Volumetrie Data Augmentation as an Effective Tool in MRI Classification Using 3D Convolutional Neural Network

Deep learning has become a default choice for many researchers in the areas of prediction and automatic diagnostics of different diseases. One of the prerequisites for achieving good results in the domain of an automatic diagnostics is the size and quality of the training set. One of the most serious challenges in the domain of a medical image analysis is the lack of manually annotated data. In this work, we use a relatively simple convolutional neural network to classify MRI images as either cognitively normal or those suffering from Alzheimer's disease. Based on a quantitative analysis of multiple techniques, we propose to use a specific set of data augmentation techniques to improve the generalization of the deep neural network model. The proposed approach was compared with three different approaches using the same neural network - one approach with no data augmentation, one with regularization not specific to visual data (dropout and weight decay) and one with extensive data augmentation specific for images. We conclude that by using the proposed set of augmentation techniques, including novel perturbed normalization, we postponed overfitting, which resulted in a better generalized model.