Brain MR Image Classification Based on Deep Features by Using Extreme Learning Machines

Magnetic Resonance Imaging (MRI) is a noninvasive medical testing procedure that can help physicians to examine internal body structures and diagnose a variety of disorders, such as tumors. MRI has some advantages over other imaging methods: mainly that there is no risk of being exposed to radiation. As a result of this many researchers from the community of computer vision and machine learning are interested in classifying or segmenting MR images to help physicians perform more detailed investigations and an automatic system for brain tumor detection and classification was proposed. Firstly, brain MR images are preprocessed by using a 5x5 Gaussian filter. Secondly, deep feature extraction was performed by using Alex Net and VGG16 models of pre-trained Convolutional Neural Network (CNN). The obtained feature vectors are combined. These feature vectors were used for MR images classification by Extreme Learning Machines (ELM) classifier. The performances of the proposed methods have been evaluated on three different data sets. Performance parameters used to assess the results are; accuracy, sensitivity, selectivity and Jaccard’s similarity index for tumor detection. The experimental results showed that the proposed system is superior in detecting and classifying brain tumors when compared with other systems.


Introduction
Brain tumors are formed due to the abnormal growth of the cells which become proliferated in an uncontrollable way [1]. Tumors can damage brain cells by pressurizing the skull which consequently begins to negatively affect human health.
Brain tumors are encountered with different characteristics and structures. Meningioma develop from the meninges, the membrane covering the brain and spinal cord [2]. Most of meningioma are benign and slow-growing tumors. Glioblastoma is the most common primary malignant brain tumor of all brain tumors. It is also one of the most difficult tumors to treat [3]. The pituitary gland is the region located in the brain base where some hormones are secreted. The benign tumors that develop in the pituitary gland are called pituitary adenomas [4]. Magnetic Resonance Imaging (MRI) is used to examine the structure of the brain tissue [5]. The tumor is composed of various biological tissues; only one protocol cannot give all the information about the tissues of brain. Therefore, MR images obtained on T1, T2 and Proton Intensity (PI) bases need to be considered together to evaluate the state of tumor [6].
Another issue is that the numbers of patients affected by brain tumors have been increasing every year according to the World Health Organization (WHO) data [7]. Yet, brain tumors may progress very rapidly and may have even more negative effects on human health. Therefore, the evaluation of images should be done quickly as it is vitally important in cancer diagnosis and treatment planning. Manual examination by experts could be time consuming, therefore, an alternative option is a dedicated automated system that helps physicians in their diagnosis and thereby speed up the treatment process. There are many computerassisted automatic detection and diagnosis systems in literature [8]. Karim et al. [9] have presented a new method by combining the Gauss Mixture Model (GMM) and the Modified Fuzzy C Means (MFCM) algorithm for tumor detection from brain MR images.
Duan et al. [10] have proposed a new method for performing brain MR image segmentation. This method involved low pass filtering, segmentation using a threshold value and morphological process.
Ahmadav et al. [11] have proposed a method in which the features from brain MR images were extracted by using wavelet transform.
Random Forest classifier was used for classification of MR images as tumor or non-tumor.
Kadam et al. [12], have proposed a method in which the features were extracted by Gray Level Co-occurrence Matrix (GLCM). Kernel Support Vector Machines (KSVM) classifier was used for classifying MR images as tumor or non-tumor. Abbasi et al. [13] have introduced a brain tumor detection method which automatically estimates tumors from volumetric MR images. The images were preprocessed using a histogram equalization method. Those structures that could be considered as tumors were segmented from the images. Some of the features extracted from these structures that have the potential of being tumors. By using the Local Binary Pattern (LBP) method along with these features, a variety of classifiers are selected, and the performances of these classifiers were compared. Mittal et al. In this study, brain MR images were classified in accordance with tumor types. The tumors are segmented using a threshold value defined by add hook. The algorithm starts with filtering out the high frequency components on the images that may be considered as noise using a Gaussian filter. Then, pre-trained CNN models were used to extract features from the images. These are the feature vectors extracted from fc6 and fc7 layers of Alex Net and Vgg16 models. Then, the combinations of these vectors were fed to ELM classifiers. The tumors were segmented from the classified images with a dedicated threshold value set by trial and error.
Finally, morphological operations and tumor masking operations were applied to eliminate mis-segmented pixels in the segmented images. The organization of this paper is as follows. In the next section, the methodology will be introduced. In section 3, the data sets used in this study will be explained. The experimental studies and the results will be explained in section 4. The study will be concluded in section 5.

Preliminaries
An automatic computer assisted system was designed for the detection of brain tumors. The proposed system was tested on MR images. Tumor detection from MR images is more efficient due to high contrast and spatial resolution, and also healthier due to low radiation. MR images provide information about the location and size of the brain tumor, but types of the tumors cannot be directly categorized from these images. In this case, experts wait for the result of the biopsy. The aim of the proposed system was to classify brain tumors according to their types using MR images and to determine the brain tumor. The proposed system consists of five main steps; pre-processing, extraction of deep features, concatenation of deep features, classification of feature vectors and detection of brain tumors. Figure 1 shows the operating principle of the proposed method. The methodology was given in the following sections.

