-
Dataset: TCIA Lymph Node Abdomen dataset, including 88 multidetector CT.
-
Model: Conditional GAN with a modified UNet generator and a CNN discriminator. Below is the overall framework:
-
Loss function: Pixel loss + perceptual loss + adversarial loss.
-
Evaluation: Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Hausdorff Distance (HD). Compared with conventional total variation (TV) denoising.
Image quality scores of motion-affected, TV-denoised, and GAN-generated images. Motion-free images are used as reference. The best qualities are bolded:
| Metrics | MSE | PSNR | SSIM | HD |
|---|---|---|---|---|
| Motion-affected | 19.9639 × 10⁻³ ± 0.0440 | 31.4970 ± 7.9058 | 0.8520 ± 0.1963 | 5.8740 ± 5.7612 |
| TV-denoised | 19.9267 × 10⁻³ ± 0.0439 | 31.2140 ± 9.2304 | 0.8498 ± 0.1945 | 6.0619 ± 5.7854 |
| GAN | 4.8628 × 10⁻³ ± 0.0135 | 33.2646 ± 5.5166 | 0.9011 ± 0.1189 | 5.8509 ± 5.3868 |
Group difference between the image quality of GAN-generated, TV-denoised, and motion-affected images:
Sample visualization:
