Computer aided diagnosis of brain tumours using deep learning methods

Abstract

Brain tumour categorisation is a crucial part of computer-aided diagnosis of brain newlinetumours. The therapeutic planning of brain tumours depends on the type of tumours. newlineThis research proposes automated systems for category-based classification of brain newlinetumours among the most incident types, glioma, meningioma, and pituitary tumours, newlineusing brain magnetic resonance imaging (MRI) images. The emergence of deep newlinelearning and machine learning algorithms have improved the performance of auto- newlinemated image classification tasks. However, the research problem of medical image newlineclassification using deep learning poses several challenges, such as limited training newlineimages, imbalance in training data for the class-specific samples, more extensive newlinecomputational requirements. Therefore, this research study proposes strategies to newlineachieve improved classification performance, considering the practical limitations in newlinethe domain. newlineThe fully automated diagnostic systems proposed in this research work are built newlineon deep learning methods and are evaluated using the publicly available dataset newlinefrom Figshare. Primarily, a convolutional neural network (CNN) is designed for newlinethe brain tumour classification task. The CNN is used as a feature extractor for newlinebrain MRI images and is used with a support vector machine (SVM) classifier for newlineimproved performance. Again, the performance enhances by applying a pre-trained newlineGoogLeNet using the technique of deep transfer learning. The performance of newlinethe tumour diagnosis is assessed using the metrics: overall classification accuracy, newlineprecision, recall, specificity, F-scores and balanced accuracy. In diagnosing the newlinetumour types, the transfer learned GoogLeNet with an SVM classifier achieves newlinestate-of-the-art performance in accuracy and F-scores. However, the system has newlineconsiderable limitations. newline