Brain tumor classification using DenseNet and U-net convolutional neural networks

Diagnosing various types of brain tumors and distinguishing between them presents complicated and challenging tasks. To address these issues, automatic methods are highly desirable. This work proposes a method for classifying tumors in T1-weighted contrast-enhanced MRI (CE-MRI) images into three of the most frequent types: meningioma, pituitary tumor, and glioma. Two scenarios are considered: first scenario, images are segmented using the U-Net Convolutional Neural Network to isolate the tumor region of interest (ROI). While, in the second scenario, two pre-trained Dense Convolutional Networks (DenseNet121) are used to extract features from the original images and the cropped images. These features are then concatenated and passed to a softmax layer to classify brain tumors. The proposed method is based on the concatenation of features at the output of pre-trained DenseNet121 models, achieving an accuracy of 98.86% on the test dataset. This outperforms existing state-of-the-art methods in brain tumor detection.