Transfer Learning Based Classification of Brain Tumors using MRI Images

Abstract

In this paper, a novel representation for Magnetic Resonance Image classification is proposed using transfer learning which exploits the classification of Brain Tumor into No Tumor, Glioma, Meningioma and Pituitary. In this work MRI images were taken. MRI scans are manually analysed by radiologists to detect abnormal conditions in the brain. It takes a long time and it is difficult to manually interpret a large number of photos. However, the complexity associated with the MRI system makes this task non-trivial. Especially, distinguishing between different types of tumors namely Glioma, Meningioma, and Pituitary is not easy and is highly subjective. To address this issue, computer-based detection helps in accurate, fast and exact diagnosis of the disease. In the proposed work,Resnet50 and VGG19 models were used.  InitiallyResnet50 and VGG19 network model all the layers were trained, the dense layer is added with the softmax classifier which classifies the brain Tumor into four types namely no tumor, Glioma, Meningioma and Pituitary, the weights are frozen before layer46 for Resnet50 and Layer15 for VGGnet19. By comparing all the results Resnet50(all the layers were trained) the accuracy is 85.64 and VGG19 (all the layers were trained) the accuracy is 88.94.In Resnet50 (Freezing of layers) the accuracy is 76.10 and the VGG19Resnet50 (Freezing of layers) the accuracy is 85.64.