High accuracy brain tumor classification with EfficientNet and magnetic resonance images

Abstract

The automatic detection of brain tumors is important for efficiently processing large amounts of data. This is a complex task due to the great variety that exists, and the inherent challenges associated with processing medical images. The aim of this work is to assess the performance of recent neural networks for the classification of brain tumors. We study the EfficientNet model, which has provided good results in many classification problems. We use two standard datasets with more than 3000 magnetic resonance images each. The classification includes four different classes with three tumor types (glioma, meningioma, pituitary tumors), and an additional class for brains without tumors. The experiments analyze three models of the EfficientNet architecture, using several techniques, such as transfer learning, early stopping and fine-tuning. The results show that the models attain an accuracy of 98.4% and 97.5% with the two datasets, which is on par with state-of-the-art methods.