School of Information Technology and Electrical Engineering, University of Queensland, Brisbane, Australia.
School of Information Technology and Electrical Engineering, University of Queensland, Brisbane, Australia.
Z Med Phys. 2023 Nov;33(4):578-590. doi: 10.1016/j.zemedi.2022.08.001. Epub 2022 Sep 2.
Background field removal (BFR) is a critical step required for successful quantitative susceptibility mapping (QSM). However, eliminating the background field in brains containing significant susceptibility sources, such as intracranial hemorrhages, is challenging due to the relatively large scale of the field induced by these pathological susceptibility sources.
This study proposes a new deep learning-based method, BFRnet, to remove the background field in healthy and hemorrhagic subjects. The network is built with the dual-frequency octave convolutions on the U-net architecture, trained with synthetic field maps containing significant susceptibility sources. The BFRnet method is compared with three conventional BFR methods and one previous deep learning method using simulated and in vivo brains from 4 healthy and 2 hemorrhagic subjects. Robustness against acquisition field-of-view (FOV) orientation and brain masking are also investigated.
For both simulation and in vivo experiments, BFRnet led to the best visually appealing results in the local field and QSM results with the minimum contrast loss and the most accurate hemorrhage susceptibility measurements among all five methods. In addition, BFRnet produced the most consistent local field and susceptibility maps between different sizes of brain masks, while conventional methods depend drastically on precise brain extraction and further brain edge erosions. It is also observed that BFRnet performed the best among all BFR methods for acquisition FOVs oblique to the main magnetic field.
The proposed BFRnet improved the accuracy of local field reconstruction in the hemorrhagic subjects compared with conventional BFR algorithms. The BFRnet method was effective for acquisitions of tilted orientations and retained whole brains without edge erosion as often required by traditional BFR methods.
背景场消除（BFR）是成功进行定量磁化率映射（QSM）所必需的关键步骤。然而，由于这些病理磁化率源引起的场的相对较大规模，在包含明显磁化率源（如颅内出血）的大脑中消除背景场是具有挑战性的。
本研究提出了一种新的基于深度学习的方法，BFRnet，用于去除健康和出血受试者的背景场。该网络基于 U-net 架构上的双频八度卷积构建，使用包含明显磁化率源的合成场图进行训练。BFRnet 方法与三种传统 BFR 方法和一种以前的深度学习方法进行比较，使用来自 4 名健康和 2 名出血受试者的模拟和体内大脑进行实验。还研究了对采集视野（FOV）方向和大脑掩模的稳健性。
对于模拟和体内实验，BFRnet 在所有五种方法中导致局部场和 QSM 结果中最好的视觉效果，具有最小的对比度损失和最准确的出血磁化率测量。此外，BFRnet 在不同大小的大脑掩模之间产生了最一致的局部场和磁化率图，而传统方法则严重依赖于精确的大脑提取和进一步的大脑边缘侵蚀。还观察到，BFRnet 在所有 BFR 方法中，在与主磁场倾斜的采集 FOV 中表现最佳。
与传统的 BFR 算法相比，所提出的 BFRnet 提高了出血受试者局部场重建的准确性。BFRnet 方法对于倾斜采集方向是有效的，并且保留了整个大脑，而无需像传统的 BFR 方法那样经常需要进行边缘侵蚀。
