Classification of Brain Tumor MR Images Using Transfer Learning and Machine Learning Models

Abstract

Brain tumors are the extra buildup of cells in the regions of brain. Any unwanted mass in the brain tissues is termed as tumor. Glioma, meningioma and pituitary tumors are the three main types of tumors. All the tumors are dangerous and cause severe damage when not treated. Proper diagnostic methods include tumor viewing using MR images among other imaging techniques. But, the manual identification of the tumor from MR images requires more time and error prone. Therefore classification of brain tumors requires the development of an effective method. Deep learning has methods have proved to be efficient in classification and identification tasks by using feature extraction methods.

In this paper deep transfer learning technique has been used for multi class tumor classification on the publicly available dataset. Eleven pretrained networks are used and experimented for a learning rates of 1e−2, 1e−3 and 1e−4 and sgdm, adam, rmsprop as optimizers. Further the same networks are experimented using classifiers support vector machine (SVM), k nearest neighbor (KNN) and decision tree (DT) for varied learning rates and optimizers. Inception-v4 network has observed an highest overall accuracy of 98.1$\%$ with KNN for a learning rate of 1e−4, 25 epoch and rmsprop as optimizer. The highest overall accuracy without the integration of classifiers is observed as 96.5$\%$ for resNet-50 with a learning rate of 1e−4 and rmsprop as optimizer. It is observed that the proposed method has observed higher accuracy with the integration of classifiers. Performance metrices used other than accuracy include precision, F1 score, micro F1 score.