Table 2. Comparison of image quality metrics for U-net, c-GAN, M-c-GAN, and the proposed method.
| U-net | c-GAN | M-c-GAN | Proposed | |
|---|---|---|---|---|
| PSNR (dB) | ||||
| Median (Q1/Q3) | 26.4 (24.8/27.5) | 26.5 (24.9/28.0) | 26.6 (24.6/28.1) | 27.5 (26.1/28.6) |
| PIR (mean ± SD) | 11.51%±5.82% | 11.69%±5.12% | 11.65%±5.10% | 13.17±5.84% |
| P value | 0.101 | 0.025 | 0.145 | <0.001 |
| NMSE (%) | ||||
| Median (Q1/Q3) | 0.309 (0.242/0.445) | 0.301 (0.209/0.443) | 0.299 (0.210/0.463) | 0.242 (0.186/0.340) |
| NIR (mean ± SD) | 70.31%±91.26% | 73.55%±86.84% | 74.82%±76.02% | 81.66%±46.35% |
| P value | <0.001 | <0.001 | <0.001 | 0.016 |
| SSIM | ||||
| Median (Q1/Q3) | 0.981 (0.974/0.986) | 0.981 (0.973/0.987) | 0.982 (0.974/0.986) | 0.985 (0.980/0.987) |
| SIR (mean ± SD) | 26.05%±20.46% | 26.02%±19.84% | 26.02±19.81% | 26.72%±20.69% |
| P value | <0.001 | <0.001 | <0.001 | 0.079 |
| CNR (%) | ||||
| Median (Q1/Q3) | 5.69 (2.64/10.9) | 5.79 (2.75/11.0) | 6.15 (2.61/11.1) | 3.88 (1.72/7.66) |
| CIR (mean ± SD) | 71.76%±49.81% | 72.41%±50.34% | 72.62%±46.70% | 80.60%±32.90% |
| P value | 0.277 | 0.177 | 0.526 | 0.214 |
P value, Wilcoxon signed-rank test in which significance level was set to be 5% and proposed bi-c-GAN was compared with U-net, c-GAN, M-c-GAN and low-dose PET. Q1/Q3, 25% and 75% percentile values of quartile. PSNR, peak signal to noise ratio; NMSE, normalized mean square error; SSIM, structural similarity; CNR, contrast noise ratio; c-GAN, conditional generative adversarial network; M-c-GAN, multiple c-GAN; PIR, improvement ratio of PSNR; SIR, improvement ratio of SSIM.