Brain Tumour Biomarkers by Deep Learning Architectures

Abstract

                Brain tumour may be detected by the use of different medical imaging modalities such as Computed Tomography (CT) and Magnetic Resonance Imaging (MRI). MRI has been shown to be effective in screening brain tumour than CT. Deep features have been proposed for brain image classification on the basis of two different architectures; Visual Geometric Group (VGG) and Inception Architectures (IA).  The need to characterize the brain images as normal or abnormal leads to different deep learning algorithms for the extraction of deep features. The MRI brain image dataset REpository of Molecular BRAin Neoplasia DaTa (REMBRANDT) is studied in this work for the classification. It contains 200 brain images with 100 related to normal and 100 to abnormal. For the analysis, same set of training and testing samples obtained via random split of 50:50 are used by the VGG-16, VGG-19, IA-V1 (GoogleNet) and IA-V3.The classification performance in percentage accuracy, sensitivity and specificity with the above architectures are recorded. Results show that IA-V3 provides best average performance of 95.1% accuracy.