Non-cartesian SENSE reconstruction of 3D UTE Cones for fast MR lung imaging
Konstantinos Zeimpekis1,2, Klaas Pruessmann2, Florian Wiesinger3, Patrick Veit-Haibach1, and Gaspar Delso4

1Nuclear Medicine, University Hospital Zurich, Zurich, Switzerland, 2Information Technology and Electrical Engineering, ETHZ, Zurich, Switzerland, 3GE Global Research, Munich, Germany, 4GE Healthcare, Waukesha, WI, United States

Synopsis

This study is about a first attempt to use CG-SENSE parallel reconstruction for non-cartesian 3D Ultra-short Echo Time Cones sequence for lung imaging. Primary goal is to test under-sampled data that reduce the scan time effectively to one quarter of the fully sampled acquisition and check if the reconstruction manages to capture lung density signal to be used for accurate PET Attenuation Correction on a PET/MRI since conventional sequences that are currently used do not capture any. We test also the possibility for high resolution lung imaging from the undersampled data reconstructed with CG-SENSE algorithm.

PURPOSE

Ultra short echo time (UTE) can provide images with lung parenchyma visualization that can be used for MR attenuation correction mapping for PET/MR [1]. Still scan times can be long for both high resolution diagnostic imaging and attenuation map extraction. We developed a conjugate-gradient SENSE non-cartesian parallel reconstruction [2] for 3D UTE Cones [3] to effectively reduce clinical scan times and preserve image quality.

METHODS

Data from healthy volunteer subjects were acquired on a GE Discovery MR750w 3T scanner. A fully-sampled and a 4x-undersampled scans were acquired. Acquisition time for the fully-sampled scan was 3 minutes and 19 seconds and for the undersampled scan was 49 seconds with a 3D UTE Cones sequence: slab selective minimum phase SLR pulse, TR/TE/FA 4ms/0.028ms/7o, FOV=30*30 cm2, 1.4 mm spatial resolution and 140 slices, BW 250 kHz, 500 points along each interleave. For both scans prospective gating was applied with 500 interleaves per trigger. The parallel reconstruction algorithm was based on CG-SENSE method and Kaiser-Bessel gridding with oversampling ratio of 1.2 and kernel width of 5. Coil sensitivity maps were acquired from the original data set through eigen-decomposition method [4]. CUDA MEX functions ran on MATLAB 8.1 on a computer with 4 GPU cores. Number of iterations - with maximum peak SNR and minimum mean squared error compared to the fully-sampled data – that yielded best results was three. Total reconstruction time around 5 minutes.

RESULTS

Figure1 shows original data set, 4x-undersampled and CG-SENSE with 3 iterations parallel reconstruction axial slices (upper images) and Maximum Intensity Projections (lower images). Figures 2 and 3 respectively show sagittal and coronal views. For quantitating the lung density signal, ROIs were drawn on the anterior and posterior of either lung and on the background as it can be shown in Figure 4. Then the lung to background signal ratio were calculated for each lung ROI with the corresponding neighboring background ROI. Figure 5 show the signal ratios for all anatomical structures (LR-left right, LA-left anterior, RP-right posterior, RA-right anterior) for all reconstructions (original – 4xundersampled – CGSENSE).

DISCUSSION

Figures 1,2 and 3 show significant SNR recovery with CG-SENSE from the 4x-undersampled images. for all tomographic planes. As expected the CG-SENSE cannot match the original SNR but yields good image quality for the specified scan time which is 49 seconds. The CG-SENSE has suppressed noise and image degradation due to under-sampling artefacts. Figure 5 shows the recovered lung density signal of the CG-SENSE compared to the under-sampled data set. Interestingly enough there is higher lung signal towards the posterior part (left and right) due to the supine positioning of the patient (lung is compressed due to gravity towards the bed).

CONCLUSION

This is a first try to implement 3D non cartesian parallel reconstruction with CGSENSE for lung imaging. From these first results we propose that CGSENSE can be used to obtain MR lung attenuation maps for PET/MR with comparable scan time to the current protocol being used - LAVA (18 seconds) which however does not capture any lung parenchyma signal. Further work needs to be done for using CG-SENSE for high resolution lung imaging perhaps with better sensitivity maps that can be extracted with a short separate scan, better coil coverage and optimisation of the sequence itself. These images need to be clinically assessed to validate the results.

Acknowledgements

No acknowledgement found.

References

[1] Zeimpekis KG, Delso G, Wiesinger F, et al. Investigation of 3D UTE MRI for lung PET attenuation correction. JNM. 2014; 55 (Supplement 1):2103

[2] Pruessmann KP, Weiger M, Börnert P, et al. Advances in sensitivity encoding with arbitrary k-space trajectories. MRM. 2001;46(4):638-51

[3] Gurney PT, Hargreaves BA, Nishimura DG. Design and analysis of a practical 3D Cones trajectory. MRM. 2006;55(3):575-582

[4] Uecker M, Lai P, Murphy MJ. ESPIRiT--an eigenvalue approach to autocalibrating parallel MRI: where SENSE meets GRAPPA. MRM. 2014;71(3):990-1001

Figures

Upper images: single axial slice for original data, 4x-undersampled and CG-SENSE reconstructed with 3 iterations. Lower images : Maximum Intensity Projection (15 axial slices) for original data, 4x-undersampled and CG-SENSE reconstructed with 3 iterations.

Upper images: single sagittal slice for original data, 4x-undersampled and CG-SENSE reconstructed with 3 iterations. Lower images : Maximum Intensity Projection (12 sagittal slices) for original data, 4x-undersampled and CG-SENSE reconstructed with 3 iterations.

Upper images: single coronal slice for original data, 4x-undersampled and CG-SENSE reconstructed with 3 iterations. Lower images : Maximum Intensity Projection (15 coronal slices) for original data, 4x-undersampled and CG-SENSE reconstructed with 3 iterations.

Original data set, 4x-undersampled and CG-SENSE with 3 iteration axial slice with ROIs drawn on the left anterior (lung and background), right anterior (lung and background), left posterior (lung and background) and right posterior (lung and background).

Lung to background ratio for all anatomical structures (left anterior-LA, right anterior-RA, left posterior-LP and right posterior-RP) for all reconstructions (red curve-original fully sampled data set, blue curve-4x-undersampled and green curve-CGSENSE with 3 iterations).



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
1608