Purpose: To develop a novel multi-parametric ultra-quality 4D-MRI technique through deep learning synthesis using a VoxelMorph-based deformable image registration (DIR) model.
Methods: Twenty-seven liver tumor patients undergoing radiotherapy were collected from Beijing Cancer Hospital with IRB approval. Each patient received a T1w 4D-MRI scan using TWIST volumetric interpolated breath-hold examination (TWIST-VIBE) sequence. Patients also received multi-parametric 3D MR scans, including T1w, T2w, and diffusion-weighted MR imaging (DWI) with b-value of 50 and 800. 4D-MRI frames at the same respiratory phases with 3D MR scans were selected, and a VoxelMorph-based model was trained to obtain the deformations between selected frames and other frames. The deformations were then applied to those 3D MRIs to synthesize multi-parametric ultra-quality 4D-MRIs at corresponding respiratory phases. Contrast-to-noise ratio and full-width-half-maximum (FWHM) of the profile along lung-liver interface were calculated on different MRIs to quantitatively evaluate image quality, and tumor motion trajectories were measured to determine motion errors and subsequently the motion accuracy of synthetic 4D-MRIs.
Results: CNR of the original T1w 4D-MRI was 6.62±3.85, and it increased to 10.54±5.64, 12.83±7.66, 25.36±14.26, and 30.42±14.02 in the synthetic ultra-quality T1w, T2w 4D-MRI, 4D-DWI (b=50), and 4D-DWI (b=800), respectively. FWHM of the liver-lung profile reduced from 6.7±2.9 mm to 4.7±2.4 mm, 3.1±1.3 mm, 6.4±4.5 mm, and 4.5±1.4 mm in the synthetic ultra-quality T1w and T2w 4D-MRI, 4D-DWI (b=50), and 4D-DWI (b=800), respectively. In addition, relative tumor motion errors were much smaller than the voxel sizes of the 3D-MRIs.
Conclusion: This study demonstrated the feasibility of a novel multi-parametric ultra-quality 4D-MRI with deep learning-based DIR. This method overcomes many limitations of current 4D-MRI techniques and holds great promises for clinical applications.
Funding Support, Disclosures, and Conflict of Interest: This research was partly supported by the following Hong Kong research grants: General Research Fund (GRF): GRF 151021/18M; Health and Medical Research Fund (HMRF): HMRF 06173276