Detection of Tumor Slice in Brain Magnetic Resonance Images by Feature Optimized Transfer Learning

This study includes investigating the presence of tumor regions in Magnetic Resonance Imaging (MRI) slices. Since the MRI taken from a patient consists of many slices, it may take time for experts to review these images. The aim of the study is to evaluate the specialist's MRI slices more quickly. The image of each MRI slice taken from the patient was applied to the Alexnet transfer learning algorithm and the properties of the image were obtained. These features are optimized with the Relieff feature selection algorithm to achieve optimum success. The highest accuracy has been achieved with the support vector machine classifier, in which optimized features are used. The study was validated with 3 different combinations by training with two datasets and testing with the other. Thus, a method that can work under different conditions were obtained. The performance metrics of the study were obtained by taking the average of the successes obtained from each data set. MRIs were trained with Alexnet transfer learning model and performance analysis was performed on the obtained classification models. The feature optimization used both increased the success to 97.55% and reduced the processing time from 0.4064 to 0.3045 seconds. The proposed model with a high success rate and a rapid classification is expected to assist the expert in both diagnosis and treatment planning.


INTRODUCTION
Computer Aided Detection (CAD) enables information communication devices to be used by experts in diagnosis and treatment. One of the application areas of CAD is the detection of brain tumors with Magnetic Resonance Imaging (MRI) [1]. MRIs are widely used to identify existing diseases and to plan treatment. They provide structural and contextual information about tumor cells, which can affect the vital functions of the patient.
MRIs are taken in slices. In manual detection, these slices are evaluated by experts. In this examination, experts diagnose the tumors in the image and plan the treatment according to the type and size of the tumor. Accordingly, the duration of diagnosis and the accuracy of classification of the tumorous slices depends on the knowledge and experience of the specialist.
Furthermore, manual examination and feature extraction is problematic in terms of early and reliable diagnosis [2]. At the same time, manual examination is a significant waste of time for experts. Therefore, it is important to detect and classify the tumor in the MRI slice by analyzing them fully automatically to assist experts detect brain tumors. In this schema, the tumor in MRI slice can be detected accurately in short time. Thus, it is aimed to contribute the early diagnosis process by saving time in tumor diagnosis by eliminating unnoticed MRI slices. Feature extraction and classification are essential in automated processing of MRIs [3]. It is the motivation of the study that transfer learning methods give more successful results than other statistics and measurement methods from the extraction of the features of the images.
Many studies have been conducted in the literature in this field in recent years with this motivation. Kaur  MRI. The first step of this approach involved image filtering. This process included image cropping, scaling and histogram equalization methods, respectively. Then, the gray level concurrence matrix method was used to extract MRI features. After the pretreatment step, the images with tumors were classified using the Random Forest method. 120 patient data were used in the evaluation and the classification accuracy of the developed model was found to be 87.62% [5]. Havaei et al. used Brats2013 data set to classify the MRIs with cascading Two-Way CNN. These CNNs provided local and global processing simultaneously with the data set.
The proposed method has achieved 88% successful tumorous MRIs [6]. Dimilliler and İlhan proposed a multi-layer perceptron classification system of brain MRIs as healthy or tumor.
First, brain MR images of 256 256 were resized to 64 64 in order to complete the training in less time and pre-processed with histogram equalization method. From the pre-processed brain MR images, tumorous images and tumor-free images were labeled by artificial neural network (ANN). The classification success was achieved as 90% [7]. Kazdal   the Alexnet transfer learning model were classified by SVM method in 3 different data sets in this study. The data set of the study was obtained using Rembrandt [16], Rider Neuro MR [17] and Brain Tumor Progression [18,19]. The data set of the study was created with 1800 axial healthy MRI slices from 170 patients from these data sets. Performance analysis was performed to obtain the most successful model in determining tumor and non-tumor MRI slices with this data set. In the training of the models, two different data sets were combined and 1200 images were obtained and verified by cross-validation method. The third data set was used for testing.
Thus, a robust method that can work under different conditions has been obtained. It was determined as the system in which Alexnet, Relieff and SVM are used together. In this classification system, which was created by integrating Relieff feature selection algorithm into Alexnet and SVM models, the test success rate was 97.55%; classification time was obtained as 0.3045 seconds.
The contributions of the study can be listed below.
 The high success rate obtained by using 3 different data sets together shows that the proposed method can work successfully on data sets with different characteristics. Thus, a more robust system assistant for experts was obtained.
 The proposed method can help experts reduce their busy schedule and assist them in identifying tumors that might be overlooked at early diagnosis.
 The accuracy of SVM classifier is increased and a more effective system is achieved by ensuring its performance with the Relieff feature optimization.

MATERIAL AND METHODS
The data set containing MR images is first converted into a format and size suitable for network structures. Firstly, MR images in DICOM format were converted to JPEG format. Then, each MR image slice was resized to 256x256. Images in the data set have 3-dimensional depth in RGB format. Then, the new format and size images obtained were used as input data in the  Fig. 1.

