Donnay Corinne, Dieckhaus Henry, Tsagkas Charidimos, Gaitán María Inés, Beck Erin S, Mullins Andrew, Reich Daniel S, Nair Govind
Translational Neuroradiology Section, National Institute of Neurological Disorders and Stroke (NINDS), National Institutes of Health, Bethesda, MD, United States.
qMRI Core, NINDS, National Institutes of Health, Bethesda, MD, United States.
Front Neuroimaging. 2023 Dec 1;2:1252261. doi: 10.3389/fnimg.2023.1252261. eCollection 2023.
Automatic whole brain and lesion segmentation at 7T presents challenges, primarily from bias fields, susceptibility artifacts including distortions, and registration errors. Here, 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), bypassing registration errors.
Brain scans Segmentation from 3T and 7T scanners were analyzed with software packages such as FreeSurfer, Classification using Derivative-based Features (C-DEF), nnU-net, and a novel 3T-to-7T transfer learning method, Pseudo-Label Assisted nnU-Net (PLAn). 3T and 7T MRIs acquired within 9 months from 25 study participants with MS (Cohort 1) were used for training and optimizing. Eight MS patients (Cohort 2) scanned only at 7T, but with expert annotated lesion segmentation, was used to further validate the algorithm on a completely unseen dataset. Segmentation results were rated visually by experts in a blinded fashion and quantitatively using Dice Similarity Coefficient (DSC).
Of the methods explored here, nnU-Net and PLAn produced the best tissue segmentation at 7T for all tissue classes. In both quantitative and qualitative analysis, PLAn significantly outperformed nnU-Net (and other methods) in lesion detection in both cohorts. PLAn's lesion DSC improved by 16% compared to nnU-Net.
Limited availability of labeled data makes transfer learning an attractive option, and pre-training a nnUNet model using readily obtained 3T pseudo-labels was shown to boost lesion detection capabilities at 7T.
7T 下的全脑自动分割和病变分割面临挑战，主要源于偏置场、包括畸变在内的磁化率伪影以及配准误差。在此，我们试图使用深度学习算法（D/L），对临床诊断为多发性硬化症（MS）的参与者在单次磁化准备快速采集梯度回波（MP2RAGE）采集中生成的多个成像对比进行颅骨剥离和全脑分割，从而绕过配准误差。
使用诸如 FreeSurfer、基于导数特征的分类（C-DEF）、nnU-net 以及一种新型的 3T 到 7T 迁移学习方法——伪标签辅助 nnU-Net（PLAn）等软件包，对来自 3T 和 7T 扫描仪的脑部扫描分割进行分析。从 25 名患有 MS 的研究参与者（队列 1）在 9 个月内获取的 3T 和 7T MRI 用于训练和优化。8 名仅在 7T 进行扫描但有专家标注病变分割的 MS 患者（队列 2），用于在完全未见过的数据集上进一步验证该算法。分割结果由专家以盲法进行视觉评分，并使用骰子相似系数（DSC）进行定量评分。
在此探索的方法中，nnU-Net 和 PLAn 在 7T 下对所有组织类别产生了最佳的组织分割效果。在定量和定性分析中，PLAn 在两个队列的病变检测方面均显著优于 nnU-Net（以及其他方法）。与 nnU-Net 相比，PLAn 的病变 DSC 提高了 16%。
标记数据的可用性有限使得迁移学习成为一个有吸引力的选择，并且使用容易获得的 3T 伪标签对 nnUNet 模型进行预训练，已被证明可以提高 7T 下的病变检测能力。
