基于卷积神经网络的低剂量CT图像去噪方法 
下载： 1379次
Low-Dose CT Image Denoising Method Based on Convolutional Neural Network
图 & 表
图 1. 网络结构示意图
Fig. 1. Schematic of the network
下载图片 查看原文
图 2. 空洞卷积示意图。(a)步长为1;(b)步长为2;(c)步长为3
Fig. 2. Schematic of dilated convolution. (a) Step size is 1; (b) step size is 2; (c) step size is 3
下载图片 查看原文
图 3. 测试图
Fig. 3. Testing images
下载图片 查看原文
图 4. 去噪效果对比图。(a)原图;(b)低剂量图;(c) RED-CNN;(d)所提网络
Fig. 4. Denoising results comparison images. (a) Original CT image; (b) low-dose CT image; (c) RED-CNN; (d) proposed network
下载图片 查看原文
图 5. 图4(a)中方框区域的放大图。(a)原图;(b)低剂量图;(c) RED-CNN;(d)所提网络
Fig. 5. Fig. 4(a) enlargement of the box region. (a) Original CT image; (b) low-dose CT image; (c) RED-CNN; (d) proposed network
下载图片 查看原文
表 1所有测试图的客观指标
Table1. Objective indexes of all the testing images
| Serial number | Index | Low-dose | RED-CNN | Proposed network |
|---|
| ① | PSNR /dB | 26.27 | 32.35 | 32.73 | | SSIM | 0.8209 | 0.9378 | 0.9378 | | RMSE | 0.0486 | 0.0241 | 0.0231 | | ② | PSNR /dB | 27.49 | 32.59 | 32.34 | | SSIM | 0.8461 | 0.9573 | 0.9536 | | RMSE | 0.0422 | 0.0235 | 0.0241 | | ③ | PSNR /dB | 30.43 | 35.75 | 36.28 | | SSIM | 0.9065 | 0.9621 | 0.9630 | | RMSE | 0.0301 | 0.0163 | 0.0153 | | ④ | PSNR /dB | 24.72 | 31.65 | 33.19 | | SSIM | 0.7711 | 0.9184 | 0.9199 | | RMSE | 0.0581 | 0.0261 | 0.0219 | | ⑤ | PSNR /dB | 22.00 | 27.07 | 28.80 | | SSIM | 0.7252 | 0.8922 | 0.8968 | | RMSE | 0.0794 | 0.0443 | 0.0363 | | ⑥ | PSNR /dB | 20.82 | 26.72 | 28.16 | | SSIM | 0.6995 | 0.8950 | 0.8975 | | RMSE | 0.0909 | 0.0461 | 0.0391 | | ⑦ | PSNR /dB | 29.30 | 36.85 | 36.69 | | SSIM | 0.8843 | 0.9660 | 0.9653 | | RMSE | 0.0343 | 0.0144 | 0.0146 | | ⑧ | PSNR /dB | 29.78 | 35.51 | 35.29 | | SSIM | 0.9191 | 0.9707 | 0.9695 | | RMSE | 0.0324 | 0.0168 | 0.0172 | | ⑨ | PSNR /dB | 27.59 | 37.75 | 37.53 | | SSIM | 0.8337 | 0.9618 | 0.9605 | | RMSE | 0.0417 | 0.0130 | 0.0133 | | ⑩ | PSNR /dB | 27.27 | 32.37 | 33.53 | | SSIM | 0.8607 | 0.9341 | 0.9348 | | RMSE | 0.0433 | 0.0241 | 0.0211 | | Average | PSNR /dB | 26.57 | 32.86 | 33.45 | | SSIM | 0.8267 | 0.9395 | 0.9399 | | RMSE | 0.0501 | 0.0249 | 0.0226 |
|
查看原文
表 2复杂度对比
Table2. Comparision of complexity
| Item | RED-CNN | Proposed network |
|---|
| Complexity | 1848000 | 216864 | | Time consumption (CPU) /s | 12.472 | 3.908 | | Time consumption (GPU) /s | 0.180 | 0.061 |
|
查看原文
表 3不同网络结构的对比
Table3. Comparision of different network structures
| Structure description | Default | Without BN and residual learning | Without concatenating feature maps | Without dilated convolution |
|---|
| PSNR /dB | 33.45 | 32.84 | 33.33 | 33.24 | | SSIM | 0.9399 | 0.9369 | 0.9382 | 0.9373 | | RMSE | 0.0226 | 0.0245 | 0.0229 | 0.0231 |
|
查看原文
表 4每个ConvBlock包含不同组数Conv-BN-ReLU时对去噪效果的影响
Table4. Impact on denoising performance when each ConvBlock contains different numbers of Conv-BN-ReLU
| Structure description | Numbers of Conv-BN-ReLU contained in each ConvBlock |
|---|
| 1 | 2 | 3 |
|---|
| PSNR /dB | 33.27 | 33.45 | 33.33 | | SSIM | 0.9290 | 0.9399 | 0.9397 | | RMSE | 0.0230 | 0.0226 | 0.0229 |
|
查看原文
章云港, 易本顺, 吴晨玥, 冯雨. 基于卷积神经网络的低剂量CT图像去噪方法[J]. 光学学报, 2018, 38(4): 0410003. Yungang Zhang, Benshun Yi, Chenyue Wu, Yu Feng. Low-Dose CT Image Denoising Method Based on Convolutional Neural Network[J]. Acta Optica Sinica, 2018, 38(4): 0410003.
PDF全文
