This study demonstrates the use of motion modelling and super resolution reconstruction (SRR) to produce an isotropic 3D image of the upper abdomen during free breathing.
Sagittal and coronal 6 mm 2D slices are acquired throughout the volume of interest. The slices are repeated with sub-voxel offsets to facilitate SRR. An interleaved navigator slice is also acquired.
The navigator slice is processed with non-rigid registration and principal component analysis, to give two motion surrogate signals. These signals are used to control the motion model. The motion model and the SRR are jointly optimised using an iterative scheme.
A volunteer was scanned, breathing freely, on a Philips Ingenia 3T system with dStream torso coil. A 2D single-shot half-Fourier turbo spin-echo (TSE) sequence was used with TE = 90 ms. Signal from fat was suppressed with spatial pre-saturation with inversion recovery (SPIR).
Contiguous stacks of sagittal and coronal 6 mm image slices were acquired through the abdomen and thorax. The stacks were repeated with a 2 mm and a 4 mm offset in the through-slice direction, to allow SRR. The in-plane acquired voxel sizes were 2.0 × 2.0 mm. Interleaved with every image slice was a sagittal navigator slice, positioned through the right hemi-diaphragm. The complete sagittal and coronal stacks were acquired five and four times, respectively, and the total acquisition time was 25 minutes.
The navigator slices were registered to a group average image using non-rigid registration1. Principal component analysis (PCA) was applied to a matrix of the deformation fields over time. The weights of the first two principal components were converted to standard scores and then interpolated with a cubic spline to the acquisition time of the image slices. The interpolated values were used as motion surrogate signals.
A method which combines respiratory motion modelling and image registration was used to fit a 3D motion model directly to the 2D image slices2. The motion model was used to perform a motion compensated SRR using the iterative back projection method3. The motion model and SRR were jointly optimised using an iterative scheme2. The voxel spacing in the motion-compensated image was an isotropic 1.875 mm.
Slices through the motion-corrected reconstructed 3D image are shown in Figure 2 and a non-motion compensated reconstruction is shown in Figure 3, for comparison. Figure 4 shows slices displayed at a higher magnification through the gall bladder and the biliary tree, with non-motion compensated images in Figure 5.
The reconstruction has effectively removed the slice stacking artefacts that would normally be seen when reconstructing a 3D volume from 2D images acquired during free breathing. The liver, gall bladder and bile ducts are displayed clearly in Figures 2 and 4, but blurred in Figure 3 and 5, indicating that the motion model has successfully compensated for the respiratory motion.
The results show that a clinically useful improvement in image quality and resolution can be obtained using the proposed method. However, further improvements are required before the method is suitable for routine clinical use.
In 2D TSE imaging, the signal from flowing blood varies, depending on the flow velocity and the direction of flow compared to the slice plane. This means that the signal from the blood is different for the sagittal and coronal images that are used for the reconstruction. This is likely to cause errors in the motion model. The inconsistent blood signal was confirmed by the radiologist in the team.
A high specific absorption rate (SAR) was experienced by the volunteer, due to the use of a TSE sequence. This was a significant limitation to the speed of the acquisition, adding 89% to the scan time. The 2D imaging sequence used was a T2-weighted half-Fourier turbo spin echo sequence, but any single-shot 2D sequence could have been used to achieve different image contrasts and, potentially, a lower SAR.
The use of an interleaved navigator slice with the same image quality as the image slices means that only half of the acquired data is used to generate the final image. The acquisition time could therefore be halved by obtaining motion surrogate signals that are independent of the imaging process or are derived from thermal noise in the receive coils4.
1. Modat M, Ridgway GR, Taylor ZA, Lehmann M, Barnes J, Hawkes DJ, Fox NC, and Ourselin S. Fast freeform deformation using graphics processing units. Comput Meth Prog Bio. 2010;98(3):278–284,.
2. McClelland JR, Modat M, Arridge S, Grimes H, D’Souza D, Thomas D, O’Connell D, Low DA, Kaza E, Collins DJ, Leach MO, and Hawkes DJ. A generalized framework unifying image registration and respiratory motion models and incorporating image reconstruction, for partial image data or full images. Phys Med Biol. 2017;62:4273–4292.
3. Irani M and Peleg S. Motion analysis for image enhancement: Resolution, occlusion, and transparency. J Vis Commun Image R. 1993;4(4):324–335
4. Andreychenko A, Raaijmakers AJE, Sbrizzi A, Crijns SPM, Lagendijk JJW, Luijten PR, and van den Berg CAT. Thermal noise variance of a receive radiofrequency coil as a respiratory motion sensor. Magn Reson Med. 2017;77(1):221–228.