Computer Vision News - May 2022

47 Vision Transformers in Medical Computer Vision There are different ways to approach the transformations: • Generative vision transformer based unsupervised MRI reconstruction architecture to increase the receptive field. • Generative non-linear mapping over latent and noisy space to improve invertibility of the model. • Cross attention to improve context of image features. Extensive experiments on accelerated multi-contrast brain MRI dataset. They proposed an ASFT network to reconstruct the high-resolution MRI scans from low resolution scans. They introduced a multi-branch features transformation and extraction (MFTE) block. They filtered out the useless information using MFTE block. Their model achieved the state-of-the-art performance for super-resolution task. Medical Image Synthesis: Tissuemorphology information acquired frommultimodal medical images play an important role in the clinical practice. GAN is a CNN based architecture that shows locality bias and spatial invariance across all the positions. Double-scale GAN showed efficient performance on benchmark IXI MRI dataset. The authors propose a dual transformer network (DTN) model for the diffeomorphic registration of MR images. DTN uses self-attention mechanisms to facilitate contextual correspondence between anatomies. DTN has two branches to learn the relevance based on the embeddings of separate one-channel images and concatenated two-channel images. DTN has two branches to learn the relevance based on the embeddings of separate one- channel images and concatenated two-channel images. DTN uses feature enhancement, based on global correspondence, to infer the velocity field and registration filed. DTN is used to optimize metric space. DTN is unsupervised. Vision Transformers (ViT) are now one of the hottest topics in the discipline of computer vision. Although CNNs are matured enough for the development of applications that can ensure an efficient and accurate diagnosis, in the medical field - where an inaccurate output might endanger lives - the concept of attention in vision transformers has paved its way for more precise outcomes. A variety of approaches have been proposed in recent years to explore and utilize the competency of vision transformers. These approaches showed excellent performance on a wide range of visual recognition tasks, including classification, lesions detection, anatomical structure segmentation, and clinical report generation. Nevertheless, the real potential of transformers for computer vision has yet to be fully explored. Next month Let’s meet next month with the continuation of the previous month’s coding article. I hope that you tried some of the ideas presented, as now we’ll see a more hands on approach. Until next month, have a great time and always be curious! 