Fig. 1. Cascaded dual-domain crossed encoding and decoding network structure
Fig. 2. Image-domain and K-domain convolution structures. (a) Image-domain convolution; (b) K-domain convolution
Fig. 3. Dual domain cascaded modes. (a) Parallel cascade; (b) crossed cascade
Fig. 4. Comparison of 4-fold under-sampled and full-sampled K-domain data visualization.(a) Under-sampled K-space; (b) full-sampled K-space
Fig. 5. Reconstruction results of under-sampled MRI. (a) 4× under-sampled zero-filled magnetic resonance images; (b) magnetic resonance images reconstructed by U-Net; (c) magnetic resonance images reconstructed by DPC-Net; (d) magnetic resonance images reconstructed by KIKI-Net; (e) full-sampled true value
Fig. 6. Reconstruction error. (a) 4× under-sampled zero filled MRI error; (b) reconstruction error of U-Net; (c) reconstruction error of DPC-Net; (d) reconstruction error of KIKI-Net
Objective indicator | DPC-Net(dual-domain parallel) | DPC-Net(dual-domain crossover) |
---|
NMSE | 0.0330 | 0.0325 | SSIM | 0.7363 | 0.7268 | PSNR /dB | 32.16 | 31.37 |
|
Table 1. Performance comparison of two cascaded methods
Condition | NMSE | SSIM | PSNR /dB |
---|
| 0.0338 | 0.7310 | 31.83 | | 0.0330 | 0.7363 | 32.16 | | 0.0341 | 0.7287 | 32.02 | | 0.0348 | 0.7213 | 31.78 | | 0.0352 | 0.7189 | 31.01 |
|
Table 2. Performance comparison of loss function weights in reconstruction
Number of cascades | NMSE | SSIM | PSNR /dB |
---|
U-Net | DPC-Net | U-Net | DPC-Net | U-Net | DPC-Net |
---|
1 | 0.0357 | 0.0350 | 0.7271 | 0.7363 | 32.15 | 32.16 | 2 | 0.0351 | 0.0341 | 0.7298 | 0.7417 | 32.23 | 32.33 | 3 | 0.0348 | 0.0340 | 0/7354 | 0.7440 | 32.67 | 32.36 | 4 | 0.0342 | 0.0338 | 0.7435 | 0.7460 | 33.03 | 32.44 | 5 | 0.0347 | 0.0326 | 0.7426 | 0.7498 | 32.78 | 32.98 | 6 | 0.0349 | 0.0318 | 0.7417 | 0.7532 | 32.53 | 33.16 | 7 | 0.0352 | 0.0324 | 0.7387 | 0.7520 | 32.47 | 33.02 |
|
Table 3. Objective indicators of reconstruction results of cascaded dual-domain codec network
Parameter | Network | 2-fold | 4-fold | 8-fold |
---|
NMSE | Zero filled | 0.0323 | 0.0522 | 0.1091 | U-Net | 0.0234 | 0.0349 | 0.0604 | KIKI-Net | 0.0231 | 0.0326 | 0.0602 | DPC-Net | 0.0221 | 0.0318 | 0.0601 | SSIM | Zero filled | 0.7820 | 0.6559 | 0.548 | U-Net | 0.826 | 0.741 | 0.623 | KIKI-Net | 0.831 | 0.748 | 0.634 | DPC-Net | 0.845 | 0.753 | 0.648 | PSNR /dB | Zero filled | 31.76 | 29.59 | 21.63 | U-Net | 34.60 | 32.53 | 27.5 | KIKI-Net | 34.98 | 33.01 | 27.9 | DPC-Net | 35.04 | 33.16 | 28.8 |
|
Table 4. Reconstruction performance of different networks under different acceleration magnifications