Table 2. Quantitative evaluation of DL-TNode-3D and cGAN-2D for 3 OCT systems. DL-TNode-3D is our method, which uses the volumetric information for speckle suppression whereas cGAN-2D uses only 2D (depth-fast axes) information for speckle suppression. CNR values that are nearest to ground truth volume CNR are highlighted in bold. The highest values of PSNR, SSIM, and MS-SSIM are highlighted in bold. CNR: contrast-to-noise ratio, PSNR: peak-signal-to-noise ratio; SSIM: structural similarity index; MS-SSIM: multi-scale structural similarity index.
| OCT System | Trained Model | CNR | PSNR (dB) | SSIM | MS-SSIM |
|---|---|---|---|---|---|
| Ophthalmic | Ground truth | 1.193 | - | - | - |
| cGAN-2D | 1.188 | 34.726 | 0.943 | 0.976 | |
| DL-TNode-3D | 1.302 | 38.076 | 0.988 | 0.996 | |
|
| |||||
| VCSEL | Ground truth | 1.623 | - | - | - |
| cGAN-2D | 1.612 | 37.175 | 0.954 | 0.982 | |
| DL-TNode-3D | 1.675 | 41.095 | 0.978 | 0.994 | |
|
| |||||
| Polygon | Ground truth | 1.394 | - | - | - |
| cGAN-2D | 1.525 | 36.876 | 0.949 | 0.985 | |
| DL-TNode-3D | 1.505 | 40.654 | 0.988 | 0.996 | |