Fig 1.
CT slice reconstruction based on the proposed approach that reconstructs 2m − 1 middle slices from two adjacent slices.
The green slices are newly reconstructed ones. Then we copy the original slice as the first slice of the reconstructed data to generate the whole CT data.
Fig 2.
Architecture of the proposed approach for CT slice reconstruction.
The neural networks deduces multiple middle slices from two neighboring slices. A parallel architecture is adopted to prevent the different target slices from influencing one another, so that the target slices are computed separately.
Fig 3.
Outline of layers of each U-net shown in Fig 2.
BN here means batch normalization layer.
Fig 4.
Histograms of the whole CT image and the liver alone.
The blue and red areas show the distributions of the CT values across the whole slice and inside the liver. The X-axis is the normalized CT values, and the Y-axis is the frequency.
Fig 5.
Training process of organ-oriented CT reconstruction.
Each neural networks is trained for each specified organ. The organ areas in the original output are then replaced with the results of these organ-oriented networks. HR: high-quality result; LR: low-quality result. Here networks means the proposed parallel U-net architecture shown in Fig 2.
Table 1.
Mean absolute error (MAE) of the reconstruction of one, two, three, four, and five slices by the proposed parallel U-net architecture.
Fig 6.
Comparison among linear interpolation, the other U-net method [2], and the proposed parallel U-net architecture (all p < 0.01).
Vertical axis shows the absolute errors of the reconstructed slices.
Table 2.
Comparison between linear interpolation, U-net architecture [2], and the proposed parallel U-net architecture: The bold numbers show the best results.
Fig 7.
Comparison of linear interpolation, the proposed parallel U-net architecture, and the proposed organ-oriented method (all p < 0.01).
Vertical axis shows the absolute errors of the liver region in the reconstructed slices.
Fig 8.
Comparison among linear interpolation, the proposed parallel U-net architecture, and the proposed organ-oriented method (all p < 0.01).
Vertical axis shows the absolute errors of the left kidney region in the reconstructed slices.
Fig 9.
Comparison among linear interpolation, the proposed parallel U-net architecture, and the proposed organ-oriented method (all p < 0.01).
Vertical axis shows the absolute errors of the right kidney region in the reconstructed slices.
Fig 10.
Comparison among linear interpolation, the proposed parallel U-net architecture, and the proposed organ-oriented method (all p < 0.01).
Vertical axis shows the absolute errors of the stomach region in the reconstructed slices.
Table 3.
Mean absolute error (MAE): Comparison between the proposed U-net architecture-based method and the organ-oriented approach.
Table 4.
Structural similarity (SSIM): Comparison between the proposed U-net architecture-based method and the organ-oriented approach.
Fig 11.
Examples of the comparison among the proposed U-net architecture-based approach, the proposed organ-oriented method, and linear interpolation.
From left to right, the images shown in grayscale are the ground truth, the result computed by linear interpolation, the result deduced by the proposed parallel U-net architecture, and the result obtained by the proposed organ-oriented reconstruction. The images shown in color maps are difference images between the compared methods’ results and the ground truth. In the color maps, dark blue indicates the smallest difference and red indicated the largest difference. The window size is adjusted to [−400, 400] for soft tissues.
Fig 12.
Examples of the comparison among the proposed U-net architecture-based approach, the proposed organ-oriented method, and linear interpolation.
From left to right, the images shown in grayscale are the ground truth, the result computed by linear interpolation, the result deduced by the proposed parallel U-net architecture, and the result obtained by the proposed organ-oriented reconstruction. The images shown in color maps are difference images between the compared methods’ results and the ground truth. In the color maps, dark blue indicates the smallest difference and red indicated the largest difference. The window size is adjusted to [−400, 400] for soft tissues.