BTFCNN: Design of a brain tumor classification model using fused convolutional neural networks

Abstract

Classification of brain tumour from magnetic resonance imaging (MRI) is a multidomain task that involves design of image denoising, segmentation, feature extraction, reduction of features, classification into different tumour types, and post-processing for application-specific use cases. Out of these tasks, feature extraction, feature reduction & classification are effectively performed via convolutional Neural Networks (CNN), but these models are highly generic, due to which their accuracy & precision performance is limited. To improve this performance, various pre-processing methods are combined with CNNs, which perform application-specific segmentation of images before their use for classification tasks. Based on this observation, a segmentation model using Saliency maps is discussed in this text, which assists in application-independent segmentation of images based on their entropy. The segmented images are given to a novel fusion-based convolutional Neural Network (CNN) model, which combines VGGNet 16, AlexNet, Inception Net and Xception Net models. The VGGNet model is used for improving efficiency for Meningioma, because it requires largely varying features, AlexNet was observed to effectively classify Pituitary Adenoma due to its low variance-based classification capabilities, while Inception Net was observed to have better performance for Craniopharyngioma, because of its visible segmentation changes, and Xception Net was able to classify Schwannoma with better accuracy because of its high-capacity of feature extraction & selection under multiple image categories. It was observed that the proposed ensemble model had 97.4% accuracy for Meningioma, 97.9% accuracy for Pituitary Adenoma, 98.2% accuracy for Craniopharyngioma, and 97.6% accuracy for classification of Schwannoma & other disease types. Due to such a high performance, the proposed model is capable of being applicable for real-time use, and can be extended for other classification applications. The model was tested under different datasets, which were taken from Kaggle, Br35H Brain Tumour Dataset, & IEEE Data Port dataset, and consistent accuracy results were achieved. Due to such a high performance, the model is highly scalable for a wide variety of brain tumour classification applications.