Dr. Hoge is actively engaged in pursuing algorithmic and computational methods to improve image quality in MRI. His focus has been on developing better methods to reconstruct accelerated MRI data, with a particular emphasis on methods to reduce inherent artifacts in functional and perfusion imaging.
Index of Current and Past Projects:
Echo-Planar Imaging Artifact Correction
Echo-Planar Imaging (EPI) is a rapid MRI data acquisition method. This high acquisition speed requires rapid magnetic field gradient switching to sample the necessary image data. However, gradient hardware errors and the temporal evolution of the spins that generate the measured signal can result in reconstructed images that look significantly different than expected.
One result of magnetic field gradient hardware errors is a mismatch between the desired and actual EPI sampling locations. This mismatch appears in the reconstructed images as a Nyquist ghost, that is shifted 1/2 a field-of-view away from the intended image location. Identifying and correcting this data mismatch has been the focus of research efforts from many different groups since the early 1990's.
Ghost Elimination via Spatial and Temporal Encoding (GESTE)
Gradient hardware artifacts in EPI depend on the polarity of the readout gradient, and tend to be equal-and-opposite when this polarity is reversed. If one acquired two EPI data sets, one with the first readout using a positive polarity (RO+) readout gradient and the second with the first readout using a negative polarity (RO-) readout, then the N/2 ghost artifacts will have opposite phase. Adding these two images together in a phase coherent fashion will cancel the ghosts (see the PLACE paper by Xiang and Ye, MRM 2007, 57:731-741). In 2010, we extended this approach using parallel imaging methods to both maintain the original EPI temporal resolution and reduce the amount of signal lost by the phase cancellation effect (Link).
Dual-Polarity GRAPPA (DPG)
In GESTE, the pMRI reconstruction step is applied to the RO+ and RO- data separately. This can be disadvantageous as it doubles the effective acceleration rate of the EPI data in the pMRI reconstruction stage, which results in a much lower pMRI SNR limit compared to standard EPI reconstruction methods. However, if one splits the pMRI convolution kernel across both the RO+ and RO- data, this problem is effectively resolved (Link). In addition, this approach also has the advantage that non-linear phase errors between the RO+ and RO- data can be automatically corrected by the pMRI reconstruction. The end result is EPI images that have significantly lower levels of Nyquist ghost artifacts, particularly at high field (>=3T) or in imaging scenarios where typical EPI reconstructions result in unsuitable images. As an example, the image below shows images acquired using a zoomed-EPI sequence in the sagital plane, using an acceleration rate of R=2 and a 32-channel reciever on a 3T magnet. Images from a traditional EPI reconstruction are shown in the top row, and exhibit a high level of Nyquist ghost artifacts throughout. The DPG reconstruction is shown on the bottom row, where the ghosts have been correctly eliminated.
Software implementations of these Nyquist ghost correction methods are available through the NC-IGT Fast Imaging Library (see below).
In regions of magnetic field in-homogeniety, the excited spins that generate the sampled signal can become out-of-phase in the time it takes to acquire EPI data. This results is signal drop-out, and is particularly problematic in regions of the brain above the paranasal sinuses and the ear canals. Z-shim methods apply a small z-gradient prior to the EPI readout, to provide spin-dephasing compensation and recover signal that would otherwise be lost. This enables functional assessment of deep brain structures that are not visible using standard EPI BOLD techniques. An example is shown in the image below.
Double-shot z-shim methods acquire two images, each with different amounts of z-shim compensation, in an alternating fashion.
In 2011, we presented a method that provides z-shim compensation in a single EPI readout (Link). The advantages include better temporal resolution and better resilience to motion, at the cost of slightly longer TR and echo spacing times.
Much of our work in the reconstruction of accelerated MRI data focused on modelling the signal acquisition system, identifying the relationship between different reconstruction approaches, and applying advanced numerical methods to the reconstruction problem.
A good review of the work highlighted below was presented in July 2010 at the SFB Workshop in Graz, Austria (Link).
GRAPPA vs. SENSE: similarities and differences
In 2006, we presented papers that identified the similarities and differences between many pMRI reconstruction methods, including SENSE and GRAPPA (Link, Link). One significant difference is that GRAPPA is more reliable in reconstructing low-spatial frequency information. This can be exploited to improve auto-calibrated SENSE methods (Link), which was a precursor to modern methods such as ESPIRiT.
Combining LSQR-Hybrid regularization with pMRI
The SENSE reconstruction problem is optimal in the least-squares sense. Thus, many least-squares optimization methods can be applied to the pMRI reconstruction problem. In 2006, we presented a paper on the application of the LSQR-Hybrid method to the pMRI reconstruction problem (Link), and in 2007 extended this approach to include the reconstruction of accelerated partial-Fourier data (Link).
Some of the methods described above have been utilized in real-time imaging applications, including
Catheter and Needle Tracking
Many of the methods developed by Dr. Hoge are implemented in the NC-IGT Fast Imaging Library, a freely distributable collection of MRI reconstruction methods available at http://ncigt.org/imaging-toolkit-fast-imaging-library