A Deep Transfer Learning framework for Multi Class Brain Tumor Classification using MRI

Recent researches proclaim that transfer learning on deep networks have performed deftly on medical diagnosis. The main intention of this work is to implement transfer learning on ResNet 50 for the classification of MR brain images to identify the type of tumors such as glioma, meningioma and pituitary. A pre-trained deep network -ResNet 50 extracts robust features and learns the structure of MR images in its convolutional layers. Then the fully connected (FC) layer of ResNet 50 is replaced with three new set of linear modules, two new set of Leaky Relu modules, two new set of dropout modules, and finally a softmax classification module to distinguish three tumor types. Thus the number of layers in ResNet 50 on transfer learning is increased from 174 to 181.This method is applied to a publically available Figshare MRI dataset, which consists of 3064 T1-weighted contrast-enhanced MR images from 233 patients with three disparate brain tumor types, which includes glioma, meningioma, and the pituitary tumors containing 1426, 708, and 930 images respectively. MR images in all the three planes such as axial, sagittal as well as coronal are incorporated in this dataset. The performance is evaluated by utilizing the five-fold cross-validation and the developed deep transfer learning framework obtained a maximum accuracy as 98.67% as compared to the state-of-art methods with hyperparameters such as dropout of 0.6, learning rate of 0.003 and optimizer as stochastic gradient descent.