Abstract
AbstractIntroductionAutomatic whole brain and lesion segmentation at 7T presents challenges, primarily from bias fields and susceptibility artifacts. Recent advances in segmentation methods, namely using atlas-free and multi-contrast (for example, using T1-weighted, T2-weighted, fluid attenuated inversion recovery or FLAIR images) can enhance segmentation performance, however perfect registration at high fields remain a challenge primarily from distortion effects. We sought to use deep-learning algorithms (D/L) to do both skull stripping and whole brain segmentation on multiple imaging contrasts generated in a single Magnetization Prepared 2 Rapid Acquisition Gradient Echoes (MP2RAGE) acquisition on participants clinically diagnosed with multiple sclerosis (MS). The segmentation results were compared to that from 3T images acquired on the same participants, and with commonly available software packages. Finally, we explored ways to boost the performance of the D/L by using pseudo-labels generated from trainings on the 3T data (transfer learning).Methods3T and 7T MRI acquired within 9 months of each other, from 25 study participants clinically diagnosed with multiple sclerosis (mean age 51, SD 16 years, 18 women), were retrospectively analyzed with commonly used software packages (such as FreeSurfer), Classification using Derivative-based Features (C-DEF), nnU-net (“no-new-Net” version of U-Net algorithm), and a novel 3T-to-7T transfer learning method, Pseudo-Label Assisted nnU-Net (PLAn). These segmentation results were then rated visually by trained experts and quantitatively in comparison with 3T label masks.ResultsOf the previously published methods considered, nnU-Net produced the best skull stripping at 7T in both the qualitative and quantitative ratings followed by C-DEF 7T and FreeSurfer 7T. A similar trend was observed for tissue segmentation, as nnU-Net was again the best method at 7T for all tissue classes. Dice Similarity Coefficient (DSC) from lesions segmented with nnU-Net were 1.5 times higher than from FreeSurfer at 7T. Relative to analysis with C-DEF segmentation on 3T scans, nnU-Net 7T had lower lesion volumes, with a correlation slope of just 0.68. PLAn 7T produced equivalent results to nnU-Net 7T in terms of skull stripping and most tissue classes, but it boosted lesion sensitivity by 15% relative to 3T, increasing the correlation slope to 0.90. This resulted in significantly better lesion segmentations as measured by expert rating (4% increase) and Dice coefficient (6% increase).ConclusionDeep learning methods can produce fast and reliable whole brain segmentations, including skull stripping and lesion detection, using data from a single 7T MRI sequence. While nnU-Net segmentations at 7T are superior to the other methods considered, the limited availability of labeled 7T data makes transfer learning an attractive option. In this case, pre-training a nnU-Net model using readily obtained 3T pseudo-labels was shown to boost lesion detection capabilities at 7T. This approach, which we call PLAn, is robust and readily adaptable due to its use of a single commonly gathered MRI sequence.
Publisher
Cold Spring Harbor Laboratory