Maruyama Tomoko, Hayashi Norio, Sato Yusuke, Ogura Toshihiro, Uehara Masumi, Watanabe Haruyuki, Kitoh Yoshihiro
Division of Radiology, Shinshu University Hospital, Matsumoto, Nagano, 390-8621, Japan.
Department of Radiological Technology, Graduate School of Radiological Technology, Gunma Prefectural College of Health Sciences, Maebashi, Gunma, 371-0052, Japan.
BMC Med Imaging. 2025 May 1;25(1):144. doi: 10.1186/s12880-025-01663-8.
Magnetic resonance imaging (MRI) is an essential tool for medical diagnosis. However, artifacts may degrade images obtained through MRI, especially owing to patient movement. Existing methods that mitigate the artifact problem are subject to limitations including extended scan times. Deep learning architectures, such as U-Net, may be able to address these limitations. Optimizing deep learning networks with batch normalization (BN) and dropout layers enhances their convergence and accuracy. However, the influence of this strategy on U-Net has not been explored for artifact removal.
This study developed a U-Net-based regression network for the removal of motion artifacts and investigated the impact of combining BN and dropout layers as a strategy for this purpose. A Transformer-based network from a previous study was also adopted for comparison. In total, 1200 images (with and without motion artifacts) were used to train and test three variations of U-Net.
The evaluation results demonstrated a significant improvement in network accuracy when BN and dropout layers were implemented. The peak signal-to-noise ratio of the reconstructed images was approximately doubled and the structural similarity index was improved by approximately 10% compared with those of the artifact images.
Although this study was limited to phantom images, the same strategy may be applied to more complex tasks, such as those directed at improving the quality of MR and CT images. We conclude that the accuracy of motion artifact removal can be improved by integrating BN and dropout layers into a U-Net-based network, with due consideration of the correct location and dropout rate.
磁共振成像（MRI）是医学诊断的重要工具。然而，伪影可能会降低通过MRI获得的图像质量，尤其是由于患者移动导致的伪影。现有的减轻伪影问题的方法存在局限性，包括扫描时间延长。深度学习架构，如U-Net，可能能够解决这些局限性。使用批量归一化（BN）和随机失活层优化深度学习网络可提高其收敛性和准确性。然而，尚未探索这种策略对用于去除伪影的U-Net的影响。
本研究开发了一种基于U-Net的回归网络用于去除运动伪影，并研究了将BN和随机失活层结合作为实现此目的的策略的影响。还采用了先前研究中的基于Transformer的网络进行比较。总共使用1200张图像（有和没有运动伪影）来训练和测试U-Net的三种变体。
评估结果表明，当实施BN和随机失活层时，网络准确性有显著提高。与伪影图像相比，重建图像的峰值信噪比大约提高了一倍，结构相似性指数提高了约10%。
尽管本研究仅限于体模图像，但相同的策略可能适用于更复杂的任务，例如旨在提高MR和CT图像质量的任务。我们得出结论，通过将BN和随机失活层集成到基于U-Net的网络中，并适当考虑正确的位置和随机失活率，可以提高运动伪影去除的准确性。
