Brain Tumour Detection and Classification using Multi-level Ensemble Transfer Learning in MRI Dataset
Michael Christopher Xenya, Zenghui Wang
Abstract
Many computer vision related works in brain tumour identification focus exclusively on either a classification model or tumor segmentation model. However, the treatment of brain tumors considers, among other factors, tumor type, shape, and size. This article proposes a framework using a segmentation model for the detection of the exact location and shape of brain tumors. The framework also uses a multi-level ensemble learning model to classify brain tumors in an MRI data set. For the geometric property of a particular tumor, a statistical measurement technique using region properties is adopted. The classification model consists of three base learners, two primary ensemble learners, and a final ensemble model. The three base learners are transferred learners fine-tuned using pre-trained VGG16, Inception-V3, and ResNet50 as base learners. The classification model is trained on three main types of MRI brain tumors as well as a normal brain, which is not so in many existing models. The use of a multi-level ensemble with the technique of re-tuning the weakest base model improves the accuracy of the final ensemble model. This work also demonstrates that a wrong segmentation of a tumor in normal MRI brain image can be disproved when tested on a deep learning model.