Technologies for Task-Specific Objective Optimization of Magnetic Resonance Imaging Acquisition and Reconstruction

Advances in technology have led to dramatic improvements in the quality of magnetic resonance imaging (MRI) data. If we had infinite time, making a high-quality MR image is easy. just collect lots of data! But every decision made in designing MRI techniques is a trade-off, given that we tailor every MR study to address specific tasks. So what constitutes a "high quality" MR image? The answer depends on two things: 1) What is the image going to be used for, and 2) what metric is used to measure "quality"? When physicists develop improved MR imaging protocols, their optimization process would ideally involve one or more ways to characterize quality objectively and quantitatively. This is true when we develop new "sequences" (i.e., ways of using MRI equipment to obtain images), or reconstruction and analysis tools (e.g., new Artificial Intelligence models to improve the images or extract information from them). There are, in fact, many so-called Image Quality Metrics (IQMs) that have been created, however almost all of these were developed to help characterize digital photos of natural scenes - that is, they were made so that when a new JPEG algorithm is created to shrink the size of photos on our phones, we can objectively quantify how good the average human viewer would think that new image looked. What makes a good quality "selfie" is, however, very different from the features in an MR image that make it useful for extracting physiologic information, and my work has shown that these IQMs often have very little to do with the usefulness of an MR image. Additionally, work from my lab has shown that the exact same general quality of an MR image can subsequently be rated as having a very different quality when rated by experts using it to do specific tasks. The image may be of high quality to answer one question, but of very low quality to answer another. This is because features in that image, required to answer the question, are impacted differently depending on the type and degree of image quality degradation. In my research program, I seek to marry the development of new MRI technologies, developed to achieve specific tasks - such as more accurately measuring the pharmacokinetics of leakage through the blood-brain-barrier - with novel objective quality metrics validated for that task. Arguably, an MR image is only actually a better image if it can change or improve on the outcome of the task that we were seeking to use it for. An MR image that is sharper might objectively be better when characterized by the metric of spatial resolution, but if we could answer the question without that resolution then it is not actually a better MRI technology (at least for that task!). Success in my research program will therefore improve MRI physicists’ and engineers’ ability to create, test and validate new MRI sequences, hardware, and image reconstruction models (and seek to commercialize those innovations through industry partnership).