Computer Vision News - August 2022

29 VORTEX: Physics-Driven Data... The reconstruction process is highly sensitive, and noise andmotion corruption are two of the perturbations most prevalent in clinical practice . Alongside a lack of fully sampled scans, these are the fundamental limitations of deploying deep learning-based methods in clinical practice and are the motivation for this work. Prior to the deep learning era, people attempted to develop solutions and algorithms tailored to a small subset of perturbations.Arjun, Beliz, andtheteamare trying to create a generalized framework robust to all these perturbations . “ Generally, with iterative methods, they’re sample-specific, but with machine learning, if you play your cards well, you can generalize to unseen things, ” Beliz points out. “ That’s very powerful. Here, we can teach our models to be prepared for what’s tocomeand leverageour physics knowledge about MRI acquisition. In machine learning methods, you usually give large amounts of data to the model and expect it to learn and MRI is a powerful non-ionizing imaging technology commonly used in clinical practice, but its downside is that it can take some time to acquire . Accelerating the scans requires dropping data points, which results in poor quality or degraded images. When data is dropped, it is referred to as being undersampled . Accelerated MRI reconstruction aims to recover high-quality images from this undersampled raw data. “ Most reconstruction methods are fully supervised , which requires many fully sampled scans to train the deep learning models, ” Arjunexplains. “ In clinical practice, we typically acquire undersampled scans. Also, deep learning and traditional iterative methods for reconstruction are sensitive to perturbations in the acquisition process. MRI, like other forms of imaging, is rooted in the physics of the hardware , and there are certain perturbations in the acquisition. We may experience noise in the image, or if a patient moves, that will cause some artifacts in the data. ”