Figure 1.
Schematic illustration of the DeepCB-XLCT network architecture: the 3D deep encoder–decoder (3D-En–Decoder) network has a 3D-Encoder network and a 3D-Decoder network. The 3D encoder consists of several convolution layers, followed by batch norm, ReLU activation function, and pooling. The 3D-Decoder has several upsampling layers followed by convolution, batch norm, and ReLU activation function. There are two full connection layers between the 3D-Encoder and the 3D-Decoder.
Figure 1.
Schematic illustration of the DeepCB-XLCT network architecture: the 3D deep encoder–decoder (3D-En–Decoder) network has a 3D-Encoder network and a 3D-Decoder network. The 3D encoder consists of several convolution layers, followed by batch norm, ReLU activation function, and pooling. The 3D-Decoder has several upsampling layers followed by convolution, batch norm, and ReLU activation function. There are two full connection layers between the 3D-Encoder and the 3D-Decoder.
Figure 2.
The cylinder phantom used in the training data simulation with two targets and three targets (a) two targets, (b) three targets, the region within the red circle is the entire reconstruction region, the region within the green rectangular box is the target region.
Figure 2.
The cylinder phantom used in the training data simulation with two targets and three targets (a) two targets, (b) three targets, the region within the red circle is the entire reconstruction region, the region within the green rectangular box is the target region.
Figure 3.
The cylinder phantom used in simulation studies. Edge-to-edge distance between the two targets: (a) 1 mm, (b) 1.5 mm, (c) 2 mm, the red circle is the boundary of the phantom, the green line is the centerline of the phantom.
Figure 3.
The cylinder phantom used in simulation studies. Edge-to-edge distance between the two targets: (a) 1 mm, (b) 1.5 mm, (c) 2 mm, the red circle is the boundary of the phantom, the green line is the centerline of the phantom.
Figure 4.
The schematic diagram of the CB-XLCT system.
Figure 4.
The schematic diagram of the CB-XLCT system.
Figure 5.
Configuration for the two-target phantom experiment. The concentrations of the two targets were both 50 mg/mL, and the edge-to-edge distances of the two targets were 2.3 mm, 1.7 mm, and 1.0 mm.
Figure 5.
Configuration for the two-target phantom experiment. The concentrations of the two targets were both 50 mg/mL, and the edge-to-edge distances of the two targets were 2.3 mm, 1.7 mm, and 1.0 mm.
Figure 6.
Tomographic images of the targets positioned at varying distances were reconstructed. The first through fifth columns present the results obtained using ADFISTA, MAP, T-FISTA, ADMLEM, and the proposed DeepCB-XLCT network, respectively. The first to third rows correspond to reconstructions with edge-to-edge distances of 2.0 mm, 1.5 mm, and 1.0 mm, respectively. The red circle is the boundary of the phantom, the green line is the centerline of the reconstruction image and the yellow circle represents the true target region.
Figure 6.
Tomographic images of the targets positioned at varying distances were reconstructed. The first through fifth columns present the results obtained using ADFISTA, MAP, T-FISTA, ADMLEM, and the proposed DeepCB-XLCT network, respectively. The first to third rows correspond to reconstructions with edge-to-edge distances of 2.0 mm, 1.5 mm, and 1.0 mm, respectively. The red circle is the boundary of the phantom, the green line is the centerline of the reconstruction image and the yellow circle represents the true target region.
Figure 7.
Profiles along the green dotted line in
Figure 5 with the edge-to-edge distances of 1.0 mm. (
a–
e) The profiles of reconstruction results achieved with ADFISTA, MAP, T-FISTA, ADMLEM, and the proposed DeepCB-XLCT network, respectively.
Figure 7.
Profiles along the green dotted line in
Figure 5 with the edge-to-edge distances of 1.0 mm. (
a–
e) The profiles of reconstruction results achieved with ADFISTA, MAP, T-FISTA, ADMLEM, and the proposed DeepCB-XLCT network, respectively.
Figure 8.
Tomographic images of the targets positioned at different distances with different noise levels were reconstructed based on the proposed DeepCB-XLCT network. The red circle is the boundary of the phantom and the yellow circle represents the true target region. The first to fourth columns are the reconstruction results with the SNRs of 30 dB, 25 dB, 20 dB, and 15 dB, respectively. The first to third row are the results reconstructed with the edge-to-edge distances of 2.0 mm, 1.5 mm, and 1.0 mm, respectively.
