When to Use Augmentation - Variability Insufficient for Cortical Thickness Estimation Improvement

Cortical thickness (CTh) is an important biomarker commonly estimated using slow-inference software packages such FreeSurfer. In recent years fast-inference, Deep Learning (DL)-based CTh estimation solutions emerged as an attractive alternative to traditional CTh estimation methods but achieved lower accuracy than traditional methods. Such performance can be attributed to inadequate variability of training samples in anatomy and appearance. In this paper, we investigate the effectiveness of uniform atrophy simulation as the augmentation strategy for DL-based CTh estimation methods. Our results indicate that uniform atrophy simulation is not an effective augmentation strategy for DL-based CTh estimation methods since the introduced variability in anatomy is insufficient. As a byproduct, we observed that increasing the appearance variability of training samples is a more effective augmentation strategy.