Deep Transferable Representation Learning For Accurate Multi-Class Brain Tumor Classification From MRI Images

Abstract

Despite the huge amount of medical imaging data available and the fact that magnetic resonance imaging (MRI) shows nearly unlimited potential for tumor imaging, accurate multi-class classification of brain tumors from MRI is still very difficult due to the heterogeneity of the tumor types, differences in appearance of the images and the limited amount of medical imaging data available to learn sufficiently discriminative features. The goal of this study is to devise a deep representation learning approach that is transferable across brain tumor datasets for automated brain tumor classification. It is an experimental subset of Brain Tumor MRI benchmark dataset containing 3265 images which includes 2609 training, 325 validation and 331 testing images from four classes including glioma, meningioma, pituitary tumor and no-tumor. Three different transfer-learning architectures (ResNet50V2, VGG16, DenseNet121, and EfficientNetB0) are explored in order to extract discriminative representations and classify the data. ResNet50V2 achieves the highest test accuracy of 94.26%, compared with VGG16 (93.66%), DenseNet121 (92.45%), and EfficientNetB0 (28.70%). ResNet50V2 also achieves macro precision of 94.46%, macro recall of 94.29% and macro F1 score of 94.37% with pituitary tumor classification attaining a recall of 100% and 98.38% macro F1 score. The study provides a new framework that integrates the different aspects of transferable representations, class-specific discrimination, and model complexity. The results prove that ResNet50V2 gives the best performance when compared to the other architectures presented for automated multi-class brain tumor MRI classification.