Another classification mechanism is proposed (Afshar et al. 2018) using the capsule network method (CapsNets). This method improves the classification accuracy in the brain tumor MR images by modifying the convolution layer. This research work claimed that the accuracy has been improved by 86.5% at convolution layer. The CapsNets achieved the accuracy by using the 64 feature map.
In (Abiwinanda et al. 2019), the authors developed the CNN model by adopted the deep learning approach which was applied to the brain MR images. They developed five classification models and among them second model shown better approximation in the classification of MR images. The proposed architecture contains the RELU layer Max Pool layer and contains the 64 hidden neurons. The proposed model achieved 98.5% accuracy at training phase and 84% accuracy at validation phase. The 2D-DWT (Jayalakshmi and Nagaraja Rao 2020) and Gabor filter are used for brain MR image classification in Ismael and Abdel-Qader (2018). This research work achieved 92% accuracy by applying the back propagation neural networks to the system.
The authors in Pashaei et al. (2018) proposed the classification model based on the neural network mechanism. This model employs the image segmentation to retrieve the tumor region from the image dataset. The authors employed different noise reduction techniques (Li et al. 2021) and transformation mechanisms to improve the prediction accuracy. They considered two data sets to determine the accuracy. The proposed model achieved 90.6% accuracy and classified the data set with different grades. To address the grading mechanism, 3D-CNN model was proposed in Sajjad et al. (2019). This model automatically grades the glioma from MR images. The proposed model has two functionalities: one is identifying the region of interest and grading the glioma. The 3D-CNN model achieved 89.5% accuracy at whole system and 92.9% accuracy in finding the region of interest.
In (Kabir Anaraki et al. 2019), the authors developed the hybrid approach by combining the genetic algorithm with convolution neural network to classify the brain MR images. The proposed system uses the genetic algorithm for choosing the CNN structure. This model achieved 90.9% accuracy in classification of Pituitary, Meningioma and Glioma images. In (Muneer et al. 2019), the authors used real-time dataset from the United States hospitals. They applied customized classification with the help of Wndchrm tool. This tool uses the deep learning approach for convolution neural networks by measuring the neighborhood distance. They achieved the accuracy of 92.8% with the proposed setup. The authors in Banerjee et al. (2019) used the CNN models to classify the MR images. They developed the ConvNet mechanisms to slice and patch the MR images. The authors also considered the existing models like VGGNet (Somasundaram and Gobinath 2019) to classify the MR images. The proposed model achieved 97% accuracy on different datasets.

Dataset
The Brain MR image Dataset contains the 3064 T1weighted images (Naseer et al. 2019). The dataset contains three types of tumours such as Glioma, Meningioma and pituitary. The resolution of the images is taken as 512 9 512. The algorithm trained the images though the Inception-V3 CNN model and compared the accuracy with pretrained-VGG16 model and ResNet-50 model. The dataset is spitted in to three segments for training phase, testing phase and validation phase. The dataset for training is used for model learning, the testing data are used for evaluating the model and the validation data set is used for tuning the parameters. Figure 1 shows the sample images of the dataset.

Proposed model
The proposed model is composed of different layers and pre-training mechanisms. Figure 2 explains about the preprocessing model, training and testing phases and brain tumor prediction. The proposed model uses the deep learning approach and hyper parameters. These parameters are optimized using the Adam Optimizer and loss function. The loss function helps the machines to model the algorithm with input data. Gradually, loss function minimizes the prediction error with the help of some optimization mechanisms. Adam Optimizer is used to optimize the parameters like sparse gradients. Figure 3 shows the Pre-trained Inception-V3 model which is adapted from GoogleNet (Szegedy et al. 2015). This network consists of 11 modules of inception and each inception module is composed of convolution layer, activation layer, Max pooling layer and Normalization layer. These inception modules help to retrieve maximum number of features from the input images. In the pre-trained Inception-V3 deep learning architecture, some of the inception modules at bottom layers are removed and concatenated the features from the inception modules from the top to perform the classification in the Brain MR Image datasets. The last inception module in the architecture is concatenated with the fully connected layers, global average pooling and extracted features of the images from inception modules. To overcome the over fitting problem in the proposed model, the dropout layer is followed after the global average polling layer. The proposed model extracts the features from the fully connected layer and forwards the extracted features to the classifier. The softmax classifier is used in the proposed model to classify the images in to multiple classes.

ReLU activation function
The proposed deep learning model uses the rectified linear unit (ReLU) activation model in convolution layer. The functionality of the activation function is to convert the input weights in to the output to the next layer (Nair and Hinton 2010). ReLU is usually employed in the convolution hidden layers. Equation 1 shows the representation of ReLU in convolution layer.  becomes 1. Hence, if the input value is 0 then the algorithm considers that the neuron is dead and it will not be considered.

