Determination of tumor sections in brain magnetic resonance images with feature optimized transfer learning

Computer-aided diagnosis (CBT) supports experts in detecting abnormalities in the clinical process. One of the areas of CBT is the detection of brain tumors using magnetic resonance images (MRI). MRI helps the specialist to diagnose and guide treatment. MRI is obtained by scanning the desired area of the body in detail using X-rays and creating sliced images. MR device during shooting; It enables images of bones, soft tissues, organs and vessels in the body to be viewed in cross-sections by scanning from different angles. Since MRI is taken in slices by hand, it is time-consuming for experts to examine these images. This can be compensated by a CBT system that will automatically detect tumor slices. In this study, we were motivated to use transfer learning algorithms in order to automatically obtain the properties of the image without the need for feature extraction. Alexnet and Resnet50 deep learning models, which are among the transfer learning methods, were preferred because they are simple to use and adapted to the rapidly developed method. However, these models cause a very high amount of features in feature extraction. In order to increase the success of the effective classification of these features in the problem space and to make the algorithm run faster; Alexnet and Resnet50 transfer learning algorithms have been innovatively optimized with Relieff and Presence Component Analysis (NCA) feature selection algorithms. The data set created with the optimized feature vector was applied to the SVM classification algorithm and tumor MRI slices were classified. In this thesis study, three different three different data sets were used in the experimental stage to show the success and effectiveness of the proposed method. These are Rembrandt, Rıder Neuro MR and Brain Tumor Progression datasets. These data sets are preferred because they are widely used in the literature and used to show the performance of the studies. In the datasets, there are 1830 axial healthy and tumor brain MRI images in total. In the experiments, the training, verification, test times and the presence or absence of tumor regions in the MRI slices were compared as a result of the SVM classification process of Alexnet and Resnet50 transfer learning algorithms optimized using Relieff and NCA feature selection algorithms. The optimized algorithm in the experiments performed both worked faster (0.28 seconds) and worked effectively to achieve a more successful classification success (98.4%). The experimental results obtained, the effective operation of the model proposed in this thesis, and its high accuracy rate are at a level that can contribute to the expert's detection of MRI slices containing tumors. In this way, the loss of time of the specialist in rapid diagnosis and treatment plan will be prevented and data that may be overlooked will be prevented. Keywords : Brain Magnetic Resonance Images, Feature Engineering (Relieff, NCA), SVM Classifier, Transfer Learning (Alexnet, Resnet50).