Figure 3
From: Denoising diffusion probabilistic models for 3D medical image generation
Comparison of the reconstruction quality of the VQ-GAN autoencoder when using different compression factors for two different samples. A latent dimension of 64 × 64 × 8 (i.e., a compression factor of 4 in each dimension, (B)) allows for a detailed reconstruction of the original image (A) by largely upholding anatomic consistency. A compression factor of 8 (i.e., a latent dimension of 32 × 32 × 4, (C)) distorts the fibular bone structure in knee MRI studies (upper row of images) and the fibroglandular tissue structure in breast MRI studies (lower row image). Adjacent, please find zoomed close-up views of the anatomic region of interest.
