Computer Vision News - April 2023

28 Medical Imaging Tools Welcome to a new article fully dedicated to tools for medical imaging! Today we will be talking about the new extension of MONAI , an extraordinary platform that we already reviewedmore than 2 years ago and that is continuing to grow, thanks to a shared and collaborative effort and to a research interest lead by facilitating use of deep learning methods in medical practice. The new extension, available here , is a prototyping repository for generative models, and since the writer is a big fan of those, I jumped at the opportunity to try it out and talk about it here. In a previous article about MONAI, our tutorial was using GANs , but the Generative Models extension introduces new state-of-the-art architectures such as Diffusion Model, Autoencoder-KL, VQ-VAE, Autoregressive transformers, (Multi-scale) Patch- GAN discriminator. Moreover, it offers a wide choice of losses: Adversarial, Spectral, and Perceptual losses, and different Diffusion Model Noise Schedulers: DDPM, DDIM, and PNDM. You can use it for applications such as Anomaly Detection, Inpainting and Super-resolution challenges. Before writing this article, we spoke to Walter Pinaya, one of the main contributors of this distribution, and a great colleague at KCL. He highlighted that the strength of MONAI Generative Models lies in the implementation and support for diffusion models in 3D medical imaging, a special feature from the team. Inspired from him, we chose to show 3 applications, two of which use the denoising diffusion probabilistic model to generate 3D synthetic images from the Brain tumour Decathlon dataset and for inpainting on the 2D MedNIST dataset . The other application uses VQVAE for 3D reconstruction . Enjoy! GENERATIVE MODELS By Marica Muffoletto (twitter) Walter Pinaya