Figure 1.
Examples that the global structures of the phantom cannot be recovered by total variation (TV) reconstructions from 60° limited-angle scanning. (a) Digital Popeye phantom; (b) the result of simultaneous algebraic reconstruction technique with TV regularization (SART-TV); (c) the result of alternating direction TV minimization (ADTVM); (d) a sinogram of 180° scanning; (e) a reconstructed sinogram by SART-TV; (f) a reconstructed sinogram by ADTVM. When two-thirds of the sinogram data is missing, the reconstructed image exhibits characteristic blurring in the missing angle direction. This can also be seen in the inpainting region of the sinograms as a loss of structure.
Figure 1.
Examples that the global structures of the phantom cannot be recovered by total variation (TV) reconstructions from 60° limited-angle scanning. (a) Digital Popeye phantom; (b) the result of simultaneous algebraic reconstruction technique with TV regularization (SART-TV); (c) the result of alternating direction TV minimization (ADTVM); (d) a sinogram of 180° scanning; (e) a reconstructed sinogram by SART-TV; (f) a reconstructed sinogram by ADTVM. When two-thirds of the sinogram data is missing, the reconstructed image exhibits characteristic blurring in the missing angle direction. This can also be seen in the inpainting region of the sinograms as a loss of structure.
Figure 2.
Schematic of sinogram-inpainting-generative adversarial network (SI-GAN). The input of the network is the limited-angle scanning sinogram. The generation of the 180° scanning sinogram is conditioned on the label sinogram data. The discriminator is also supplied with the input information (the grey rounded rectangle) from the sinogram embedding. In the network structure, we achieve the supervised sinogram inpainting task by adding the loss of sinogram domain and reconstructed image domain. The blue and green rectangles indicate the encoder and decoder of the generator, respectively. The pink rectangles indicate the discriminator. The yellow rounded rectangles are the filter and back-projection operations, which can quickly calculate the reconstruction loss.
Figure 2.
Schematic of sinogram-inpainting-generative adversarial network (SI-GAN). The input of the network is the limited-angle scanning sinogram. The generation of the 180° scanning sinogram is conditioned on the label sinogram data. The discriminator is also supplied with the input information (the grey rounded rectangle) from the sinogram embedding. In the network structure, we achieve the supervised sinogram inpainting task by adding the loss of sinogram domain and reconstructed image domain. The blue and green rectangles indicate the encoder and decoder of the generator, respectively. The pink rectangles indicate the discriminator. The yellow rounded rectangles are the filter and back-projection operations, which can quickly calculate the reconstruction loss.
Figure 3.
Network structure for the generator (up) and the discriminator (down). The generation of the output sinogram is conditional on the input limited-angle sinogram. The discriminator is also supplied with the conditional information from the limited-angle sinogram embedding. Given the skip connection operation, the channels of the decoder layers are twice as large as those in corresponding encoder layers. In addition, we expect the estimated sinogram patches by the generator to fool the discriminator as much as possible. The blue boxes indicate the image blocks generated by the network layers in the generator and discriminator.
Figure 3.
Network structure for the generator (up) and the discriminator (down). The generation of the output sinogram is conditional on the input limited-angle sinogram. The discriminator is also supplied with the conditional information from the limited-angle sinogram embedding. Given the skip connection operation, the channels of the decoder layers are twice as large as those in corresponding encoder layers. In addition, we expect the estimated sinogram patches by the generator to fool the discriminator as much as possible. The blue boxes indicate the image blocks generated by the network layers in the generator and discriminator.
Figure 4.
Real data experimental phantom: Chengdu Dosimetric Phantom, CPET Co. Ltd., Chengdu, China.
Figure 4.
Real data experimental phantom: Chengdu Dosimetric Phantom, CPET Co. Ltd., Chengdu, China.
Figure 5.
Average RMSE of sinograms for different values of and . The color of squares indicates the RMSE value of the estimated sinograms. The RMSE value increases as the color becomes lighter.
Figure 5.
Average RMSE of sinograms for different values of and . The color of squares indicates the RMSE value of the estimated sinograms. The RMSE value increases as the color becomes lighter.
Figure 6.
Sinogram inpainting results for different 90° test data. (a) 180° sinogram data as the ground truth; (b) 90° limited-angle sinogram data; (c) estimated 180° sinogram data by patch-GAN; (d) estimated 180° sinogram data by SI-GAN; (e) error data of (a,c); (f) error data of (a,d). The display window of (a–d) is [0, 1]. The display window of (e,f) is [−0.1, 0.1].
Figure 6.
Sinogram inpainting results for different 90° test data. (a) 180° sinogram data as the ground truth; (b) 90° limited-angle sinogram data; (c) estimated 180° sinogram data by patch-GAN; (d) estimated 180° sinogram data by SI-GAN; (e) error data of (a,c); (f) error data of (a,d). The display window of (a–d) is [0, 1]. The display window of (e,f) is [−0.1, 0.1].
Figure 7.
