Does Deep Learning Require Image Registration for Early Prediction of Alzheimer’s Disease? A Comparative Study Using ADNI Database

Image registration is the process of using a reference image to map the input images to match the corresponding images based on certain features. It has the ability to assist the physicians in the diagnosis and following up on the patient’s condition. One of the main challenges of the registration is that it takes a huge time to be computationally efficient, accurate, and robust as it can be framed as an optimization problem. In this paper, we introduce a comparative study to investigate the influence of the registration step exclusion from the preprocessing pipeline and study the counter effect of the augmentation on the model performance. We achieved the goal of the study through three experiments using Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset. Finally, a T-statistical test is applied to validate our hypothesis with a p-value of 0.027, in which case the null hypothesis should be rejected. Our proposed approach of using augmentation without any registration outperforms the other experiments for the AD vs CN task with an AUC score of 94.62%. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.