Detail Matters: High-Frequency Content for Realistic Synthetic MRI Generation

Deep Learning (DL)-based segmentation methods have been quite successful in various medical imaging applications. The main bottleneck of these methods is the scarcity of quality-labelled samples needed for their training. The lack of labelled training data is often addressed by augmentation methods, which aim to synthesise realistic samples with corresponding labels. While the synthesis of realistic samples remains a challenging task, little is known about the impact of fine detail in synthetic data on the performance of DL-based segmentation models. In this work, we investigate whether, and to what extent, the high-frequency (HF) detail in synthetic brain MR images (MRIs) impacts the performance of DL-based segmentation methods. To assess the impact of HF detail, we generate two synthetic datasets, with and without HF detail and train corresponding segmentation models to evaluate the impact on their performance. The results obtained demonstrate that the presence of HF detail in synthetic brain MRIs, used during training, significantly improve the Dice score up to 1.73% for Gray Matter (GM), 1.34% for White Matter (WM) and 4.41% for Cerebrospinal Fluid (CSF); and therefore justify the need for synthesising realistic-looking MRIs.