Respiratory motion can impact the efficacy and safety of radiation therapy in the thorax and abdomen. For treatments using an MRI-guided linac, free-breathing 4D-MRI is a promising alternative to 4D-CT for motion management, providing excellent soft-tissue contrast with no ionizing radiation. High-quality MR images free from motion artefacts are needed to delineate lesions from normal tissue. Currently, however, MR-based approaches require multiple scans with substantial scan times.

To meet these needs, Sihao Chen, Hongyu An and colleagues at Washington University in St. Louis are developing a way to use a single MRI scan for motion detection, motion-resolved 4D-MRI and motion-integrated 3D-MRI reconstruction. Speaking at last week’s AAPM Annual Meeting, Chen showed that this is possible with an acquisition time of less than a minute, using a self-navigated MR method with deep learning-based image reconstruction.

The three-stage technique begins with a self-navigated respiratory motion detection sequence called CAPTURE, which is a variant of the stack-of-stars MRI sequence. The researchers implemented CAPTURE on the 0.35 T ViewRay MRI-guided linac and evaluated their proposed technique by imaging a respiratory motion phantom and 12 healthy volunteers. They performed regular MRI scans using 2000 radial spokes, with an acquisition time of 5–7 min. They evaluated the full scan (2000 radial spokes), as well as the first 10% of the data, which took just 30–40 s.

Chen shared some example CAPTURE-detected respiratory curves, which demonstrated CAPTURE’s ability to detect respiratory motion despite different respiratory patterns between subjects and during individual scans. The corresponding frequency spectra clearly identified the individual frequency components.

Next, the team used the measured respiratory signals to create 4D-MRIs via three reconstruction techniques: multi-coil non-uniform inverse fast Fourier transform (MCNUFFT); compressed sensing; and deep learning-based Phase2Phase (P2P) reconstruction.

In a motion phantom study, the team reconstructed 4D-MR images using either 5 min or 30 s of data. The CAPTURE motion detection improved the visibility of embedded spheres in the phantom to the level seen in ground truth images. In the short MRI scan, P2P reconstruction restored image sharpness and reduced undersampling artefacts compared with the non-corrected baseline.

For the patient scans, the researchers used the first 200 spokes for short-scan (30 s) reconstruction, observing that P2P clearly outperformed the other two methods for 4D-MRI reconstruction. They then used 4D-MRIs created from both the 30 s and 5 min scans to derive motion vector fields. Chen noted that the difference between the two was “moderate compared with the overall motion range”.

In the final step, these motion vector fields are employed to reconstruct 3D-MRIs using a motion integrated reconstruction (MOTIF) model. 3D-MR images of the phantom demonstrated that MOTIF reduced motion artefacts and improved image quality. In the patient study, short-scan images (200 spokes) reconstructed by MOTIF had better signal-to-noise ratio and fewer motion artefacts than the non-corrected baseline, and demonstrated “modest image quality” compared with regular-scan images (2000 spokes) reconstructed by MOTIF.

Read more

Best in physics: multidimensional MRI and FLASH proton therapy

The team also performed a blinded radiological review of the 12 subjects. Images reconstructed by MOTIF using the entire data set scored over 8/10 points when rated for sharpness, contrast and lack of artefacts. “For short scans, MOTIF with P2P received a relatively satisfactory review score of 5/10, whereas no motion correction scored less than 3/10,” said Chen.

Chen concluded that a rapid single MRI scan, used with CAPTURE, P2P and MOTIF, can generate high-quality 4D-MR images for lesion motion range determination and 3D-MR images for lesion delineation on a low-field MRI-guided linac.