基于磁共振深度学习重建算法缩短扫描时间的可行性分析：水模研究

doi:10.16150/j.1671-2870.2024.02.006

摘要/Abstract

摘要：

目的：旨在评估深度学习重建（deep learning reconstruction, DLR）算法在缩短磁共振成像（magnetic resonance imaging, MRI）扫描时间方面的应用潜力。方法：利用水模，采用控制变量法描绘扫描时间随着激励次数（number of excitation，NEX）、矩阵、l-to-noise ratio，SNR）及主观图像质量变化（包括锐利度及细节清晰度、失真度的四分制评估），并描绘变化曲线、计算拟合曲线。结果：在传统重建和不同降噪水平的DLR重建中，NEX和分辨率与MRI扫描时间和SNR之间存在正相关性。在相同的NEX和分辨率条件下，传统重建、DLR_L、DLR_M和DLR_H的SNR依次升高。以主观评价3或者4分为令人满意的图像，当矩阵固定为512×512时，不同NEX值下，DLR重建的图像在锐利度、失真度和细节显示方面均表现出色，且在NEX为3、5、7和11时，图像细节显示最佳，同时显著缩短了扫描时间。当NEX为2、4、5、6时，图像的失真令人满意，当NEX为3、5、7和11时，可获得满意的细节显示。以上所有组合，可节省扫描时间31~244 s。随着分辨率的增加，图像质量在锐利度、失真度和细节显示方面均有所提升，失真度较低。当NEX固定为6时，DLR_H、DLR_M、DLR_L及传统重建的图像分别在矩阵为320×320、384×384、448×448及640×640时，即扫描时间分别为141 s、141 s、187 s及232 s时，可获得令人满意的锐利度。DLR_H和DLR_M在512×512矩阵下实现了较小的失真度，而DLR_L和传统重建需要更高的成像矩阵和更长的扫描时间以获得类似的图像质量。对于细节显示的清晰度，DLR_H在512×512矩阵下的表现尤为突出，扫描时间少于DLR_M、DLR_L及传统重建。结论：DLR，特别是DLR_H，可在降低NEX和分辨率以缩短MRI扫描时间的同时，不仅能保持令人满意的SNR和图像细节显示，还有可能实现更高的图像清晰度和更低的失真度。

关键词: 深度学习重建, 激励次数, 图像质量

Abstract:

Objective To explore the feasibility of deep learning reconstruction (DLR) in shortening the scanning time through water phantom experiments. Methods The control variable method based on phantom, was adopted to depict the curves of scanning time varying with the number of excitation (NEX) ，matrix，and resolution. The signal-to-noise ratio (SNR) and subjective image quality of different DLR with high, medium and low noise reduction levels (DLR_H, DLR_M, DLR_L) and traditional reconstruction (ConR) were analyzed, including the four-point assessment of sharpness and detail clarity, and distortion degree, and the curves of changes were depicted and the fitting curves were calculated. Results The positive correlation between NEX and resolution with MRI scanning time and SNR was consistent in ConR and DLR reconstructions with different noise reduction levels. Specifically, under the same NEX and resolution conditions, the SNR of ConR, DLR_L, DLR_M, and DLR_H increased sequentially. When the matrix was fixed at 512×512 while images with subjective evaluation score of 3 or 4 were taken as satisfactory ones, the images reconstructed by DLR can obtain satisfactory sharpness, distortion degree and detail display, and when NEX were 3, 4, 5 and 7 and 11, image details were displayed best while scanning time was significantly reduced. Meanwhile, the distortion of images achieved satisfactory results with NEX of 2, 4, 5, and 6. Also, satisfactory detail display was obtained when NEX was 3, 5, 7 and 11. All the above combinations of NEX and resolution saved scanning time from 31 to 244 seconds. Similarly, as the resolution increased, the image scores of the sharpness and detail display gradually increased and. distortion degree decreased. When NEX was fixed at 6, the images reconstructed by DLR_H, DLR_M, DLR_L and ConR obtained satisfactory sharpness ；When the matrix was 320×320, 384×384, 448×448 and 640×640,, the scanning time was 141 seconds, 141 seconds, 187 seconds and 232 seconds, respectively. DLR_H and DLR_M achieved a smaller distortion degree at the 512×512 matrix, while DLR_L and ConR required a higher imaging matrix and longer scanning time to obtain similar image quality. For detail clarity, DLR_H achieved satisfactory detail display when the matrix was 512×512, and the scanning time was less than that of DLR_M, DLR_L and ConR. Conclusions DLR, especially DLR_H, while reducing NEX and resolution to shorten the MRI scanning time, may not only maintain a satisfactory SNR and image detail clarity, but also achieve higher image clarity and lower distortion degree.

Key words: Deep learning reconstruction, Number of excitations, Image quality

中图分类号:

R445.2

吕晓宇, 冯威铭, 周慧赟, 李纪强, 董海鹏, 黄娟. 基于磁共振深度学习重建算法缩短扫描时间的可行性分析：水模研究[J]. 诊断学理论与实践, 2024, 23(02): 131-138.

LÜ Xiaoyu, FENG Weiming, ZHOU Huiyun, LI Jiqiang, DONG Haipeng, HUANG Juan. Feasibility of reducing scan time based on deep learning reconstruction in magnetic resonance imaging: a phantom study[J]. Journal of Diagnostics Concepts & Practice, 2024, 23(02): 131-138.