Figure 8.
Tomographic images of the targets positioned at different distances with different noise levels were reconstructed based on the proposed DeepCB-XLCT network. The red circle is the boundary of the phantom and the yellow circle represents the true target region. The first to fourth columns are the reconstruction results with the SNRs of 30 dB, 25 dB, 20 dB, and 15 dB, respectively. The first to third row are the results reconstructed with the edge-to-edge distances of 2.0 mm, 1.5 mm, and 1.0 mm, respectively.
Figure 9.
Tomographic images of the targets positioned at different distances were reconstructed in the ablation experiment. The red circle is the boundary of the phantom and the yellow circle represents the true target region. The first to fourth rows are the results obtained based on the proposed DeepCB-XLCT network without skip connection and ROILoss, the proposed DeepCB-XLCT network without skip connection, the proposed DeepCB-XLCT network without ROILoss, and the proposed DeepCB-XLCT network, respectively. The first to third columns are the results reconstructed with the edge-to-edge distances of 2.0 mm, 1.5 mm, and 1.0 mm, respectively.
Figure 9.
Tomographic images of the targets positioned at different distances were reconstructed in the ablation experiment. The red circle is the boundary of the phantom and the yellow circle represents the true target region. The first to fourth rows are the results obtained based on the proposed DeepCB-XLCT network without skip connection and ROILoss, the proposed DeepCB-XLCT network without skip connection, the proposed DeepCB-XLCT network without ROILoss, and the proposed DeepCB-XLCT network, respectively. The first to third columns are the results reconstructed with the edge-to-edge distances of 2.0 mm, 1.5 mm, and 1.0 mm, respectively.
Figure 10.
Tomographic images of three targets were reconstructed using various methods. The red circle is the boundary of the phantom and the yellow circle represents the true target region. The first column depicts the true locations of the targets. The second to sixth columns present results obtained using ADFISTA, MAP, T-FISTA, ADMLEM, and the proposed DeepCB-XLCT network, respectively. The first to third rows show the reconstructions for the three targets positioned at different locations.
Figure 10.
Tomographic images of three targets were reconstructed using various methods. The red circle is the boundary of the phantom and the yellow circle represents the true target region. The first column depicts the true locations of the targets. The second to sixth columns present results obtained using ADFISTA, MAP, T-FISTA, ADMLEM, and the proposed DeepCB-XLCT network, respectively. The first to third rows show the reconstructions for the three targets positioned at different locations.
Figure 11.
Tomographic images of the targets positioned at varying distances were reconstructed based on different algorithms for phantom experiments. The first to third rows display the fused XLCT/CT tomographic images for the EEDs of 2.3, 1.7, and 1.0 mm, respectively. The reconstructions obtained from ADFISTA, MAP, T-FISTA, ADMLEM, and the proposed DeepCB-XLCT network are presented in the first through fifth columns, respectively.
Figure 11.
Tomographic images of the targets positioned at varying distances were reconstructed based on different algorithms for phantom experiments. The first to third rows display the fused XLCT/CT tomographic images for the EEDs of 2.3, 1.7, and 1.0 mm, respectively. The reconstructions obtained from ADFISTA, MAP, T-FISTA, ADMLEM, and the proposed DeepCB-XLCT network are presented in the first through fifth columns, respectively.
Figure 12.
Tomographic images reconstructed based on different algorithms for mice experiments. (a) The CT projection image, the region between the two red lines was used for reconstruction; (b) the reconstructed CT slice at the height of the green line of (a); (c–g) reconstructions obtained using ADFISTA, MAP, T-FISTA, ADMLEM, and the proposed DeepCB-XLCT network, respectively; (h–l) the corresponding fusion results of XLCT and CT.
Figure 12.
Tomographic images reconstructed based on different algorithms for mice experiments. (a) The CT projection image, the region between the two red lines was used for reconstruction; (b) the reconstructed CT slice at the height of the green line of (a); (c–g) reconstructions obtained using ADFISTA, MAP, T-FISTA, ADMLEM, and the proposed DeepCB-XLCT network, respectively; (h–l) the corresponding fusion results of XLCT and CT.
Table 1.
