Analysis of Deep Learning based Ensemble Models for Brain Tumor Detection

Brain tumor is one of the severe diseases that can lead to human death. The probability of success for human treatment increases if the tumor is detected at an early stage. This paper works on the transfer learning models for detection of brain tumor. Transfer learning is one of the most useful deep learning techniques which offers reutilization of the pre-trained models on new tasks. This paper analyzed the performance of pre-trained models, i.e., VGG16, Xception, VGG19, and ResNet50, on three different MRI based datasets of varying sizes. On the basis of the performance of these models, this work continues the analysis using different ensembles of VGG16, ResNet50, and Xception. Overall, this work analyzes twelve ensembles, which include three weighted ensembles for VGG16+ResNet50, VGG16+Xception, Xception+ResNet50, and four weighted ensembles of VGG16+Resnet50+Xception. This work shows that the VGG16+ResNet50 ensemble with 25% weightage to ResNet50 and 75% weightage of VGG16, gives 94.04%,98.6%, and 99.4% accuracy on dataset1, dataset2, and dataset3, respectively, to outperform all other individual and ensemble models.