Results of (a) ground truth, (b) filtered-back projection (FBP), (c) SART-TV, (d) patch-GAN, (e) SI-GAN + FBP, and (f) SI-GAN + SART-TV. (g) Error map 1, which is the difference image of (a,e). (h) Error map 2, which is the difference image of (a,f). The display window of (a–f) is [0, 0.255]. The display window of (g,h) is [−0.1, 0.1].
Figure 7.
Results of (a) ground truth, (b) filtered-back projection (FBP), (c) SART-TV, (d) patch-GAN, (e) SI-GAN + FBP, and (f) SI-GAN + SART-TV. (g) Error map 1, which is the difference image of (a,e). (h) Error map 2, which is the difference image of (a,f). The display window of (a–f) is [0, 0.255]. The display window of (g,h) is [−0.1, 0.1].
Figure 8.
Zoomed-in ROIs of the second and fourth slices. The first column is the ground truth. The third and fifth columns are the reconstructed image by the comparison method and the proposed method, respectively. The second, fourth and sixth columns are the enlarged ROIs of the first, third and fifth columns, respectively. The display window is [0, 0.255].
Figure 8.
Zoomed-in ROIs of the second and fourth slices. The first column is the ground truth. The third and fifth columns are the reconstructed image by the comparison method and the proposed method, respectively. The second, fourth and sixth columns are the enlarged ROIs of the first, third and fifth columns, respectively. The display window is [0, 0.255].
Figure 9.
Sinogram inpainting results for different 60° test data. (a) 60° limited-angle sinogram data; (b) estimated 180° sinogram data by patch-GAN; (c) estimated 180° sinogram data by SI-GAN; (d) sinogram error map of (b); (e) sinogram error map of (c). The display window of (a–c) is [0, 1]. The display window of (d,e) is [−0.1, 0.1].
Figure 9.
Sinogram inpainting results for different 60° test data. (a) 60° limited-angle sinogram data; (b) estimated 180° sinogram data by patch-GAN; (c) estimated 180° sinogram data by SI-GAN; (d) sinogram error map of (b); (e) sinogram error map of (c). The display window of (a–c) is [0, 1]. The display window of (d,e) is [−0.1, 0.1].
Figure 10.
Results of 60° limited angle reconstruction. (a) Ground truth, (b) SART-TV, (c) patch-GAN, (d) SI-GAN + SART-TV and (e) image error map. The display window of (a–d) is [0.01, 0.255]. The display window of (b) is [−0.1, 0.1].
Figure 10.
Results of 60° limited angle reconstruction. (a) Ground truth, (b) SART-TV, (c) patch-GAN, (d) SI-GAN + SART-TV and (e) image error map. The display window of (a–d) is [0.01, 0.255]. The display window of (b) is [−0.1, 0.1].
Figure 11.
The reference images of two representative test slices; (a,b) are the slice 1 and 2 reconstructed using the SART-TV method with full 360 projections. The yellow rectangular boxes are the ROIs. The display windows are [0.002, 0.012].
Figure 11.
The reference images of two representative test slices; (a,b) are the slice 1 and 2 reconstructed using the SART-TV method with full 360 projections. The yellow rectangular boxes are the ROIs. The display windows are [0.002, 0.012].
Figure 12.
Sinogram inpainting results for two test data of the anthropomorphic head phantom. (a) Collected 180° sinogram data as the ground truth; (b) collected 80° limited-angle sinogram data; (c) estimated 180° sinogram data by patch-GAN; (d) estimated 180° sinogram data by SI-GAN; (e) error of (a,c); (f) error of (a,d). The display window of (a–d) is [0, 1]. The display window of (e) and (f) is [−0.1, 0.1].
Figure 12.
Sinogram inpainting results for two test data of the anthropomorphic head phantom. (a) Collected 180° sinogram data as the ground truth; (b) collected 80° limited-angle sinogram data; (c) estimated 180° sinogram data by patch-GAN; (d) estimated 180° sinogram data by SI-GAN; (e) error of (a,c); (f) error of (a,d). The display window of (a–d) is [0, 1]. The display window of (e) and (f) is [−0.1, 0.1].
Figure 13.
Image reconstruction of the anthropomorphic head phantom in 80° limited-angle scanning. From left to right in each row: (a) are the reference images, (b–f) are the images reconstructed from the FBP, SART-TV, patch-GAN, SI-GAN + FBP, and SI-GAN + SART-TV methods. The display window of (a,c–f) is [0.002, 0.012]. The display window of (b) is [0.000, 0.012].
Figure 13.
Image reconstruction of the anthropomorphic head phantom in 80° limited-angle scanning. From left to right in each row: (a) are the reference images, (b–f) are the images reconstructed from the FBP, SART-TV, patch-GAN, SI-GAN + FBP, and SI-GAN + SART-TV methods. The display window of (a,c–f) is [0.002, 0.012]. The display window of (b) is [0.000, 0.012].
Figure 14.
Reconstructed ROIs in slice 1 of the anthropomorphic head phantom. The display window is [0.002, 0.012].
Figure 14.
Reconstructed ROIs in slice 1 of the anthropomorphic head phantom. The display window is [0.002, 0.012].
Table 1.
Establishment of the training dataset.
Table 1.
Establishment of the training dataset.
Procedure: Establishment of the training dataset |
- Step 1.
