Table 2. Evaluative comparison across deep learning based, classical and our proposed methods on the simulated foam phantom (Pelt et al., 2022 ▸), LoDoInd (Shi et al., 2024 ▸) and real-world experimental dataset TomoBank (De Carlo et al., 2018 ▸). Performance metrics, specifically PSNR / SSIM, are averaged across all slices and highlighted in bold for the best outcomes. Classical methods include an additional post-processing step before reconstruction, with parameters optimized on training data. All methods were designed to have similar numbers of trainable parameters for the same network architecture and were trained for a consistent number of epochs to ensure fairness in comparison.
| Parameters | MBIR/Kazantsev | Münch | Miqueles | Vo | ||||
|---|---|---|---|---|---|---|---|---|
| I0 / Pring / Pzinger | Corrupted | Post-proc. | Sinogram proc. | Followed by post-proc. | Ours | |||
| Dataset: foam (512, 512, 512) | ||||||||
| 30 / 0.1 / 0.001 | 1.14 / 0.23 | 19.55 / 0.70 | 17.97 / 0.44 | 19.24 / 0.70 | 19.72 / 0.71 | 20.34 / 0.72 | 20.37 / 0.71 | 21.80 / 0.76 |
| 100 / 0.1 / 0.001 | 4.07 / 0.27 | 21.10 / 0.69 | 19.58 / 0.51 | 21.81 / 0.75 | 22.26 / 0.76 | 23.70 / 0.77 | 23.57 / 0.78 | 24.24 / 0.79 |
| 100 / 0.1 / 0 | 4.97 / 0.29 | 22.67 / 0.77 | 19.63 / 0.51 | 23.00 / 0.77 | 22.39 / 0.76 | 23.32 / 0.77 | 23.38 / 0.77 | 24.41 / 0.79 |
| 100 / 0.2 / 0 | 3.04 / 0.26 | 22.21 / 0.76 | 19.59 / 0.51 | 22.50 / 0.76 | 22.25 / 0.75 | 22.78 / 0.76 | 23.20 / 0.77 | 24.09 / 0.78 |
| 100 / 0 / 0.002 | 5.58 / 0.29 | 23.10 / 0.77 | 19.66 / 0.51 | – | 24.26 / 0.78 | – | 23.86 / 0.77 | 24.87 / 0.79 |
| Dataset: LoDoInd (2000, 1250, 1250) | ||||||||
| 500 / 0.1 / 0.001 | 5.91 / 0.21 | 36.30 / 0.91 | 36.22 / 0.90 | 36.11 / 0.91 | 36.93 / 0.92 | 36.93 / 0.92 | 36.50 / 0.91 | 38.65 / 0.93 |
| Dataset: TomoBank (2160, 2560, 2560) | ||||||||
| NA | 17.77 / 0.24 | 35.64 / 0.77 | 35.33 / 0.77 | 35.53 / 0.77 | 35.77 / 0.78 | 35.79 / 0.78 | 35.81 / 0.78 | 36.55 / 0.79 |