Compressed Sensing

MRI is a fascinating imaging technology, but acquisition speed still remains a challenge, especially for patients who are anxious, can't keep still or those who have limited press hold capacity. These challenges can be solved with compressed sensing. The technology that helps us to go beyond speed by drastically decreasing acquisition times in MRI without sacrificing image quality. My name is crystal form and I'm part of the research and development team at Siemens Healthineers that developed the first clinical application based on compressed sensing and I'm happy to explain the technology to you. Let's simplify the sophisticated technique by comparing it to everyday things to understand the essence of compressed sensing. Just imagine taking an MRI scan. Of an Apple, traditionally we would sample all raw data points of this Apple in the so called case space. This guarantees the highest image quality you can think of. But this also makes MRI scans long compress sensing changes the game based on 3 Golden rules, incoherent subsampling, transform sparsity, and nonlinear iterative reconstruction. Let's focus on number one, incoherent subsampling. We will look at both words separately and focus on subsampling. First subsampling what we all know from existing acceleration techniques means that we sample just apart of all raw data points. However, subsampling typically is accompanied by a degradation of image quality, for example, either low resolution image or an image within folding artifacts. At this point in coherence the new part of this technique comes into the game. Incoherent means that the data points are sampled randomly, so both combined. Enable two essential things. First, subsampling increases speed. The fewer raw data points we sample, the faster the acquisition and 2nd with incoherence we avoid distinct aliasing artifacts. Instead, incoherent subsampling leads to noise like artifact superimposed over the image. Let's find out how we can remove this noise from the image. Rule #2 is transformed sparsity first question what is sparsity and image is considered sparse when it's informational content is reflected in as few data points as possible. The best example of a sparse Mr image is from angiography with white vessels surrounded by black background. Although the background is often not exactly black, know important information is lost by setting the background data to zero. What means to completely black. So what does transform sparsity mean? Let's go back to the example of our noisy Apple, which we acquired with incoherent subsampling. If you have a closer look and select one line of the image, the image intensity in this line clearly shows that the actual relevant information is overlaid with noise, which we ideally would like to separate from the valuable information. Therefore, we transform the image into another representation where it's easier to distinguish between useless and useful. Information. For example, this can be performed with the so-called wavelet transform. If you now cut the same line, it's very easy to define a threshold and say all pixels below this threshold contain information that is irrelevant. And what happens by removing all pixels below this threshold? Much of the noise has disappeared and the image has been clearly improved when we transform it back to its original representation. Of course it is not that easy to separate the noise from the valuable information in case of a real Mr Image. That's why we need the third ingredient of compressed sensing. The nonlinear iterative reconstruction. Our goal is to achieve an optimum balance of data consistency and sparsity. That means we want to remove as much noise as possible, but not too remove useful image information. Let's look what this would mean for our known example. If we overweight transform sparsity, we would get a completely black image as we would have removed all the image information and as you can see there's nothing to see. Let's look at it the other way round. If we overweight data consistency, we filter out too little noise, or in the extreme case, nothing at all. This means data consistency will be at 100%, but there's no improvement in image quality. Actually, just the image we started with a noisy Apple. The iterative process ensures a balance of both and after a defined number of iterations, data consistency and sparsity have been increased and have come to an ideal balance. That means we've brought speed and quality in harmony thanks to the three Golden rules of compressed sensing. Incoherent subsampling for high acquisition speed transforms sparsity to separate and remove the noise from the image content and nonlinear iterative reconstruction to balance data consistency and sparsity. Let's look at the potential that compress sensing has. Take discard exam. This compress sensing cardiac scene in our first clinical application based on compressed sensing, we enable high resolution cardiac sonimage ingin free breathing. This opens up cardiac MRI to a larger patient group such as those with arrhythmia. All those who cannot hold their breath. As you can see there is great potential in this disruptive technology. It opens up the way to go beyond speed and joy discovering it.

Compressed Sensing subsampling Speed and quality in harmony "Relevant" information Easier separation Compressed Sensing 3. Non-linear iterative increases distinct of noise and overlaid with noise 2. Transform Transform 2. Remove noisefrom aliasing speed 2. Transform 2. Transform image content 2. Transform Acquisition speed sparsity image content sparsity Transform 2. Transform is a challenge s a challenge sparsity artifacts Compressed Sensing sparsity Transform sparsity sparsity Transform Data low 1. Incoherent low 1. Incoherent 1. Incoherent 1. High acquisition cs consistency subsampling sparsity 2. Transform Data net subsampling subsampling speed cs improvement Transform Beyond speed. reconstruction Incoherent subsampling consistency sparsity Transform fascinating for those who have Data sparsity consistency subsampling subsampling 3. Non-linear iterative 3. Balance data consistency for patients who Compressed Sensing Cardiac Cine* center only 50 % of lines reconstruction Christoph Forman and sparsity sparsity 3. Non-linear iterative are anxious anxious High-resolution in free-breathing reconstruction Research & Development at 3. Non-linear iterative non-linear reconstruction reconstruction high 0.8 0.6 0.2 high Siemens Healthineers 100 150 250