ExRAN: Deep Ensemble Majority Voting using Transfer Learning for Brain tumor Identification from Magnetic Resonance Imaging
Arnab Chakraborty, D. Vetrithangam
Abstract
Brain Tumor diagnosis at early stage can play a critical role in the prognosis of the patient. Magnetic resonance (MR) imaging is regarded as the present state-of-the-art imaging technique for brain tumor diagnosis. There has been significant work done to identify brain tumor using computer-aided diagnostic model via both hand-crafted an deep learning based approaches. However, due to the variability in contrast and distortion present in the MR images, the identification of BT becomes challenging. The ExRAN model is an ensemble of all the pretrained network i.e. InceptionNet, VGG19, VGG16, Xception, InceptionResNet, ResNet101, ResNet152V2, DenseNet121 are trained individually on the dataset. The single network suffers from poor generalization and convergence whereas multiple single networks stacked using ensemble techniques provide more robust and generalized features. Our proposed ensemble ExRAN model identifies the deep discriminative feature via ensemble majority voting from the trained multiple pre-trained deep learning models. This study focuses on validation of the proposed ExRAN model on open-source MR imaging dataset. Additionally, the ExRAN model surpasses all other existing state-of-the-art and individual deep learning pre-trained models in terms of performance rate and attaining faster convergence rate over subsequent epochs.