Show abstract - PDF - Reviews

Recent development in magnetic resonance (MR) harmonization has facilitated the synthesis of varying MR image contrasts while preserving the underlying anatomical structures. This enables an investigation into the impact of different T1-weighted (T1-w) MR image contrasts on the performance of deep learning-based algorithms, allowing the identification of optimal MR image contrasts for pretrained algorithms. In this study, we employ image harmonization to examine the influence of diverse T1-w MR image contrasts on the state-of-the-art ventricle parcellation algorithm, VParNet. Our results reveal the existence of an optimal operating contrast~(OOC) for VParNet ventricle parcellation, achieved by synthesizing T1-w MR images with a range of contrasts. The OOC for VParNet is not of the same MR image contrast of any of the training data. Experiments conducted on healthy subjects and post-surgical NPH patients demonstrate that adjusting the MR image contrast to the OOC significantly enhances the performance of a pretrained VParNet, thereby improving its clinical applicability.
Hide abstract