-
基于深度学习的磁共振成像(magnetic resonance imaging, MRI)方法需要大规模、高质量的病患数据样本集进行预训练. 然而, 由于病患隐私及设备等因素限制, 获取大规模、高质量的磁共振数据集在实际临床应用中面临挑战. 本文提出一种新的基于深度学习的欠采样磁共振图像重建方法, 该方法无需预训练、不依赖训练数据集, 而是充分利用待重建的目标MR图像的结构先验和支撑先验, 并将其引入深度图像先验(deep image prior, DIP)框架, 从而削减对训练数据集的依赖, 提升学习效率. 基于参考图像与目标图像的相似性, 采用高分辨率参考图像作为深度网络输入, 将结构先验信息引入网络; 将参考图像在小波域中幅值大的系数索引集作为目标图像的已知支撑集, 构造正则化约束项, 将网络训练转化为网络参数的最优化求解过程. 实验结果表明, 本文方法可由欠采样 k空间数据重建得到更精确的磁共振图像, 且在保留组织特征、细节纹理方面具有明显优势.Magnetic resonance imaging (MRI) method based on deep learning needs large-quantity and high-quality patient-based datasets for pre-training. However, this is a challenge to the clinical applications because it is difficult to obtain a sufficient quantity of patient-based MR datasets due to the limitation of equipment and patient privacy concerns. In this paper, we propose a novel undersampled MRI reconstruction method based on deep learning. This method does not require any pre-training procedures and does not depend on training datasets. The proposed method is inspired by the traditional deep image prior (DIP) framework, and integrates the structure prior and support prior of the target MR image to improve the efficiency of learning. Based on the similarity between the reference image and the target image, the high-resolution reference image obtained in advance is used as the network input, thereby incorporating the structural prior information into network. By taking the coefficient index set of the reference image with large amplitude in the wavelet domain as the known support of the target image, the regularization constraint term is constructed, and the network training is transformed into the optimization process of network parameters. Experimental results show that the proposed method can obtain more accurate reconstructions from undersampled k-space data, and has obvious advantages in preserving tissue features and detailed texture.
-
Keywords:
- magnetic resonance imaging/
- undersampled image reconstruction/
- deep image prior/
- support prior
[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] -
参数 图像 目标图像1 目标图像2 目标图像3 网络超参数 学习率 0.005 0.003 0.003 L 6 6 6 $n_{\rm{d}}$ [32, 64, 64, 64, 128, 128] [16, 32, 64, 64, 128, 128] [16, 32, 64, 64, 128, 128] $n_{\rm{u}}$ [32, 64, 64, 64, 128, 128] [16, 32, 64, 64, 128, 128] [16, 32, 64, 64, 128, 128] $n_{\rm{s}}$ [16, 16, 16, 16, 16, 16] [16, 16, 16, 16, 16, 16] [16, 16, 16, 16, 16, 16] $k_{\rm{d}}$ [3, 3, 3, 3, 3, 3] [3, 3, 3, 3, 3, 3] [3, 3, 3, 3, 3, 3] $k_{\rm{u}}$ [3, 3, 3, 3, 3, 3] [3, 3, 3, 3, 3, 3] [3, 3, 3, 3, 3, 3] $k_{\rm{s}}$ [1, 1, 1, 1, 1, 1] [1, 1, 1, 1, 1, 1] [1, 1, 1, 1, 1, 1] 迭代次数 5000 3000 3000 小波参数 小波函数 Haar Haar Haar 分解层数 7 8 8 P 13500 55000 52000 λ 1 × 10–10 1 × 10–6 1 × 10–9 
下载:
