-
摘要:
虽然现有基于深度学习的压缩感知磁共振成像(CS-MRI)方法已经取得了较好的效果,但这些方法的可解释性仍然面临挑战,并且从理论分析到网络设计的过渡并不够自然。为解决上述问题,提出深度双域几何蒸馏特征自适应网络(DDGD-FANet)。该深度展开网络将磁共振成像重建优化问题迭代展开成3个子部分:数据一致性模块、双域几何蒸馏模块和自适应网络模块,不仅可以补偿重建图像丢失的上下文信息,恢复更多的纹理细节,还可以去除全局伪影,进一步提高重建效果。在公开数据集使用3种不同的采样模式进行实验,结果表明:DDGD-FANet在3种采样模式下均取得了更高的峰值信噪比和结构相似性指数,在笛卡儿10%压缩感知(CS)比率下,峰值信噪比较迭代收缩阈值算法(ISTA-Net+)、快速ISTA(FISTA)-Net和DGDN模型分别提高了5.01 dB、4.81 dB和3.34 dB。
Abstract:Although the existing compressed sensing-magnetic resonance imaging (CS-MRI) methods based on deep learning have achieved good results, the interpretability of these methods still faces challenges, and the transition from theoretical analysis to network design is not natural enough. In order to solve the above problems, this paper proposed a deep dual-domain geometry distillation feature adaptive network (DDGD-FANet). The deep unfolding network iteratively expanded the MRI reconstruction optimization problem into three sub-modules: data consistency module, dual-domain geometry distillation module, and adaptive network module. It could compensate for the lost context information of the reconstructed image, restore more texture details, remove global artifacts, and further improve the reconstruction effect. Three different sampling modes were used in the public dataset. The results show that DDGD-FANet achieves a higher peak signal-to-noise ratio and structural similarity index in all three sampling modes. At the Cartesian 10% compressed sensing(CS )ratio, the peak signal-to-noise ratio is increased by 5.01 dB, 4.81 dB, and 3.34 dB, respectively, higher than that of iterative shrinkage-thresholding algorithm (ISTA)-Net +, fast ISTA (FISTA)-Net, and DGDN models.
-
表 1 不同CS比率笛卡儿采样模式下大脑数据集重建效果对比
Table 1. Comparison of reconstruction effects in different CS ratios of Cartesian sampling pattern in brain dataset
掩码类型 方法 PSNR/dB SSIM CS比率为10% CS比率为20% CS比率为30% CS比率为10% CS比率为20% CS比率为30% Zero-filling[32] 23.84 26.40 31.14 0.574 3 0.661 1 0.786 5 PBDW[10] 26.82 32.41 35.05 0.735 5 0.866 0 0.911 6 PANO[11] 28.98 34.62 36.75 0.789 7 0.893 3 0.922 2 DC-CNN[24] 30.04 35.75 38.48 0.823 5 0.920 5 0.949 3 笛卡儿 ISTA-Net[22] 30.28 36.58 39.17 0.825 6 0.933 4 0.956 8 ISTA-Net+[22] 30.86 37.03 39.63 0.845 5 0.935 5 0.959 7 FISTA-Net[23] 31.06 37.29 39.86 0.844 4 0.939 6 0.960 7 DGDN[25] 32.53 38.31 40.61 0.886 3 0.949 9 0.965 6 DDGD-FANet 35.87 39.55 41.41 0.932 0 0.959 3 0.970 1 表 2 不同CS比率伪径向和二维随机采样模式下大脑数据集建效果对比
Table 2. Comparison of reconstruction effects in different CS ratios of pseudo-radial and 2D random sampling pattern in brain dataset
掩码类型
方法PSNR/dB CS比率为10% CS比率为20% CS比率为30% 伪径向 Zero-filling[32] 26.64 30.32 32.93 PBDW[10] 32.48 36.53 38.95 PANO[11] 33.61 37.01 39.29 DC-CNN[24] 34.32 38.43 40.75 ISTA-Net[22] 34.70 38.74 41.01 ISTA-Net+[22] 34.83 38.75 41.00 FISTA-Net[23] 35.10 39.05 41.19 DGDN[25] 35.91 39.33 41.34 DDGD-FANet 37.08 39.67 41.54 二维随机 Zero-filling[32] 27.71 29.36 31.13 PBDW[10] 35.68 38.49 40.83 PANO[11] 36.60 39.57 42.03 DC-CNN[24] 37.65 40.81 43.40 ISTA-Net[22] 38.22 41.40 43.97 ISTA-Net+[22] 38.41 41.66 44.20 FISTA-Net[23] 38.36 41.75 44.29 DGDN[25] 38.85 41.91 44.38 DDGD-FANet 39.25 42.13 44.53 表 3 本文模型消融实验结果
Table 3. Results of the proposed model in ablation experiment
表 4 数据一致性层消融实验结果