图/表 8

图1

表1

图2

图3

图4

图5

图6

图7

参考文献 16

[1]	RUSSO V, LOVATO L, LIGABUE G. Cardiac MRI: technical basis[J]. Radiol Med, 2020, 125(11):1040-1055.
[2]	DONOHO D L. Compressed sensing[J]. IEEE Transactions on Information Theory, 2006, 52(4):1289-1306.
[3]	DESHMANE A, GULANI V, GRISWOLD M A, et al. Parallel MR imaging[J]. J Magn Reson Imaging, 2012, 36(1):55-72. doi: 10.1002/jmri.23639 pmid: 22696125
[4]	郭效宾, 柯腾飞, 刘一帆, 等. 磁共振压缩感知技术在肿瘤影像检查中应用[J]. 放射学实践, 2020, 35(12):1635-1638.
	GUO X B, KE T F, LIU Y F, et al. The application of magnetic resonance compressed sensing technology in tumor imaging examination.[J] Radiol Pract, 2020, 35(12): 1635-8.
[5]	LIBERMAN G, SOLOMON E, LUSTIG M, et al. Multiple-coil k-space interpolation enhances resolution in single-shot spatiotemporal MRI[J]. Magn Reson Med, 2018, 79(2):796-805. doi: 10.1002/mrm.26731 pmid: 28556180
[6]	严福华. 深度学习MRI重建算法的临床应用和发展前景[J]. 磁共振成像, 2023, 14(5):8-10.
	YAN F H. The clinical application and development prospect of deep learning MRI reconstruction algorithm[J]. Chin J Magn Reson Imag, 2023, 14(5):8-10.
[7]	刘高平, 曲太平, 许强, 等. 基于深度学习重建常规头部2D T1WI超分辨率图像质量[J]. 中国医学影像技术, 2022, 38(3): 326-331.
	LIU G P, QU T P, XU Q, et al. Imaging quality of super-resolution reconstruction of conventional head 2D T1WI based on deep learning[J]. Chin J Med Imag Technol, 2022, 38(3): 326-331.
[8]	宣锴, 王乾. 面向磁共振图像重建的k空间降采样优化[J]. 模式识别与人工智能, 2021, 34(4):367-374. doi: 10.16451/j.cnki.issn1003-6059.202104009
	XUAN K, WANG Q. Optimizing K-space subsampling pattern toward MRI reconstruction[J]. Pattern Recognit Artif Intell, 2021, 34(4):367-374.
[9]	LEBEL R M. Performance characterization of a novel deep learning-based MR image reconstruction pipeline[EB/OL]. [2023-05-13]. https://www.baidu.com/link?url=XDD_QK0-r2gYithJyjWW_vbNytceiRHd9g_IMsUUeNIYFyIephWTH9uhbAyTeOTG&wd=&eqid=d3683c33001cda870000000566456e76.
[10]	KIM M, KIM H S, KIM H J, et al. Thin-slice pituitary MRI with deep learning-based reconstruction: diagnostic performance in a postoperative setting[J]. Radiology, 2021, 298(1):114-122. doi: 10.1148/radiol.2020200723 pmid: 33141001
[11]	YASAKA K, TANISHIMA T, OHTAKE Y, et al. Deep learning reconstruction for 1.5 T cervical spine MRI: effect on interobserver agreement in the evaluation of degenerative changes[J]. Eur Radiol, 2022, 32(9):6118-6125.
[12]	SPRAWLS P. Magnetic resonance imaging: principles, methods, and techniques[M]. Medical Physics Publishing Madison, 2000.
[13]	周楠, 花立春, 刘杰, 等. 深度学习重建法在MRI重建中的应用进展[J]. 中国医疗设备, 2023, 38(12):165-169.
	ZHOU N, HUA L C, LIU J, et al. Application of deep learning reconstruction in MRI reconstruction[J] China Med Devices, 2023, 38(12):165-169.
[14]	XIE Y, TAO H, LI X, et al. Prospective comparison of standard and deep learning-reconstructed turbo spin-echo MRI of the shoulder[J]. Radiology, 2024, 310(1):e231405.
[15]	LEE K L, KESSLER D A, DEZONIE S, et al. Assessment of deep learning-based reconstruction on T2-weighted and diffusion-weighted prostate MRI image quality[J]. Eur J Radiol, 2023, 166:111017.
[16]	ZERUNIAN M, PUCCIARELLI F, CARUSO D, et al. Artificial intelligence based image quality enhancement in liver MRI: a quantitative and qualitative evaluation[J]. Radiol Med, 2022, 127(10):1098-1105. doi: 10.1007/s11547-022-01539-9 pmid: 36070066

Item	Matrix	In-plane voxel size (mm)	In-plane voxel volume(mm³)	NEX	Scan time(s)
Part 1	512×512	0.47	0.66	1	35
	512×512	0.47	0.66	2	65
	512×512	0.47	0.66	3	95
	512×512	0.47	0.66	4	126
	512×512	0.47	0.66	5	156
	512×512	0.47	0.66	6	187
	512×512	0.47	0.66	7	217
	512×512	0.47	0.66	8	247
	512×512	0.47	0.66	9	278
	512×512	0.47	0.66	10	308
	512×512	0.47	0.66	11	339
	512×512	0.47	0.66	12	369
	512×512	0.47	0.66	13	399
	512×512	0.47	0.66	14	430
	512×512	0.47	0.66	15	460
	512×512	0.47	0.66	16	491
Part 2	128×128	1.88	10.55	6	50
	192×192	1.25	4.69	6	95
	256×256	0.94	2.64	6	95
	320×320	0.75	1.69	6	141
	384×384	0.63	1.17	6	141
	448×448	0.54	0.86	6	141
	512×512	0.47	0.66	6	187
	640×640	0.38	0.42	6	232
	768×768	0.31	0.29	6	232
	896×896	0.27	0.22	6	278
	1 024×1 024	0.23	0.16	6	647