Alexnet Transfer Learning Model
Deep learning is a machine learning technique that is inspired by the human brain that has the ability to analyze, observe, make decisions for complex problems and process them in a short time. It has taken its place in academic studies as an application that can extract features with or without supervision, can perform operations such as transformation and classification by using large amounts of data and is frequently used in today's technology.
AlexNet is the first large-scale deep network capable of image classification. This architecture is generally one of those architectures that move away from stacking convolution and pooling layers on top of each other in a consecutive structure. This is because stacking all layers and adding multiple filters costs a calculation and memory and increases the possibility of memorization. Alexnet has used modules connected in parallel to overcome this situation [20].
Network architecture of AlexNet is given in Fig. 2. The Alexnet architecture is previously trained with millions of images and has the ability to classify in 100 object categories. Before the softmax layer, the network operation was interrupted and it was used to extract the properties of the images of brain MRIs.

Relieff Feature Selection Algorithm
Feature selection is a subset of features belonging to a class and obtained by feature extraction models. The relieff feature selection algorithm, with filter type feature selection, gives the best result in estimating the feature significance for models that use double distances between observations to predict the response, directing the distance [21]. properties that are greater than a certain threshold value are selected [22].  parameters. In training sets, inputs and outputs are matched to obtain decision functions that will classify input data [23]. It is widely used in classification problems because of its lack of congruence problems, its ability to work with more than one independent variable, and its success in modeling complex boundaries.

RESULTS AND DISCUSSIONS
The data set of the study was obtained from Rembrandt [16], Rider Neuro [17] and Brain Tumor Progression (BTP) [18,19] MRI data sets, which are used in the literature and whose reliability has been proven by scientific studies, which are in the National Institutes of Health (NIH) Cancer Imaging Archive (TCIA). Each of the images in the sets consists of axial region brain MRIs. Sets and their contents are as shown in Table 1. Three different data sets were created in the study by separating the tumorous MRI slice and without tumorous MRI slice shown in Table 1 with the help of an expert. The data sets, which were formed as a double group, were divided into two different groups, healthy and tumor. Three different data sets were used in the study, as shown in Figure 3. The first group consists of 50% healthy MRI data and 50% tumor MRI data of Remrandt and Rider Neuro MR data set.
The second group consists of 51.6% healthy and 48.4% tumor MRI data of the BTP and Rider Neuro MR data set. The third group of Rembrandt and BTP data set consists of MRI data of 49.2% healthy and 50.8% tumorous. In order to avoid the disadvantage of the unbalanced dataset, the dataset was organized by taking approximately equal numbers of normal and tumor images.

Healthy Tumor
Feature vectors created by obtaining attributes belonging to healthy and tumor classes using Alexnet deep learning architecture. It was optimized with the Relieff feature selection algorithm and applied to the SVM classifier and MRI slices were classified as tumor or healthy.

Evaluation of the Proposed Method
In this study, Alexnet model is used as a transfer learning architecture. Three different data sets,  Table 2.
According to the complexity matrices obtained with the experimental results, performance analysis was made between 3 data set combinations and six different classification models. As a result of the analysis made. The hybrid model obtained by adding the Relieff feature selection algorithm to the Alexnet deep learning model, which was trained using the Rembrandt data set, was the most successful model. In addition, when the test times of the classification algorithms given in Table 2 are compared, the model obtained by optimizing Alexnet with Relieff, which is the study using the Rembrandt data set, was determined as the fastest model. The average accuracy of proposed method is obtained as 97.55%. The time taken for the test was obtained as 0.3045 seconds.
In Figure 4, the variables obtained from the complexity matrix of the most successful model and their values are given.

Comparison with the State-of-the-Art Methods
The proposed model and deep learning studies conducted in the literature and their success rates are given in that similar methods and datasets. The accuracy of proposed method is gathered similar to or better than other methods.  [26].
Along with the experimental results; The hybrid system, in which the Alexnet deep learning model and the Relieff feature selection algorithm are used together, has been identified as the most successful model in terms of performance and speed in general. The average accuracy of the proposed method in the experiments performed was 97.55%. The time taken for the test was obtained as 0.3045 seconds. The ROC curve of the most successful system is given in Fig. 5.

CONCLUSION
In this study, an automated system for classifying brain cancer MRI slices using Magnetic Resonance Imaging (MRI) taken from 3 different datasets is presented. The test of the proposed method is done at the level of classifying SVM and the other dataset by optimizing the features from Alexnet transfer learning, which is trained with images taken from 2 different datasets, with the Relieff method. Accuracy, sensitivity, specificity, sensitivity and f1 score performance metrics are used to evaluate the proposed method. In the experiments, it was determined that feature optimization increases the success and shortens the working time. The proposed methodology can be used as a helpful tool for busy professionals to examine MRI slices with tumor detection only, rather than whole MRI slices. The results of the experiments reveal the superior performance of the proposed model compared to previous methods. This study will help to examine brain tumors accurately on MRI slices.