Quantitative susceptibility reconstruction guided by a multi-scale T1-weighted image attention mechanism

doi:10.12122/j.issn.1673-4254.2026.09.23

Abstract

Abstract:

Objective To improve the accuracy and structural consistency of quantitative susceptibility mapping (QSM) reconstruction through a deep learning network guided by multi-scale T1-weighted image attention. Methods A multi-scale T1-weighted attention-driven deep QSM network (T1w-ADQSM) was proposed to enhance the precision and structural fidelity of single-orientation QSM reconstruction. The method incorporates structural features from T1-weighted images and employs an attention module to guide the network to focus on the anatomical boundaries and key regions. Experimental comparisons were conducted among T1w-ADQSM,truncated k-space division (TKD) method, morphology enabled dipole inversion (MEDI) and QSMnet. Quantitative evaluations were performed using the metrics including high-frequency error norm (HFEN), structural similarity index measure (SSIM),normalized root mean square error (NRMSE), and peak signal-to-noise ratio (PSNR). Results In healthy volunteers, T1w-ADQSM achieved the highest PSNR (43.12±1.19) and the lowest NRMSE (51.98±3.65) compared with TKD, MEDI, and QSMnet. T1w-ADQSM outperformed QSMnet across all the 4 quantitative metrics with statistically significant differences (P<0.05). Further validation based on the COSMOS reference reconstructed from multi-orientation data showed that T1w-ADQSM exhibited better structural consistency than QSMnet, and the reconstruction results were closer to the COSMOS reference. Conclusion The proposed multi-scale T1-weighted image attention-driven deep learning reconstruction method improves the accuracy and structural consistency of QSM.

Key words: quantitative susceptibility mapping, image reconstruction, deep learning

Piaoran LI, Quan TAO, Fusheng OYANG, Baoliang GUO, Qiugen HU, Yanqiu FENG. Quantitative susceptibility reconstruction guided by a multi-scale T1-weighted image attention mechanism[J]. Journal of Southern Medical University, 2026, 46(9): 2258-2265.

Figures/Tables 7

Fig.1 Overview of the proposed T1w-ADQSM method.

Fig.2 Generation of the attention gate (A) and multi-scale T1-weighted images (B).

Tab.1 Means and standard deviation of the quantitative performance metrics for T1w-ADQSM with different T1w data types

Types	HFEN (%)	NRMSE (%)	PSNR	SSIM
T1w_norm	59.43±3.56	58.34±4.28	41.56±0.83	0.932±0.033
T1w_edge	58.21±1.67	59.31±3.43	39.82±1.06	0.928±0.024
T1w_fused	56.69±2.96	57.76±3.05	42.06±0.94	0.936±0.011

Fig.3 Comparison of single-orientation QSM results from one volunteer using non-deep learning dipole deconvolution methods (TKD and MEDI) and deep learning-based methods (QSMnet and T1w-ADQSM). The yellow arrows in the figure indicate prominent differences between QSMnet and T1w-ADQSM.

Fig.4 Axial view of the QSM map (first row) and the corresponding difference maps (second row) of the subject in Fig.3. The last two rows show zoomed-in views. The red arrows indicate regions with obvious differences between QSMnet and T1w-ADQSM.

Tab.2 Means and standard deviation of the quantitative performance metrics

Methods	HFEN (%)	NRMSE (%)	PSNR	SSIM
MEDI	62.39±4.69	79.36±7.03	37.16±0.88	0.952±0.016
TKD	59.23±4.18	69.49±5.30	39.25±0.97	0.914±0.023
QSMnet	57.47±3.28	58.89±3.14	41.39±1.05	0.931±0.008
T1w-ADQSM	56.69±2.96	57.76±3.05	42.06±0.94	0.936±0.011

Fig.5 Comparison of single-orientation QSM reconstructions (QSMnet and T1w-ADQSM) with the corresponding COSMOS results. In terms of reconstructed structural fidelity, T1w-ADQSM outperforms the QSMnet method (red arrows).

