Inception Architecture for Brain Image Classification

Abstract

A non-invasive diagnostic support system for brain cancer diagnosis is presented in this study. Recently, very deeper convolution neural networks are designed for computerized tasks such as image classification, natural language processing. One of the standard architecture designs is the Visual Geometric Group (VGG) models. It uses a large number of small convolution filters (3x3) connected serially. Before applying max pooling, convolution filters are stacked up to four layers to extract features’ abstraction. The main drawback of going deeper is over fitting, and also updating gradient weights is very hard. These limitations are overcome using the inception module, which is wider rather than deeper. Also, it has parallel convolution layers with 3x3, 5x5, and 1x1 filters that reduce the computational complexity due to stacking, and the outputs are concatenated. This study’s experimental results show the usefulness of inception architecture for aiding brain image classification on Repository of Molecular Brain Neoplasia DaTa (REMBRANDT) Magnetic Resonance Imaging (MRI) images with an average accuracy of 95.1%, sensitivity of 96.2%, and specificity of 94%.