Motion Field Integrated Reconstruction Using Deep Learning for MR and PET/MR Respiratory Motion Correction
Since motion is unavoidable in clinical human imaging studies, motion correction in PET and MRI has long been of interest for the liver/lung imaging community. Respiratory motion, the most crucial source of motion in body imaging, affects thoracic organs and the upper/lower abdomen. In MRI, motion can severely blur the images and create artifacts due to incorrect sampling of k-space data in the Fourier domain. In PET imaging, the periodic respiratory motion negatively impacts the detection of small lesions and quantification of PET tracer uptake values. Typically, breath-holding is used in clinical MR scans. However, breath-holding limits spatial coverage and image quality, and it can sometimes be extremely difficult to perform for some patients with comorbidities. In this aspect, free-breathing respiratory motion correction methods are highly desirable. The objective of this dissertation was to develop motion field integrated reconstruction using deep learning to resolve respiratory motion for MRI and PET/MRI.
To correct deformable respiratory motion using free-breathing methods, two key components are needed: 1) respiratory motion detection for imaging data re-binning, 2) deformable motion vector field to correct the motion. Respiratory motion detection and re-binning was proposed to be derived from a self-navigated MR sequence called CAPTURE during free-breathing. Respiratory motion-resolved 4D-MRIs can be re-binned based on the respiratory motion curve and then reconstructed using either a Fourier transform reconstruction for long scans (~5 minutes) or a deep-learning based Phase2Phase reconstruction for short scans (~30 seconds). Subsequently, motion vector fields are derived from those 4D-MRIs and are served as the motion model for respiratory motion correction. The resulting motion vector fields are then used in the proposed motion integrated reconstruction to transform all k-space data to the reference motion state (end-of-expiration) during reconstruction.
First, the proposed respiratory motion correction method was developed for a low-field 0.35 T MRI-guided linear accelerator. Phase2Phase network and motion field reconstruction was tested for 4D-MRI and 3D-MRI respectively using 10% of the nominal acquisition time, while yielding comparable spatial and dynamic motion information of the liver. Second, respiratory motion information derived from the simultaneously acquired MRI motion scan during PET/MRI imaging session was used to re-bin PET list mode data. MR-assisted motion correction using motion compensated image reconstruction was evaluated in both physical motion phantom and patient studies. Motion corrected PET images using either short or long MRI modeling scans exhibited significant improvements in SUVs and lesion volume compared to the motion uncorrected PET images. Lastly, deep-learning prior was proposed as regularization in the motion field integrated reconstruction for continuous free-breathing dynamic contrast-enhanced MRIs. The proposed deep-learning based motion integrated reconstruction demonstrated superb image quality and artifact-freeness from severely undersampled MRI data, which may allow for improved dynamic contrast enhancement quantification and easiness for diagnosis.