One of the main limitations of MRI is the trade-off between spatial and temporal resolution. Sampling more k-space is needed for higher spatial resolution which in turns requires time. This limitation affects a variety of applications from static to dynamic MRI. For e.g, a 3D MP-RAGE or 3D fast spin echo acquisition can take 5-7 min. depending on the spatial resolution desired which can be subject to motion artifacts in pediatric or anxious patients. The limitation is even more obvious in dynamic MRI applications like post contrast liver imaging for example. Multiple temporal phases are acquired after injection of a contrast agent but each phase to be acquired within a short duration to capture angiographic, arterial and venous phases. Another example is functional MRI where both temporal resolution to track fast changes and spatial resolution to observe changes at a finer scale are important. Often, compromises are made to make one better at the cost of the other.

Various approaches have been proposed to address this limitation. The earliest is the use of non-Cartesian k-space trajectories like echo-planar or spiral readouts to traverse k-space more efficiently than Cartesian line-by-line traversal. The next major advance in the field to address this limitation was parallel imaging. Parallel imaging utilizes redundant information from multiple coils to enable undersampling the k-space data and fill in the missing lines using coil sensitivity information. This can be done in image space, k-space or hybrid-space. Most clinical protocols employ parallel imaging in at least some applications to achieve 2-3X reduction, which can be used to improve scan time or temporal resolution or to improve spatial resolution or both. Other approaches involving view sharing where the central k-space data is acquired and the peripheral k-space data is shared between different temporal phases have been proposed for dynamic MRI especially MR angiography. More sophisticated approaches like HYPR that involve constrained reconstruction have also been successfully demonstrated.

Compressed sensing (CS) exploits the inherent sparsity in MR data to reconstruct images with minimal artifacts from specially designed k-space sampling schemes. It overturns the traditional model of acquiring large amounts of (redundant) data, reconstructing them and then compressing them and instead acquires data sparsely and uses specialized image reconstruction algorithms. In contrast to other acceleration schemes, CS acquires the data using a sampling scheme that ensures incoherent point-spread-functions (PSF). This is often achieved using simple pseudo-random sampling or more sophisticated Poisson disk sampling schemes. An incoherent PSF already is beneficial in making the under-sampling artifacts non-coherent. CS reconstruction is non-linear and ensures that of the many solutions possible to fill the missing k-space, the one that yields the most sparse image is selected. Note that the sparsity imposed could be spatial or temporal. In dynamic MRI like post-contrast liver imaging or angiography, there is sparsity in the temporal domain since only a small fraction of voxels change from one temporal phase to the next. This can be exploited to achieve speed-ups. Lastly, CS can be combined with parallel imaging to make it even more efficient.

In this talk, we will briefly cover traditional methods for improving spatio-temporal resolution but largely focus on the basics of CS reconstruction and the more advanced CS methods that combine it with parallel imaging, concluding with real world clinical examples.