Pre-Processing
Depending on the application, a variety of linear, non-linear, fixed, adaptive or pixel-based pre-processing methods are usually employed before classification to enhance the sensitivity [7,17].
In some conditions where the differentiation between normal and abnormal tissue is complex due to high noise level, experts may naturally make mistakes in interpretation. On the other hand, minor differences between normal and abnormal tissues can also be masked by noise. Therefore, it is necessary to remove the possible noises with preprocessing MR images. Also, the enhancement in the visual quality of the image provides great benefits to experts. In this study, MR images were pre-processed with a 5 × 5 Gaussian filter using Matlab environment. The used filter is given in Eq. 1, where σ represents the bandwidth of the filter. (1)

Feature Extraction With Pre-Trained CNN Models
Two pre-trained CNN models, Alex Net and VGG16, were used in this study for feature extraction. Alex Net is known as the first deep CNN structure which consists of twenty-five layers, eight of which contribute to learning by adjusting weights. Five of these layers are convolution layers while remaining three are fully connected. In Alex Net architecture, max-pooling layers come after the convolution layers, and convolution layers use varying kernel sizes [18]. Another deep CNN model is the VGG16 model proposed by Simonyan et al. [19]. The VGG16 model consists of 41 layers, 16 layers of which have adjustable weights. These layers are composed of 13 convolution and 3 fully connected layers. The VGG16 model uses only 3 × 3 dimensional kernels in all layers of convolution.
Similar to Alex Net, max-pooling layers follow the convolution layers [18,19]. Activations (fc6, fc7) in the first and second fully connected layers are used to extract feature vectors. fc6 and fc7 vectors contain in total 4096 features.

Extreme Learning Machine
The Extreme Learning Machine (ELM) was firstly designed by Huang et al. [20]. ELM is a single hidden layered feed forward neural network whose input weights are calculated randomly, and output weights are calculated analytically [20][21][22]. This approach has several advantages over conventional learning algorithms like Back Propagation (BP) [23]. A sample ELM model is shown in

Segmentation
The segmentation process was performed to separate brain texture from other textures in the MR images. Thus, structures such as skulls, backgrounds, scalps and eyes were removed from the MR images. The basic principles of segmentation algorithm rely on region enlargement, deformable templates, clustering with thresholding and pattern recognition techniques. In this study, the structures that might be tumors were segmented by a threshold value determined as described in Eq. (2).

Datasets
Three data sets which are open access related to brain MR images were used in this study. The first data set was RIDER. The Cancer Imaging Archive (TCIA) organized RIDER brain image data that

Experimental Studies and Results
Experiments were conducted to test the performance of the proposed system. A computer with Intel Core i5-4810 CPU and 8 GB memory is used in the experiments and programs are written in MATLAB 2017-b environment.

Classification Results
Brain MR Images were resized to 227 × 227 and 224 × 224 sizes respectively to use for Alex Net and VGG16 models. The features of brain MR images were removed from the layers fc6 and fc7.        Table 3 shows the results for REMBREDANT data set. Table   3 shows that all deep feature combinations produce acceptable accuracy and the combination of all feature sets provides the highest accuracy. The highest accuracy was 98.46% and the feature numbers was 16384. The second highest accuracy was 98.30% for the feature combinations of Alex Net fc6 and VGG16 fc6, and the feature numbers was 8192. The lowest accuracy was calculated as 97.10% from the feature combinations of Alex Net fc6 and Alex Net fc7, and the number of features was 8192. The highest accuracy rates obtained for each data set were compared with other methods in the literature. In Table 4, proposed method was compared with Amin et al.'s method that used RIDER data set. As seen in Table 4, our proposed method is superior then Amin's method with 10fold cross validation [6]. Table 5

Segmentation Results
During the segmentation phase of the image, global thresholding was used, and a threshold value was selected for this. When the pixel value of interested region was less than the threshold value, it was ignored, and the image was transformed into binary image. Thresholding process gives faster results than the methods performing segmentation by using generally more than one feature of the image. Since brightness, especially in the majority of biomedical images, is a distinctive feature in terms of segmentation, it becomes an application area where the thresholding process can be used. After the thresholding process, morphological operations and windowing technique were used. When the performance evaluation based on accuracy, sensitivity and specificity were performed, it was observed that the best threshold was T = 155.
The performance values in determining the threshold were shown in Table 7. The rows show the tested threshold values while the columns show the calculated numerical performance measures in Table 7. As can be seen from These parameters are given in Eq.
(3) to (6).  Figure 3 is reviewed visually, it is seen that tumor regions have been successfully segmented close to real. When the tumor detection results of the data sets were examined, the tumor performance was on the acceptable level for three data sets. It was seen that the best success in tumor detection was obtained from RIDER data set. The results given in Table 8 confirm the visual evaluation. While rows indicate the names of the data sets, columns present the calculated numerical performance criteria in Table 8. As seen from

Conclusion
Firstly, the effects of combined deep features on image classification have been examined in this article. In other words, the best way to combine deep feature sets in the task of classifying brain MR images is investigated. Two pre-trained CNN models were used to extract feature vectors from three different data sets.
Results indicate that proposed brain tumor classification yields better accuracy in comparison to those available in the literature for the selected datasets. In future studies, we are planning to investigate the classification effect of feature combination of early layers. In this study, tumor detection from brain MR images was performed with segmentation by means of thresholding, followed by morphological operations. In the future studies, it is planned to perform tumor detection with new segmentation techniques.