Each CT image was subjected to value normalization. The image value was rescaled to [0, 0.255]. The normalized images were taken as generating sinogram samples. - Step 2.
We applied Siddon’s ray-tracing algorithm [ 51] to simulate the fan-beam geometry. We generated sinograms for 512 views in 180° with 512 linear detectors with the same size of image pixels. The generated sinograms were taken as labels of the SI-GAN. - Step 3.
To generate the limited-angle sinograms for the inputs of the SI-GAN, we deleted 120° projection data (341 views) with different angle directions. For each sinogram, the sinogram data deletion positions in 180° were 1°–120°, 16°–135°, 31°–150°, 46°–165° and 61°–180°, respectively. Additionally, we added noise to illustrate the practicality of the method. The noise is modeled as Gaussian zero-mean and variance [ 52]: , where indexes the pixels in the projection data and denotes the measured sinogram with added Gaussian noise; the background noise variance was set to .
|
On the basis of the above procedure, 5000 pairs of input and label sinograms with size of 512 × 512 were prepared. |
Table 2.
Parameters set in the real data study.
Table 2.
Parameters set in the real data study.
Parameters | For SI-GAN Training | For SI-GAN Testing |
---|
Detector elements | 512 | 512 |
Detector bin size (mm) | 0.831 | 0.831 |
Distance of source to object (mm) | 483.41 | 462.66 |
Distance of source to detector (mm) | 796.49 | 870.96 |
Tube voltage (kVp) | 120 | 120 |
Tube current (A) | 209 | 210 |
Number of projections | 512 | 512 |
Scanning range (°) | 180 | 180 |
Reconstruction size | 512 × 512 | 512 × 512 |
Table 3.
Quantitative evaluations of results by different algorithms for 90° limited-angle scanning (50 testing images).
Table 3.
Quantitative evaluations of results by different algorithms for 90° limited-angle scanning (50 testing images).
| avg. PSNR | avg. RMSE | avg. NMAD | avg. SSIM |
---|
FBP | 17.234 | 0.0553 | 1.5684 | 0.2631 |
SART-TV | 18.792 | 0.0317 | 0.6512 | 0.7479 |
patch-GAN | 28.369 | 0.0131 | 0.1828 | 0.9433 |
SI-GAN () + FBP | 27.230 | 0.0164 | 0.3493 | 0.8513 |
SI-GAN () + SART-TV | 28.122 | 0.0139 | 0.1933 | 0.9466 |
SI-GAN + FBP | 29.209 | 0.0114 | 0.2689 | 0.8657 |
SI-GAN + SART-TV | 31.052 | 0.0093 | 0.1264 | 0.9648 |
Table 4.
Quantitative evaluations of results for 60° limited-angle scanning (50 testing images).
Table 4.
Quantitative evaluations of results for 60° limited-angle scanning (50 testing images).
| avg. PSNR | avg. RMSE | avg. NMAD | avg. SSIM |
---|
SART-TV | 15.117 | 0.0407 | 0.9306 | 0.6149 |
patch-GAN | 27.460 | 0.0141 | 0.2033 | 0.9327 |
SI-GAN + SART-TV | 29.820 | 0.0097 | 0.1467 | 0.9588 |
Table 5.
Evaluations of the reconstructed images using different algorithms in the anthropomorphic head study.
Table 5.
Evaluations of the reconstructed images using different algorithms in the anthropomorphic head study.
| | PSNR | RMSE | NMAD | SSIM |
---|
Slice 1 | FBP | 13.6388 | 2.52 × 10−3 | 0.9820 | 0.9564 |
SART-TV | 21.9823 | 1.02 × 10−3 | 0.2843 | 0.9939 |
patch-GAN | 29.6305 | 4.04 × 10−4 | 0.1002 | 0.9983 |
SI-GAN + FBP | 24.4512 | 9.55 × 10−4 | 0.3874 | 0.9929 |
SI-GAN + SART-TV | 35.3856 | 2.25 × 10−4 | 0.0504 | 0.9989 |
Slice 2 | FBP | 11.4794 | 2.44 × 10−3 | 0.9954 | 0.9589 |
SART-TV | 23.5963 | 8.71 × 10−4 | 0.2603 | 0.9953 |
patch-GAN | 29.8019 | 4.01 × 10−4 | 0.1162 | 0.9982 |
SI-GAN + FBP | 24.4064 | 9.07 × 10−4 | 0.3882 | 0.9935 |
SI-GAN + SART-TV | 35.1920 | 2.41 × 10−4 | 0.0714 | 0.9987 |
Table 6.
Quantitative evaluations of estimated sinograms by different sinogram inpainting method for tests (10 testing sinograms).
Table 6.
Quantitative evaluations of estimated sinograms by different sinogram inpainting method for tests (10 testing sinograms).
| | avg. RMSE | avg. NMAD |
---|
patch-GAN | Test one (90°) | 0.01094 | 0.02636 |
Test two (60°) | 0.01227 | 0.03570 |
SI-GAN | Test one (90°) | 0.00547 | 0.01297 |
Test two (60°) | 0.00601 | 0.01790 |