题目: Analysis of deep complex-valued convolutional neural networks for MRI reconstruction and phase-focused applications【深度复值卷积神经网络在MRI重建和相位聚焦应用中的分析】
对我有用的部分:
1. 对比了不同复激活函数的性能
2. 对UNet在里面的重要性做了评估
3. 关注了油水分离等对相位信息有高需求的任务
4. 代码开源
5. 依然提出了DL生成的图像平滑的问题
6. 认为复数网络结构也适用于超声
7. 测试了网络的深度和宽度带来的影响
网络结构:
FIGURE 1 A, One iteration of the unrolled network based on the iterative shrinkage-thresholding algorithm.7,8 This consists of an update block, which uses the MRI model to enforce data consistency with the physically measured k-space samples. Then, a residual structure block
is used to denoise the input image to produce the output image ym+1. Each convolutional layer except for the last is followed by a ReLU and a
complex-valued activation function (see Methods section B). B, The second reconstruction network architecture, which is based on the original
U-Net for segmentation.15 Every orange box depicts a multi-channel feature map. The number of channels is denoted on top of each feature map representation. Each arrow denotes a different operation, as depicted by the right-hand legend. This network uses contracting and expanding paths to capture information. Note: ReLU, rectified linear unit
结果
1. 激活函数 :CReLU最好
2. 网络的深度和宽度对实验结果的影响
实数和复数网络都是随着网络宽度和深度的增加,性能更好
Supporting Information Figure S2. Here, we replotted Figure 3a by making the x-axis the number of total feature maps instead of the number of total parameters. We applied a linear trendline to each series. Here, we can see that the performance of the real-valued model actually never surpasses the complex-valued model, despite having the same or greater number of feature maps. Additionally, the feature map size, feature map properties, and number of unrolled iterations were equivalent for these experiments.
FIGURE 3B, Performance of the unrolled network as a function of network depth on a test dataset. Here, the number of feature maps is kept constant at 128 and 90 for the complex and real networks, respectively, whereas the number of iterations is varied for each network. The number of iterations in compressed sensing does not change; however, its performance is plotted for reference.
3. UNet(在复数和实数网络的不同表现)
指标 | 复数值CNN (±标准差) | 实数值CNN (±标准差) | 比较结果 |
---|---|---|---|
PSNR (dB) | 35.28 ± 2.34 | 35.00 ± 1.95 | 复数值CNN更优 |
NRMSE | 0.16 ± 0.03 | 0.17 ± 0.02 | 复数值CNN更优 |
SSIM | 0.90 ± 0.05 | 0.90 ± 0.02 | 两者相当 |