MedVAE

Variational Autoencoders (VAEs) are widely used in imaging tasks such as image generation, reconstruction, and representation learning. However, most open-source VAEs are trained on natural images and are not suitable for 3D medical CT data. This creates a gap for researchers who need reliable VAE models that can handle real clinical CT volumes.

To address this problem, we benchmark MedVAE, a VAE pre-trained on large-scale 3D medical CT data. The goal is to evaluate whether MedVAE can better preserve anatomical structures and intensity distributions compared to general-purpose VAEs.

Benchmarks

Name	VAE Type	# Patients	LPIPS	SSIM	PSNR	DSC Abdominal
stable-diffusion-v1-5	KL-VAE	0
stable-diffusion-3.5-large	KL-VAE	0
2D-MedVAE_KL-smooth-tianyu-2K	KL-VAE	2k
2D-MedVAE_KL-smooth-kuma-10K	KL-VAE	10k
2D-MedVAE_KL-sharp-kuma-10K	KL-VAE	10k
2D-MedVAE_KL-sharp-kuma-20K	KL-VAE	20k
2D-MedVAE_KL-sharp-kuma-100K	KL-VAE	100k
2D-MedVAE_KL-sharp-kuma-300K	KL-VAE	300k
3D-MedVAE_MAISI	KL-VAE	40K
3D-MedVAE_KL-sharp-kuma-100K (ongoing)	KL-VAE	100K
3D-MedVAE_KL-sharp-kuma-1M (ongoing)	KL-VAE	1M
3D-MedVAE_VQ	VQ-VAE	100K
3D-MedVAE_VQ-kuma-100K (ongoing)	VQ-VAE	100K

Citation

@article{liu2025see,
  title={See More, Change Less: Anatomy-Aware Diffusion for Contrast Enhancement},
  author={Liu, Junqi and Wu, Zejun and Bassi, Pedro RAS and Zhou, Xinze and Li, Wenxuan and Hamamci, Ibrahim E and Er, Sezgin and Lin, Tianyu and Luo, Yi and Płotka, Szymon and others},
  journal={arXiv preprint arXiv:https://www.arxiv.org/abs/2512.07251},
  year={2025},
  url={https://github.com/MrGiovanni/SMILE}
}