[1]江立雯,谢海滨,杨光.基于两阶段级联神经网络的乳腺体素内不相干运动参数图重建[J].中国医学物理学杂志,2026,43(9):1201-1208.[doi:DOI:10.3969/j.issn.1005-202X.2026.09.007]
　JIANG Liwen,XIE Haibin,et al.Two-stage cascaded neural network for breast intravoxel incoherent motion parameter map reconstruction[J].Chinese Journal of Medical Physics,2026,43(9):1201-1208.[doi:DOI:10.3969/j.issn.1005-202X.2026.09.007]
点击复制

基于两阶段级联神经网络的乳腺体素内不相干运动参数图重建()

《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]

卷:
43卷
期数:
2026年第9期
页码:
1201-1208
栏目:
医学影像物理
出版日期:
2026-09-30

文章信息/Info

Title:
Two-stage cascaded neural network for breast intravoxel incoherent motion parameter map reconstruction
文章编号:
1005-202X（2026）09-1201-08
作者:
江立雯1; 2; 谢海滨1; 2; 杨光1; 2
1.华东师范大学物理学院上海市磁共振重点实验室， 上海 200062；2.华东师范大学医学磁共振与分子影像技术研究院， 上海200062
Author(s):
JIANG Liwen1;  2;  XIE Haibin1;  2;  YANG Guang1;  2
1. Shanghai Key Laboratory of Magnetic Resonance, School of Physics, East China Normal University, Shanghai 200062, China 2. Institute of Magnetic Resonance and Molecular Imaging in Medicine, East China Normal University, Shanghai 200062, China
关键词:
扩散加权磁共振成像; 体素内不相干运动; 级联U-Net; 去噪; 重建
Keywords:
Keywords: diffusion weighted imaging intravoxel incoherent motion cascaded U-Net denoising reconstruction
分类号:
R318
DOI:
DOI:10.3969/j.issn.1005-202X.2026.09.007
文献标志码:
A
摘要:
提出一种两阶段级联U-Net用于体素内不相干运动（IVIM）参数估计。该方法将去噪与重建解耦，先预训练去噪网络，再联合训练去噪网络和重建网络，并辅以差异化学习率防止预训练去噪特征退化，从而提升低信噪比下的重建质量。在2024年AAPM组织的IVIM-dMRI重建挑战赛的乳腺模拟数据集上进行训练与测试，实验结果表明，全乳腺区域的相对均方根误差（rRMSE）为0.050±0.007，而肿瘤感兴趣区域（ROI）的rRMSE达到0.038±0.006，肿瘤ROI的rRMSE相较于传统最小二乘拟合方法（0.512±0.284）、单一U-Net（0.122±0.035）以及未进行去噪网络预训练的级联结构（0.051±0.010）分别降低92.6%、68.9%和25.5%，显著提升噪声环境下的重建精度。
Abstract:
Abstract: A two-stage cascaded U-Net is developed for the estimation of intravoxel incoherent motion (IVIM) parameters. This method decouples denoising and reconstruction. The denoising network is pre-trained, followed by joint training of the denoising network and the reconstruction network. Differentiated learning rates are adopted to prevent the degradation of pre?rained denoising features, thereby enhancing reconstruction performance under low signal-to-noise ratio conditions. Experiments conducted on the breast simulation dataset from the AAPM IVIM-dMRI Reconstruction Challenge 2024 show that the relative root mean square error is 0.050±0.007 for the whole breast region, and 0.038±0.006 for the tumor region of interest (ROI). For the tumor ROI, the proposed method achieves rRMSE reductions of approximately 92.6%, 68.9%, and 25.5% compared with the traditional least-squares fitting method (0.512±0.284), the single U-Net architecture (0.122±0.035), and the cascaded structure without denoising pre-training (0.051±0.010), respectively, demonstrating improved reconstruction precision under noisy conditions.

相似文献/References:

[1]刘洪源,李勤. IVIM扩散加权成像在前列腺疾病中的研究进展[J].中国医学物理学杂志,2018,35(6):666.[doi:DOI:10.3969/j.issn.1005-202X.2018.06.009]
　LIU Hongyuan,LI Qin. Research advances of intravoxel incoherent motion diffusion-weighted imaging in prostatic diseases [J].Chinese Journal of Medical Physics,2018,35(9):666.[doi:DOI:10.3969/j.issn.1005-202X.2018.06.009]
[2]王毅,赖清泉,黄美铃,等. 磁共振体素内不相干运动在肺良恶性肿瘤的诊断效能及肺癌化疗疗效评估的应用[J].中国医学物理学杂志,2019,36(6):682.[doi:DOI:10.3969/j.issn.1005-202X.2019.06.012]
　WANG Yi,LAI Qingquan,HUANG Meiling,et al. Application of MR-IVIM in the diagnosis of benign and malignant lung tumors and the evaluation of chemotherapeutic efficacy in lung cancer[J].Chinese Journal of Medical Physics,2019,36(9):682.[doi:DOI:10.3969/j.issn.1005-202X.2019.06.012]

备注/Memo

备注/Memo:
【收稿日期】2026-05-11 【基金项目】上海市中央引导地方科技发展资金（YDZX20263100004014002）；“大零号湾项目”（202509-KW-1.1.1-001） 【作者简介】江立雯，硕士研究生，研究方向：医学图像处理、人工智能，E-mail: jlwkmust@163.com 【通信作者】杨光，博士，副研究员，研究方向：医学图像处理、人工智能，E-mail: gyang@phy.ecnu.edu.cn
更新日期/Last Update: 2026-09-30