References 37

[1]	Schweser F, Deistung A, Reichenbach JR. Foundations of MRI phase imaging and processing for Quantitative Susceptibility Mapping (QSM)[J]. Z Für Med Phys, 2016, 26(1): 6-34. doi：10.1016/j.zemedi.2015.10.002
[2]	Wang Y, Liu T. Quantitative susceptibility mapping (QSM): Decoding MRI data for a tissue magnetic biomarker[J]. Magn Reson Med, 2015, 73(1): 82-101. doi：10.1002/mrm.25358
[3]	Buelo CJ, Velikina J, Mao L, et al. Multicenter, multivendor validation of liver quantitative susceptibility mapping in patients with iron overload at 1.5 T and 3 T[J]. Magn Reson Med, 2025, 93(1): 330-40. doi：10.1002/mrm.30251
[4]	李明月, 高洁, 袁毓, 等. 铁代谢异常对高血压性脑出血的老年患者微创血肿清除术后神经功能的影响[J]. 中华老年心脑血管病杂志, 2025, 27(9): 1169-72. doi：10.3969/j.issn.1009-0126.2025.09.008
[5]	Schumacher K, Prince MR, Blumenfeld JD, et al. Quantitative susceptibility mapping for detection of kidney stones, hemorrhage differentiation, and cyst classification in ADPKD[J]. Abdom Radiol, 2024, 49(7): 2285-95. doi：10.1007/s00261-024-04243-6
[6]	Sun HF, Klahr AC, Kate M, et al. Quantitative susceptibility mapping for following intracranial hemorrhage[J]. Radiology, 2018, 288(3): 830-9. doi：10.1148/radiol.2018171918
[7]	Duyn J. MR susceptibility imaging[J]. J Magn Reson, 2013, 229: 198-207. doi：10.1016/j.jmr.2012.11.013
[8]	Kim HG, Park S, Rhee HY, et al. Quantitative susceptibility mapping to evaluate the early stage of Alzheimer's disease[J]. NeuroImage Clin, 2017, 16: 429-38. doi：10.1016/j.nicl.2017.08.019
[9]	Voon CC, Wiltgen T, Wiestler B, et al. Quantitative susceptibility mapping in multiple sclerosis: a systematic review and meta-analysis[J]. NeuroImage Clin, 2024, 42: 103598. doi：10.1016/j.nicl.2024.103598
[10]	Xie Y, Zhang Y, Wu SL, et al. Atrophy-independent and dependent iron and myelin changes in deep gray matter of multiple sclerosis: a longitudinal study using χ‑separation imaging[J]. Acad Radiol, 2025, 32(2): 988-99. doi：10.1016/j.acra.2024.07.031
[11]	朱立科, 詹逸珺, 裴建. 定量磁化率成像在阿尔茨海默病脑铁沉积中的应用进展[J]. 放射学实践, 2026, 41(3): 317-22.
[12]	Yang YQ, Yuan YT, Ren BX, et al. Deep learning-based diffusion MRI tractography: Integrating spatial and anatomical information[J]. Neuroimage, 2025, 317: 121314. doi：10.1016/j.neuroimage.2025.121314
[13]	QSM Consensus Organization Committee, Bilgic B, Costagli M, et al. Recommended implementation of quantitative susceptibility mapping for clinical research in the brain: a consensus of the ISMRM electro-magnetic tissue properties study group[J]. Magn Reson Med, 2024, 91(5): 1834-62. doi：10.1002/mrm.30006
[14]	Li M, Chen C, Xiong Z, et al. Quantitative susceptibility mapping via deep neural networks with iterative reverse concatenations and recurrent modules[J]. Med Phys, 2025, 52(6): 4341-54. doi：10.1002/mp.17747
[15]	Liu T, Spincemaille P, de Rochefort L, et al. Calculation of susceptibility through multiple orientation sampling (COSMOS): a method for conditioning the inverse problem from measured magnetic field map to susceptibility source image in MRI[J]. Magn Reson Med, 2009, 61(1): 196-204. doi：10.1002/mrm.21828
[16]	Shmueli K, de Zwart JA, van Gelderen P, et al. Magnetic susceptibility mapping of brain tissue in vivo using MRI phase data[J]. Magn Reson Med, 2009, 62(6): 1510-22. doi：10.1002/mrm.22135
[17]	Li W, Wang N, Yu F, et al. A method for estimating and removing streaking artifacts in quantitative susceptibility mapping[J]. Neuroimage, 2015, 108: 111-22. doi：10.1016/j.neuroimage.2014.12.043
[18]	Liu T, Liu J, de Rochefort L, et al. Morphology enabled dipole inversion (MEDI) from a single-angle acquisition: comparison with COSMOS in human brain imaging[J]. Magn Reson Med, 2011, 66(3): 777-83. doi：10.1002/mrm.22816