CNR and DICE metrics for the reconstruction results of two targets with varying EEDs obtained from numerical simulations.
Table 1.
CNR and DICE metrics for the reconstruction results of two targets with varying EEDs obtained from numerical simulations.
Methods | 2 mm | 1.5 mm | 1.0 mm |
---|
CNR | DICE | CNR | DICE | CNR | DICE |
---|
ADFISTA | 2.47 | 0.44 | 2.58 | 0.41 | 2.66 | 0.42 |
MAP | 2.83 | 0.38 | 3.01 | 0.37 | 3.16 | 0.38 |
T-FISTA | 1.06 | 0.26 | 2.39 | 0.47 | 2.31 | 0.45 |
ADMLEM | 2.19 | 0.44 | 2.39 | 0.46 | 2.56 | 0.48 |
Proposed method | 4.32 | 0.89 | 4.69 | 0.97 | 4.49 | 0.96 |
Table 2.
CNR and DICE metrics for the reconstruction results of two targets with varying EEDs under different noise levels from numerical simulations.
Table 2.
CNR and DICE metrics for the reconstruction results of two targets with varying EEDs under different noise levels from numerical simulations.
Noise Level | 2.0 mm | 1.5 mm | 1.0 mm |
---|
CNR | DICE | CNR | DICE | CNR | DICE |
---|
30 dB | 4.32 | 0.89 | 4.69 | 0.97 | 4.49 | 0.96 |
25 dB | 4.08 | 0.81 | 4.49 | 0.95 | 3.94 | 0.38 |
20 dB | 3.60 | 0.78 | 4.04 | 0.90 | 3.76 | 0.45 |
15 dB | 2.93 | 0.69 | 3.39 | 0.80 | 3.24 | 0.81 |
Table 3.
CNR and DICE metrics for the reconstruction results of two targets with varying EEDs in the ablation experiment.
Table 3.
CNR and DICE metrics for the reconstruction results of two targets with varying EEDs in the ablation experiment.
Methods | 2.0 mm | 1.5 mm | 1.0 mm |
---|
CNR | DICE | CNR | DICE | CNR | DICE |
---|
Without skip connection and ROILoss | 1.99 | 0.47 | 2.19 | 0.58 | 2.20 | 0.57 |
Without skip connection | 4.05 | 0.86 | 4.47 | 0.96 | 4.19 | 0.94 |
Without ROILoss | 3.31 | 0.85 | 3.69 | 0.90 | 3.55 | 0.91 |
Proposed method | 4.32 | 0.89 | 4.69 | 0.97 | 4.49 | 0.96 |
Table 4.
CNR and DICE metrics for the reconstruction results of two targets with varying EEDs in phantom experiments.
Table 4.
CNR and DICE metrics for the reconstruction results of two targets with varying EEDs in phantom experiments.
Methods | 2.3 mm | 1.7 mm | 1.0 mm |
---|
CNR | DICE | CNR | DICE | CNR | DICE |
---|
ADFISTA | 3.94 | 0.56 | 4.03 | 0.48 | 4.76 | 0.41 |
MAP | 3.62 | 0.49 | 4.88 | 0.56 | 4.36 | 0.45 |
T-FISTA | 3.41 | 0.57 | 7.64 | 0.73 | 3.86 | 0.53 |
ADMLEM | 7.64 | 0.75 | 7.92 | 0.78 | 4.86 | 0.58 |
Proposed method | 7.57 | 0.56 | 11.09 | 0.79 | 12.06 | 0.71 |
Table 5.
CNR and DICE metrics for the reconstruction results of two targets in mouse experiments.
Table 5.
CNR and DICE metrics for the reconstruction results of two targets in mouse experiments.
| ADFISTA | MAP | T-FISTA | ADMLEM | Proposed Method |
---|
CNR | 3.15 | 1.61 | 2.55 | 3.63 | 4.33 |
DICE | 0.32 | 0.16 | 0.24 | 0.35 | 0.60 |
Table 6.
Reconstruction time with different methods in mice experiments.
Table 6.
Reconstruction time with different methods in mice experiments.
Methods | ADFISTA | MAP | T-FISTA | ADMLEM | Proposed Method |
---|
Reconstruction time/s | 14.9 | 48.1 | 48.6 | 15.8 | 1 |