Loss function
Loss function is used to compute the error between the true label values and predicted values. Later error is reduced using the optimization mechanism. In this research work, the cross entropy loss function is used to calculate the error (Mannor et al. 2005). As the proposed model is performing the multiclass classification of the MR images so it considered multi class cross entropy. Equation 2 shows the multiclass cross entropy.
where X i represents the input value and the Y i represents the one-hot encoding target value (y i1, y i2 ,…y ic ).

Optimization mechanism
In the proposed deep learning mechanism, the proposed model used adaptive moment estimation (Adam) optimizer In (Robbins and Munro 1951), the SGD method was explained with Eq. 5.
where dK represents the derivative of the weights, dg represents the derivatives of bias for every epoch.
Equation 6 shows the SGD with movement Z, the gradients mean movement is range from 0 and 1.

VGG-16 Model
The pre-trained VGG-16 CNN model is used in our research work. VGG-16 model is modified by removing some of the layers to overcome the over fitting issue. In this model, it considered 16 convolution layers which are proposed in Simonyan and Zisserman (1409). The image input size is taken as 224 9 224 9 3 and considered 16 convolution layers with 3 9 3 filter size. This network considered 5 max pooling layer of size 2 9 2. Figure 4 shows the architecture of the VGG-16 model and it contains 2 fully connected layers along with softmax to classify the images at the top of the model.

ResNet50 Architecture
The ResNet50 is a deep CNN model which is developed by Microsoft corporation in 2015 (Ji et al. 2019). It contains the 50 layers with 26 million parameters in the network. In ResNet, the learning process is carried out from residuals which are subtracted features from input layers. Figure 5 shows the Resnet 50 architecture where it skip link to forward the information across layers. This research work used the pre-trained ResNet50 architecture to compare the performance with the proposed architecture.

Results and discussion
The performance of the proposed model is tested with Dataset-255 which is having 255 pathological brain medical images of size 256 9 256 resoultion. This data set contains different type of pathological images such as Giloma, Meningioma, Pituitary and Normal. The proposed model selects the 25 images of each pathological image. In the considered dataset, 220 are the diseased images and 35 are the normal images.
To evaluate the performance of the proposed model, the algorithm used the following parameters such as accuracy (A), precision (P), recall (R) and F1-score. These parameters measure the predicted and true values from Eqs. (7-10), respectively.
where T p represents true positive, T N represents true negative, F p represents false positive and F N represents the false negative. The proposed model is trained with different hyper-parameters which are given in Table 1. Tensorflow and Keros with backend have been used to train the deep learning model in the proposed architecture. The 80% of the data is allocated for training, and 20% data is allocated for testing. The computation of evaluation metric is done by using the ground truth labels and predicted labels. Table 2 shows the precision, recall and F1-score of the different blocks in Inception-v3 model. The blocks available at the bottom layers in the Inception-v3 extracts the features and forwards them to the classifier to classify the brain tumor images. Table 3 shows the comparison of accuracy of proposed and existing algorithms. The proposed Inception-v3 CNN architecture achieved 99.34% accuracy on the training data and 89% accuracy on the validation data. By using the transfer learning process, the pre-trained VGG-16 model achieved 91% accuracy on the training data and 86% Fig. 4 Pre-Trained VGG-16 Architecture (Simonyan and Zisserman 1409) accuracy on the validation data, ResNet-50 achieved 92% accuracy on the training data and 77% accuracy on the validation data. Figures 6, 7, and 8 show the accuracy of training and validation of Inception-v3, VGG-16 and ResNet-50. Figure 9 shows the region of convergence (ROC) of the proposed and existing models. The ROI of the proposed Inception-v3 model achieved good results for each class of brain tumor (0, 1, 2, 3), class-0 represents the Glioma, class-1 represents the Meningioma, class-2 represents the pituitary and class-3 represents the normal brain.

Conclusion
This paper explained the deep learning mechanism for detection of brain tumors in MR images. In the proposed pre-trained Inception-V3 deep learning architecture, some of the inception modules at bottom layers are removed and concatenated the features from the inception modules from the top to perform the classification in the Brain MR Image datasets. The last inception module in the architecture is concatenated with the fully connected layers, global average pooling and extracted features of the images from inception modules. The proposed model extracts the features from the fully connected layer and forwards the extracted features to the classifier. The softmax classifier is used in the proposed model to classify the images in to multiple classes. Tensorflow and Keros with backend have  The accuracy of the proposed model is recorded as 99.34% which is high compared to the VGG-16 and ResNet-50 models. In future, the research work is concentrating on the combination of DensNet model and Inception-v3 model to improve the accuracy.