导出CSV
待重建MR图像 重建方法 10% 20% 相对误差/% PSNR/dB SSIM 相对误差/% PSNR/dB SSIM 目标图像1 零填充 35.64 18.7365 0.5543 16.14 25.6185 0.7135 CS-WS 34.53 19.0101 0.5590 13.74 27.0139 0.7998 DIP 32.51 19.5344 0.6666 12.49 27.8440 0.8989 本文方法 19.89 23.8004 0.8172 9.43 30.2852 0.9793 目标图像2 零填充 19.66 22.5625 0.7550 13.36 25.9204 0.8065 CS-WS 17.14 23.7530 0.7654 10.34 28.1426 0.8507 DIP 15.09 24.9087 0.8473 5.41 33.7814 0.9619 本文方法 7.51 30.9288 0.9454 3.48 37.5915 0.9788 目标图像3 零填充 17.64 23.5716 0.7857 12.76 26.3807 0.8266 CS-WS 15.24 24.8395 0.8139 9.87 28.6144 0.8774 DIP 14.90 25.0660 0.8585 5.43 33.8264 0.9607 本文方法 7.31 31.2228 0.9491 3.53 37.5488 0.9806 待重建MR图像 重建方法 30% 40% 相对误差/% PSNR/dB SSIM 相对误差/% PSNR/dB SSIM 目标图像1 零填充 11.53 28.5361 0.7657 7.71 32.0337 0.8087 CS-WS 9.40 30.3126 0.8689 6.68 33.2736 0.9136 DIP 9.05 30.6386 0.9342 6.98 32.8933 0.9605 本文方法 7.06 32.7970 0.9590 5.64 34.7395 0.9728 目标图像2 零填充 4.61 35.1677 0.8677 3.46 37.6630 0.8830 CS-WS 2.61 40.0883 0.9377 1.95 42.6338 0.9438 DIP 2.80 39.4992 0.9858 2.33 41.1008 0.9892 本文方法 2.08 42.0469 0.9905 1.73 43.6437 0.9931 目标图像3 零填充 4.31 35.8182 0.8813 3.24 38.2961 0.8997 CS-WS 2.45 40.7193 0.9444 1.84 43.1958 0.9601 DIP 3.22 38.5632 0.9824 2.54 40.4324 0.9885 本文方法 2.05 42.2484 0.9908 1.64 44.2051 0.9939 下载:
导出CSV
待重建MR图像 采样模板(采样率) 重建方法 相对误差/% PSNR/dB SSIM 目标图像1 径向(20%) 零填充 14.34 26.6426 0.7852 CS-WS 12.11 28.1129 0.8519 DIP 10.95 28.9845 0.9169 本文方法 7.92 31.7960 0.9547 变密度(30%) 零填充 15.03 26.2373 0.7735 CS-WS 11.15 28.8264 0.8662 DIP 9.26 30.4441 0.9321 本文方法 6.43 33.6199 0.9648 目标图像2 径向(10%) 零填充 8.05 30.3202 0.7960 CS-WS 5.73 33.2803 0.9010 DIP 4.40 35.5814 0.9662 本文方法 3.54 37.4434 0.9760 变密度(20%) 零填充 6.47 32.2213 0.8778 CS-WS 3.56 37.4195 0.9625 DIP 3.15 38.4764 0.9821 本文方法 2.57 40.2243 0.9850 目标图像3 径向(10%) 零填充 7.03 31.5576 0.8177 CS-WS 5.18 34.2111 0.9158 DIP 4.73 35.0303 0.9629 本文方法 3.55 37.4852 0.9751 变密度(20%) 零填充 5.82 33.2106 0.8978 CS-WS 3.27 38.2161 0.9675 DIP 3.01 38.9786 0.9819 本文方法 2.54 40.4229 0.9852 下载:
导出CSV
采样模板(采样率) 重建方法 相对误差/% PSNR/dB SSIM 径向 (20%) DIP + Ref 8.88 30.9180 0.9478 DIP + Sup 10.35 29.4739 0.9264 DIP + Ref + Sup 8.24 31.5530 0.9512 DIP + Ref + Sup + Cor 7.92 31.7960 0.9547 变密度 (30%) DIP + Ref 7.38 32.4116 0.9583 DIP + Sup 9.08 30.6104 0.9583 DIP + Ref + Sup 6.71 33.2360 0.9620 DIP + Ref + Sup + Cor 6.43 33.6199 0.9648 下载:
导出CSV
待重建MR图像 重建方法 计算时间 10% 20% 30% 40% 目标图像1 CS-WS 46 s 46 s 49 s 46 s DIP 2 min 53 s 2 min 43 s 2 min 47 s 2 min 45 s 本文方法 3 min 55 s 3 min 56 s 3 min 54 s 3 min 56 s 图11(d)所示目标图像 CS-WS 42 s 42 s 45 s 44 s DIP 2 min 33 s 2 min 35 s 2 min 35 s 2 min 34 s 本文方法 3 min 14 s 3 min 15 s 3 min 15 s 3 min 13 s 下载:
导出CSV
-
[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34]
下载:
计量
- 文章访问数:4103
- PDF下载量:64
- 被引次数:0