[19]	de Rochefort L, Liu T, Kressler B, et al. Quantitative susceptibility map reconstruction from MR phase data using Bayesian regularization: Validation and application to brain imaging[J]. Magn Reson Med, 2010, 63(1): 194-206. doi：10.1002/mrm.22187
[20]	Langkammer C, Bredies K, Poser BA, et al. Fast quantitative susceptibility mapping using 3D EPI and total generalized variation[J]. NeuroImage, 2015, 111: 622-30. doi：10.1016/j.neuroimage.2015.02.041
[21]	Liu Z, Spincemaille P, Yao YH, et al. MEDI+0: Morphology enabled dipole inversion with automatic uniform cerebrospinal fluid zero reference for quantitative susceptibility mapping[J]. Magn Reson Med, 2018, 79(5): 2795-803. doi：10.1002/mrm.26946
[22]	Yoon J, Gong EH, Chatnuntawech I, et al. Quantitative susceptibility mapping using deep neural network: QSMnet[J]. Neuroimage, 2018, 179: 199-206. doi：10.1016/j.neuroimage.2018.06.030
[23]	Bollmann S, Rasmussen KGB, Kristensen M, et al. DeepQSM- using deep learning to solve the dipole inversion for quantitative susceptibility mapping[J]. NeuroImage, 2019, 195: 373-83. doi：10.1016/j.neuroimage.2019.03.060
[24]	Jung W, Yoon J, Ji S, et al. Exploring linearity of deep neural network trained QSM: QSMnet+[J]. NeuroImage, 2020, 211: 116619. doi：10.1016/j.neuroimage.2020.116619
[25]	Chen YC, Jakary A, Avadiappan S, et al. QSMGAN: Improved Quantitative Susceptibility Mapping using 3D Generative Adversarial Networks with increased receptive field[J]. NeuroImage, 2020, 207: 116389. doi：10.1016/j.neuroimage.2019.116389
[26]	Wei HJ, Cao S, Zhang YY, et al. Learning-based single-step quantitative susceptibility mapping reconstruction without brain extraction[J]. Neuroimage, 2019, 202: 116064. doi：10.1016/j.neuroimage.2019.116064
[27]	Jung W, Bollmann S, Lee J. Overview of quantitative susceptibility mapping using deep learning: Current status, challenges and opportunities[J]. NMR Biomed, 2022, 35(4): e4292. doi：10.1002/nbm.4292
[28]	Zhang JW, Nguyen TD, Solomon E, et al. mcLARO: Multi-contrast learned acquisition and reconstruction optimization for simultaneous quantitative multi-parametric mapping[J]. Magn Reson Med, 2024, 91(1): 344-56. doi：10.1002/mrm.29854
[29]	Gao Y, Xiong Z, Shan SS, et al. Plug-and-Play latent feature editing for orientation-adaptive quantitative susceptibility mapping neural networks[J]. Med Image Anal, 2024, 94: 103160. doi：10.1016/j.media.2024.103160
[30]	Wang YD, He NY, Zhang CY, et al. An automatic interpretable deep learning pipeline for accurate Parkinson's disease diagnosis using quantitative susceptibility mapping and T1-weighted images[J]. Hum Brain Mapp, 2023, 44(12): 4426-38. doi：10.1002/hbm.26399
[31]	Fiscone C, Sighinolfi G, Manners DN, et al. Multiparametric MRI dataset for susceptibility-based radiomic feature extraction and analysis[J]. Sci Data, 2024, 11: 575. doi：10.1038/s41597-024-03418-6
[32]	Fischl B. FreeSurfer[J]. NeuroImage, 2012, 62(2): 774-81. doi：10.1016/j.neuroimage.2012.01.021
[33]	Chartsias A, Joyce T, Giuffrida MV, et al. Multimodal MR synthesis via modality-invariant latent representation[J]. IEEE Trans Med Imaging, 2018, 37(3): 803-14. doi：10.1109/tmi.2017.2764326
[34]	Brignoni-Pérez E, Dubner SE, Ben-Shachar M, et al. White matter properties underlying reading abilities differ in 8-year-old children born full term and preterm: a multi-modal approach[J]. NeuroImage, 2022, 256: 119240. doi：10.1016/j.neuroimage.2022.119240
[35]	Kingma DP, Ba J. Adam: a method for stochastic optimization[EB/OL]. 2014: arXiv: 1412.6980.
[36]	Glorot X, Bordes A, Bengio Y. Deep sparse rectifier neural networks[C]//International Conference on Artificial Intelligence and Statistics., 2011.
[37]	Langkammer C, Schweser F, Shmueli K, et al. Quantitative susceptibility mapping: Report from the 2016 reconstruction challenge[J]. Magn Reson Med, 2018, 79(3): 1661-73. doi：10.1002/mrm.26830