Go to AFOMP Pulse Contents/Index Page
(XIAO Haonan, Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong SAR, China)
Background: Four-dimensional (4D) magnetic resonance imaging (MRI) is becoming popular inliver cancer radiationtherapy (RT)for its excellent soft-tissue contrast in the abdominal region. However, most available 4D-MRI techniques suffer from insufficient image quality, long acquisition time, or require specially designed sequences or hardware not available in the clinic. These limitations have greatly hindered the clinical implementation of 4D-MRI.
Purpose: This study aims to solve the abovementioned deficiencies of existing 4D-MRI techniques by developing novel ultra-quality (UQ) 4D-MRI methods capable of motion management and real-time tumor tracking in liver RT using a commercially available 4D-MRI sequence and deep learning-based registration models.
Methods: In the first part, an UQ 4D-MRI method was developed. Thirty-nine patients receiving RT for liver tumors were included, each received 4D-MRI scan and multi-parametric (Mp) 3D-MRI scans as prior images. UQ 4D-MRI at any instant was considered a deformation of the prior images, and the deformations was obtained via a dual-supervised deformation estimation model (DDEM). The registration accuracies of DDEM, VoxelMorph (normalized cross-correlation (NCC) supervised), VoxelMorph (end-to-end point error (EPE) supervised), and parametric total variation (pTV) algorithm were compared. Tumor motion on UQ 4D-MRI was evaluated using region-of-interest (ROI) tracking errors, while image quality was evaluated using the contrast-to-noise ratio (CNR), lung–liver edge sharpness, and perceptual blur metric (PBM). In the second part, the proposed UQ 4D-MRI method was further extended on temporal efficiency to be ultra-fast high-quality (UFHQ) Mp 4D-MRI by obtaining deformations from retrospectively downsampled 4D-MR images via dual-supervised downsampling-invariant deformable registration (D3R) model. Besides all the above-mentioned evaluation metrics, the registration robustness of the D3R model was compared to iterative registration methods, including Demons, Elastix, and pTV algorithm.
Results: The registration accuracy of the DDEM was significantly betterthan all the other methods, with an inference time of 69.3±5.9 ms.The registration robustness of the D3R model was also significantly better than all the iterative methods, giving stable DVF prediction and higher image similarities at downsampling factors up to 500. UQ 4D-MRI yielded ROI tracking errors of 0.79±0.65, 0.50±0.55, and 0.51±0.58 mm in the superior-inferior (SI), anterior-posterior (AP), and mid-lateral (ML) directions, respectively. From the original 4D-MRI to UQ 4D-MRI, the CNR increased from 7.25±4.89 to 18.86±15.81; the lung–liver edge full-width-at-half-maximum decreased from 8.22±3.17 to 3.65±1.66 mm in the in-plane direction and from 8.79±2.78 to 5.04±1.67 mm in the cross-plane direction, and the PBM decreased from 0.68±0.07 to 0.38±0.01.The UFHQ Mp 4D-MRI yielded ROI tracking error of 1.18±1.20, 0.52±0.55, and 0.41±0.47 mm in the SI, AP, and ML directions and similar image quality improvement as the UQ 4D-MRI.
Conclusion: We have successfully demonstrated novel 4D-MRI techniques for liver RT. Compared with the original images, UQ 4D-MR images provided versatile image contrast, improved image quality, and accurate tumor motion trajectories within short processing times. The UFHQ Mp 4D-MRI technique further enhanced the temporal efficiency, making the imaging frequency greater than 3 Hz. These methods show great promise to expand the clinical implementation of 4D-MRI for motion management in liver RT.