The Effect of Noisy and Blurry Data on Deep Learning: Application in Brain Image Classification

Convolutional Neural Networks (CNN) is one of the best Deep Learning algorithms commonly used for computer vision tasks, including medical image analysis. CNN can learn the representational features from images starting from the lower to complex features. However, noisy data can affect the generalization of the networks, which we can often find in medical images, such as Magnetic Resonance Imaging (MRI). In this paper, we intend to find the correlation between noisy data and the performance of CNN models. We build automatic CNN-based classifiers for normal brain MR images based on axial view by setting up three different data scenarios to train the classifiers: 1) original data, 2) blurred data, and 3) noisy data. We also evaluate the relationship between the prediction accuracy and kernel size of the convolutional layers. Based on our investigation, deeper layers and smaller kernels in the CNN models give better generalization.