Machine learning based tumor detection and classification in brain magnetic resonance images

Meningiomas are the most common primary brain tumors in the central nervous system. Although most of them are benign, they can damage many vital functions due to their size and location. There are multi-class brain tumor classification techniques in the literature, including the meningioma class. In this study, the proposed method aims to increase the success of meningioma tumor detection by using binary classification instead of multiple classification methods. The proposed method increases the classification success by applying a three-step classification method and combining the power of four different machine learning algorithms in the final classification stage. Figshare brain tumor dataset, which includes T1-weighted contrast-enhanced brain Magnetic Resonance (MR) images of meningiomas and other tumor types, was used in the study. First, Gray-Level Co-Occurance Matrix algorithm (GLCM) features were extracted from the gray-level converted training images and these data were trained with Linear SVM. After filtering the training images of meningioma with the Non-Local Means (NLM) filter, they were segmented using k-Means Clustering and the GLCM features and contour features of the regions in the cluster where the tumor was located were extracted. Then, the contour features were trained with eXtreme Gradient Boosting (XGBoost). The GLCM features were trained with XGBoost, Light Gradient Boosting Machine (LGBM), Linear Support Vector Machine (Linear SVM), Extra Trees. Then, the necessary classifications were made on the test images in the order determined in the algorithm and the result was determined as meningioma or not meningioma. When the obtained results were compared with the literature, it was observed that the proposed method performed better.