Table 4. Ablation experiment results of data consistency layer
模型 PSNR/dB SSIM DDGD-FANet_NODC 34.73 0.9215 DDGD-FANet 35.87 0.9320 -
[1] BIAN W Y. Optimization-based deep learning methods for magnetic resonance imaging reconstruction and synthesis[D]. Gainesville: University of Florida, 2022. [2] LUSTIG M, DONOHO D L, SANTOS J M, et al. Compressed sensing MRI[J]. IEEE Signal Processing Magazine, 2008, 25(2): 72-82. doi: 10.1109/MSP.2007.914728 [3] DONOHO D L. Compressed sensing[J]. IEEE Transactions on Information Theory, 2006, 52(4): 1289-1306. doi: 10.1109/TIT.2006.871582 [4] QU X B, GUO D, CHEN Z, et al. Compressed sensing MRI based on nonsubsampled contourlet transform[C]//Proceedings of the IEEE International Symposium on IT in Medicine and Education. Piscataway: IEEE Press, 2008: 693-696. [5] STALIDIS G, MAGLAVERAS N, DIMITRIADIS A, et al. Parametric 4D modeling of myocardial surface motion applied to MRI data[C]//Proceedings of the Computers in Cardiology. Piscataway: IEEE Press, 1997: 319-322. [6] SONAWANE S S, AGNIHOTRI S P. Compressed MR imaging using wavelet transform[C]//Proceedings of the International Conference on Micro-Electronics and Telecommunication Engineering. Piscataway: IEEE Press, 2016: 259-264. [7] LINGALA S G, JACOB M. Blind compressive sensing dynamic MRI[J]. IEEE Transactions on Medical Imaging, 2013, 32(6): 1132-1145. doi: 10.1109/TMI.2013.2255133 [8] LIANG D, WANG H F, YING L. SENSE reconstruction with nonlocal TV regularization[C]//Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society. Piscataway: IEEE Press, 2009: 1032-1035. [9] OUYANG Y, CHEN Y, WU Y. A spatial regularization framework of orientation diffusion functions using total variation and wavelet[C]//Proceedings of the IEEE International Symposium on Biomedical Imaging: From Nano to Macro. Piscataway: IEEE Press, 2011: 272-275. [10] QU X B, GUO D, NING B D, et al. Undersampled MRI reconstruction with patch-based directional wavelets[J]. Magnetic Resonance Imaging, 2012, 30(7): 964-977. doi: 10.1016/j.mri.2012.02.019 [11] QU X B, HOU Y K, LAM F, et al. Magnetic resonance image reconstruction from undersampled measurements using a patch-based nonlocal operator[J]. Medical Image Analysis, 2014, 18(6): 843-856. doi: 10.1016/j.media.2013.09.007 [12] BECK A, TEBOULLE M. A fast iterative shrinkage-thresholding algorithm for linear inverse problems[J]. SIAM Journal on Imaging Sciences, 2009, 2(1): 183-202. doi: 10.1137/080716542 [13] BOYD S, PARIKH N, CHU E, et al. Distributed optimization and statistical learning via the alternating direction method of multipliers[J]. Foundations and Trends in Machine Learning, 2010, 3(1): 1-122. [14] LECUN Y, BENGIO Y, HINTON G. Deep learning[J]. Nature, 2015, 521(7553): 436-444. doi: 10.1038/nature14539 [15] WANG T H, LEI Y, FU Y B, et al. Medical imaging synthesis using deep learning and its clinical applications: a review[EB/OL]. (2020-04-21)[2023-02-01]. https://arxiv.org/abs/2004.10322v1. [16] WANG S S, SU Z H, YING L, et al. Accelerating magnetic resonance imaging via deep learning[C]//Proceedings of the IEEE 13th International Symposium on Biomedical Imaging. Piscataway: IEEE Press, 2016: 514-517. [17] YANG G, YU S M, DONG H, et al. DAGAN: deep de-aliasing generative adversarial networks for fast compressed sensing MRI reconstruction[J]. IEEE Transactions on Medical Imaging, 2018, 37(6): 1310-1321. doi: 10.1109/TMI.2017.2785879 [18] YANG Y, SUN J, LI H B, et al. ADMM-CSNet: a deep learning approach for image compressive sensing[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(3): 521-538. doi: 10.1109/TPAMI.2018.2883941 [19] WANG S S, XIAO T H, LIU Q G, et al. Deep learning for fast MR imaging: a review for learning reconstruction from incomplete k-space data[J]. Biomedical Signal Processing and Control, 2021, 68: 102579. doi: 10.1016/j.bspc.2021.102579 [20] AGHABIGLOU A, EKSIOGLU E M. Deep unfolding architecture for MRI reconstruction enhanced by adaptive noise maps[J]. Biomedical Signal Processing and Control, 2022, 78: 104016. doi: 10.1016/j.bspc.2022.104016 [21] YANG Y, SUN J, LI H B, et al. Deep ADMM-net for compressive sensing MRI[C]//Proceedings of the 30th International Conference on Neural Information Processing Systems. New York: ACM, 2016: 10-18. [22] ZHANG J, GHANEM B. ISTA-net: interpretable optimization-inspired deep network for image compressive sensing[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2018: 1828-1837. [23] XIANG J X, DONG Y G, YANG Y J. FISTA-net: learning a fast iterative shrinkage thresholding network for inverse problems in imaging[J]. IEEE Transactions on Medical Imaging, 2021, 40(5): 1329-1339. doi: 10.1109/TMI.2021.3054167 [24] SCHLEMPER J, CABALLERO J, HAJNAL J V, et al. A deep cascade of convolutional neural networks for dynamic MR image reconstruction[J]. IEEE Transactions on Medical Imaging, 2018, 37(2): 491-503. doi: 10.1109/TMI.2017.2760978 [25] FAN X H, YANG Y, ZHANG J P. Deep geometric distillation network for compressive sensing MRI[C]//Proceedings of the IEEE EMBS International Conference on Biomedical and Health Informatics. Piscataway: IEEE Press, 2021: 1-4. [26] LU C, JIANG B R, MU Y D. Fast fourier convolution[C]//Proceedings of the Neural Information Processing Systems. Vancouver: Curran Associates Inc, 2020: 4479-4488. [27] SINGH N M, IGLESIAS J E, ADALSTEINSSON E, et al. Joint frequency and image space learning for MRI reconstruction and analysis[J]. The Journal of Machine Learning for Biomedical Imaging, 2022, 2022: 018. [28] WOO S, PARK J, LEE J Y, et al. CBAM: convolutional block attention module[C]//Proceedings of the Computer Vision–ECCV 2018. Berlin: Springer, 2018: 3-19. [29] CHENG J, WANG H F, ZHU Y J, et al. Model-based deep MR imaging: the roadmap of generalizing compressed sensing model using deep learning[EB/OL]. (2029-09-05)[2023-02-01]. https://arxiv.org/abs/1906.08143. [30] CUI Z X, ZHU Q Y, CHENG J, et al. Deep unfolding as iterative regularization for imaging inverse problems[EB/OL]. (2022-11-24)[2023-02-01]. https://arxiv.org/abs/2211.13452v1. [31] JIANG K, WANG Z Y, YI P, et al. Deep distillation recursive network for remote sensing imagery super-resolution[J]. Remote Sensing, 2018, 10(11): 1700. doi: 10.3390/rs10111700 [32] BERNSTEIN M A, FAIN S B, RIEDERER S J. Effect of windowing and zero-filled reconstruction of MRI data on spatial resolution and acquisition strategy[J]. Journal of Magnetic Resonance Imaging, 2001, 14(3): 270-280. doi: 10.1002/jmri.1183 -