Do Pre-processing and Augmentation Help Explainability? A Multi-seed Analysis for Brain Age Estimation

DaeHyun Cho, Christian Wallraven

Abstract: The performance of predicting biological markers from brain scans has rapidly increased over the past years due to the availability of open datasets and efficient deep learning algorithms. There are two concerns with these algorithms, however: they are black-box models, and they can suffer from over-fitting to the training data due to their high capacity. Explainability for visualizing relevant structures aims to address the first issue, whereas data augmentation and pre-processing are used to avoid overfitting and increase generalization performance. In this context, critical open issues are: (i) how robust explainability is across training setups, (ii) how a higher model performance relates to explainability, and (iii) what effects pre-processing and augmentation have on performance and explainability. Here, we use a dataset of 1,452 scans to investigate the effects of augmentation and pre-processing via brain registration on explainability for the task of brain age estimation. Our multi-seed analysis shows that although both augmentation and registration significantly boost loss performance, highlighted brain structures change substantially across training conditions. Our study highlights the need for a careful consideration of training setups in interpreting deep learning outputs in brain analysis.