A Transformer reconstruction method of Magnetic Resonance Image (MRI) based on multi-head self-attention and convolution feature fusion is studied. The U-shaped network structure is adopted to improve the reconstruction performance by learning the multi-scale features of the image. The Swin Transformer structure of deep separation convolution and multi-head self-attention fusion is adopted to improve the feature learning ability of the network. Simulation experiments based on multiple sampling modes are carried out under CC359 Brain dataset, and the results show that the proposed method is effective in improving the quality of MRI reconstruction and reducing the system complexity.
在网络的架构上,相比较卷积结构的局部感受野特性,通过自注意力机制捕获全局依赖信息的Transformer也在视觉领域迅速流行.FENG等[21]使用任务型Transformer架构,共享MR重构任务和MR超分任务之间的关键特征表示.在基于多模态的任务中,FENG等[22]使用融合交叉注意力机制的Transformer架构迁移备用模态的特征到目标模态上,进行多模MR重构.但是由于自注意力机制通过计算每个像素点与其他像素点的相关性,计算复杂度与图像大小的二次方成正比,具有很高的计算代价.Swin Transformer[16]通过基于平移窗口的多头注意力(Shifted Windows based Multi-Head Self-Attention,W-MSA/SW-MSA)结构,将注意力计算限制在了一个个非重叠窗口内,将计算复杂度从输入尺寸的二次方降低为线性,有效地降低了计算代价.YAN等[23]基于Swin Transformer提出了用于MR图像超分辨率的Transformer网络.SwinMR[14]通过将SwinIR[24]引入到MR图像重构任务,在重构图像的质量和泛化性上都优于基于卷积的网络,但是SwinMR[14]的计算复杂度远高于CNN.SDAUT[25]将Swin Transformer[16]与U-Net[15]结合,同时结合可变形的注意力提升重构质量,在一定程度上降低网络了的复杂性和计算代价.但是这些网络主要利用了自注意力结构进行特征学习,缺乏对局部特征的感知,阻碍了MR图像精细结构的恢复.
本文使用峰值信噪比(Peak Signal to Noise Ratio,PSNR)和结构相似性(Structural Similarity Index,SSIM)作为评价指标.基于图像像素点之差的PSNR通过利用最大信号功率与信号噪声之比来衡量图像的精细程度.而SSIM从图像的亮度、对比度和结构方面来测量图像的相似度.在网络的复杂度和运行速度方面,本文在给定输入数据尺寸为1 × 1 × 256 × 256以及不包含鉴别器参数的条件下,统计重构网络的参数量和计算量(Floating Point Operations,Flops),并且在所有测试切片上,计算所有网络的平均测试时间.
DONOHOD L. Compressed sensing[J]. IEEE Transactions on Information Theory, 2006, 52(4): 1289-1306.
[2]
PRUESSMANNK P, WEIGERM, BÖRNERTP, et al. Advances in sensitivity encoding with arbitrary k-space trajectories[J]. Magnetic Resonance in Medicine, 2001, 46(4): 638-651.
[3]
GRISWOLDM A, JAKOBP M, HEIDEMANNR M, et al. Generalized autocalibrating partially parallel acquisitions (GRAPPA)[J]. Magnetic Resonance in Medicine, 2002, 47(6): 1202-1210.
[4]
LUSTIGM, PAULYJ M. SPIRiT: Iterative self‐consistent parallel imaging reconstruction from arbitrary k‐space[J]. Magnetic Resonance in Medicine, 2010, 64(2): 457-471.
LUSTIGM, DONOHOD L, SANTOSJ M, et al. Compressed sensing MRI[J]. IEEE Signal Processing Magazine, 2008, 25(2): 72-82.
[7]
LIANGD, LIUB, WANGJ, et al. Accelerating SENSE using compressed sensing[J]. Magnetic Resonance in Medicine, 2009, 62(6): 1574-1584.
[8]
LUSTIGM, DONOHOD, PAULYJ M. Sparse MRI: The application of compressed sensing for rapid MR imaging[J]. Magnetic Resonance in Medicine, 2007, 58(6): 1182-1195.
[9]
YANGY, SUNJ, LIH, et al. ADMM-CSNet: A deep learning approach for image compressive sensing[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 42(3): 521-538.
YANGG, YUS, DONGH, et al. DAGAN: Deep de-aliasing generative adversarial networks for fast compressed sensing MRI reconstruction[J]. IEEE transactions on medical imaging, 2017, 37(6): 1310-1321.
[13]
WANGS, SUZ, YINGL, et al. Accelerating magnetic resonance imaging via deep learning[C]//2016 IEEE 13th International Symposium on Biomedical Imaging (ISBI). Prague: IEEE, 2016: 514-517.
[14]
HUANGJ, FANGY, WUY, et al. Swin transformer for fast MRI[J]. Neurocomputing, 2022, 493: 281-304.
[15]
RONNEBERGERO, FISCHERP, BROXT. U-net: Convolutional networks for biomedical image segmentation[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Munich: Springer, 2015: 234-241.
[16]
LIUZ, LINY, CAOY, et al. Swin transformer: Hierarchical vision transformer using shifted windows[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal: IEEE, 2021: 9992-10022.
QUANT M, NGUYEN-DUCT, JEONGW K. Compressed sensing MRI reconstruction using a generative adversarial network with a cyclic loss[J]. IEEE Transactions on Medical Imaging, 2018, 37(6): 1488-1497.
[21]
FENGC M, YANY, FUH, et al. Task transformer network for joint MRI reconstruction and super-resolution[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Strasbourg: Springer, 2021: 307-317.
[22]
FENGC M, YANY, CHENG, et al. Accelerated multi-modal MR imaging with transformers[J]. arXiv Preprint arXiv: 2106.14248, 2021, 3.
[23]
YANC, SHIG, WUZ. SMIR: A transformer-based model for MRI super-resolution reconstruction[C]// 2021 IEEE International Conference on Medical Imaging Physics and Engineering (ICMIPE). Hefei: IEEE, 2021: 1-6.
[24]
LIANGJ, CAOJ, SUNG, et al. SwinIR: Image restoration using swin transformer[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal: IEEE, 2021: 1833-1844.
[25]
HUANGJ, XINGX, GAOZ, et al. Swin Deformable Attention U-Net Transformer (SDAUT) for explainable fast MRI[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Singapore: Springer, 2022: 538-548.
[26]
WANGZ, CUN X, BAOJ, et al. Uformer: A general u-shaped transformer for image restoration[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans: IEEE, 2022: 17683-17693.
[27]
SOUZAR, LUCENAO, GARRAFAJ, et al. An open, multi-vendor, multi-field-strength brain MR dataset and analysis of publicly available skull stripping methods agreement[J]. NeuroImage, 2018, 170: 482-494.
[28]
LAIW S, HUANGJ B, AHUJAN, et al. Fast and accurate image super-resolution with deep Laplacian Pyramid Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 41(11): 2599-2613.