Table 3.
Quantitative image generation comparisons.
| Dataset | Model | FID ↓ | KID ↓ | Precision ↑ | Recall ↑ |
|---|---|---|---|---|---|
| AIROGS | StyleGAN-3 | 20.43 | 0.019 | 0.43 | 0.19 |
| Medfusion | 11.63 | 0.008 | 0.70 | 0.40 | |
| CRCDX | cGAN | 49.26 | 0.036 | 0.64 | 0.02 |
| StyleGAN-3 | 18.83 | 0.014 | 0.57 | 0.24 | |
| Medfusion | 30.03 | 0.021 | 0.66 | 0.41 | |
| CheXpert | ProGAN | 84.31 | 0.127 | 0.30 | 0.17 |
| StyleGAN-3 | 28.69 | 0.032 | 0.68 | 0.08 | |
| Medfusion | 17.28 | 0.020 | 0.68 | 0.32 |
Models include Generative Adversarial Networks (StyleGAN-3, cGAN, and ProGAN) and our proposed Medfusion model. Metrics for the best-performing model are indicated in bold.
FID fréchet inception distance, KID kernel inception distance.