Hierarchical Probabilistic Ultrasound Image Inpainting via Variational Inference

Med-cDiff: Conditional Medical Image Generation with Diffusion Models

“…Vahdat and Kautz [37] developed Nouveau VAE (NVAE), a hierarchical VAE that is able to generate highly realistic images. Hung et al [2] adapted some of the features from NVAE into their hierarchical conditional VAE for ultrasound image inpainting. Cui et al [38] adopted NVAE in positron emission tomography (PET) scan image denoising and uncertainty estimation.…”

“…We compared our method against other inpainting methods such as pix2pixGAN, HPUNet [2], and UP-GAN using the LPIPS and FID metrics. Furthermore, we performed a 2AFC paradigm [65] to measure how well trainees can discriminate real images from the generated ones.…”

“…Existing generative models are able to perform some of these jobs decently; e.g., the Hierarchical Probabilistic UNet (HPUNet) [2] for ultrasound image inpainting, and the progressive Generative Adversarial Network (GAN) [3] and SMORE [4] for medical image super-resolution. These methods work to some extent, but they are tailored to specific applications or imaging modalities, making it difficult for researchers to adapt them to different tasks or modalities.…”

“…Models based on Variational Autoencoders (VAE) can be effective in some medical applications [2,7], but the generated images tend to be blurry [8]. Although GAN-based models can generate high-quality medical images [5,9], they suffer from unstable training due to vanishing gradient, convergence, and mode collapse [10].…”