Online image reconstruction via Multiple Orthogonal Reference Sensitivity Encoding (MORSE).

MORSE: A New Efficient Method for Real-Time MRI Image Reconstruction

MRI scans often use "shortcuts" to speed up imaging. However, these shortcuts can cause blurry or distorted images. A new method called MORSE helps fix these distortions much faster than current tools. This makes it possible to clean up images instantly while the patient is still in the scanner.

Solving the Speed-Accuracy Trade-off in Parallel Imaging

The fundamental challenge in modern Magnetic Resonance Imaging (MRI) is the tension between temporal resolution and spatial resolution. To make scans faster, physicists employ parallel imaging. This technique samples data below the Nyquist frequency (the minimum sampling rate required to prevent signal distortion). This essentially takes fewer "snapshots" of the signal to save time.

However, this under-sampling creates aliasing. This is a phenomenon where signals from different parts of the body overlap. It creates "foldover" artifacts that look like ghostly shadows superimposed on the anatomy. To untangle these overlapping signals, the scanner must know the "sensitivity profile" of its receiver coils. These are the unique ways each physical antenna in the coil array perceives signal strength at different points in space.

Sophisticated algorithms exist to calculate these profiles and "unfold" the images. However, they are often so computationally heavy that they cannot be run in real-time. This forces a compromise. Clinicians must choose between fast, noisy, artifact-ridden scans or slow, high-quality reconstructions that happen long after the patient has left the room.

The Mechanics of Coil Sensitivity

To understand how MORSE addresses this, one must first understand the SENSE (Sensitivity Encoding) framework. In a typical parallel imaging setup, multiple receiver coils are placed around the patient. Because these coils are physically positioned differently, they each "see" the internal magnetization with a different weighting.

If we know exactly how much each coil weights a specific voxel (a three-dimensional volume element), we can solve for the true underlying magnetization. The accuracy of this solution depends entirely on the precision of the coil sensitivity maps.

Traditionally, these maps are estimated using a separate, fully sampled "reference" scan. However, at ultra-high fields (UHF), such as 7 Tesla (7T) scanners, several complications arise. Rapidly varying sensitivities and chemical shift artifacts occur frequently. Chemical shifts happen when signals from fat and water appear spatially displaced. Field inhomogeneities also make it difficult for a single, simple sensitivity map to describe a voxel accurately. If the map is too simple, the reconstruction fails. This leaves behind the very artifacts the technology was meant to eliminate.

Decoding the MORSE Framework

The authors propose Multiple Orthogonal Reference Sensitivity Encoding (MORSE) to move beyond these limitations. Instead of assuming one sensitivity profile per voxel, MORSE estimates multiple orthogonal sensitivities. This allows the algorithm to account for complex scenarios. For example, it can handle a single voxel containing both water and fat signals that appear to originate from different locations.

The MORSE pipeline, illustrated in, begins by transforming the reference data into a "virtual coil" space. Rather than performing massive calculations on every individual physical coil, the method uses Singular Value Decomposition (SVD). SVD is a mathematical technique that identifies the most important patterns in a dataset. This compresses the coil data into a smaller number of principal modes. This reduction in dimensionality is the engine of the method's efficiency.

Once in this reduced space, MORSE applies a flexible spatial weighting using a Gaussian kernel. Unlike previous methods that treated all voxels equally, MORSE can prioritize local information. This makes it robust to patient motion between the reference scan and the actual imaging sequence. The core innovation lies in the voxel-wise SVD performed on a weighted outer product of the coil signals. This step extracts not just the primary sensitivity, but also "higher-order" sensitivities. As shown in [, E], these higher-order vectors capture the nuances of the field that a standard single-map approach would miss.

Finally, these estimates are fed into a regularized SENSE framework. The authors utilize Tikhonov regularization (a method that stabilizes mathematical solutions by penalizing extreme values). This prevents the reconstruction from amplifying noise. Crucially, the singular values from the SVD are used to derive data-driven regularization terms. This means the algorithm automatically adjusts its "smoothing" based on its confidence in the sensitivity estimates at any given location.

High-Speed, High-Fidelity Imaging

The implications of MORSE are most visible when comparing it to existing methods like ESPIRiT, ENLIVE, or LORAKS. In testing on 7T functional MRI (fMRI) data, the authors find that MORSE produces a significantly higher temporal signal-to-noise ratio (tSNR) than its competitors. As demonstrated in, while GRAPPA and ESPIRiT suffer from signal instability or residual foldover artifacts, MORSE maintains a highly homogeneous and clean signal.

The disparity in reconstruction speed is perhaps the most critical finding. In one benchmark using a 7T high-resolution protocol, the ESPIRiT method required over 15 minutes to reconstruct a single volume. In contrast, MORSE completed the same task in just 24 seconds. For high-resolution structural imaging, where methods like ENLIVE can take several days to process, MORSE completes the job in under six minutes. This brings the possibility of "online" deployment to the forefront. This refers to the ability to see high-quality, corrected images almost immediately after the scanner finishes its sequence.

Beyond the brain, the authors demonstrate the method's versatility in liver and knee imaging. In these contexts, MORSE successfully handles rapid sensitivity changes at the edges of the anatomy. It also manages the complex chemical shifts of fat and water. This provides artifact-free images that are ready for clinical review without lengthy post-processing delays.

Limits of the Approach

Despite its advantages, MORSE is not a universal panacea. The method relies on several free parameters. These include the number of sensitivity orders ($N_{order}$), the smoothing kernel width ($w$), and the regularization factor ($\lambda$). The authors note that the optimal values for these parameters are spatially specific and application-dependent. A setting that works perfectly for a brain scan might fail for a liver scan.

Furthermore, the method faces a ceiling of computational complexity. While much faster than older methods, the speed is still governed by the number of reference coils and the chosen sensitivity order. Finally, users must be wary of "over-regularization." If the $\lambda$ parameter is set too high, the algorithm may introduce signal biases. It may also fail to correct certain foldover artifacts. In these cases, the system trades accuracy for an overly smooth, potentially misleading image.

Figures from the paper