Synthetic 4D CT for Adaptive MR-Guided Radiotherapy Treatment on an MR-Linac
Loading...
Embargo End Date
2025-04-07
ICR Authors
Authors
Goodburn, R
Document Type
Thesis or Dissertation
Date
2024-10-07
Date Accepted
Abstract
Purpose: To investigate and develop synthetic CT (synCT) methods for adaptive MR-guided radiotherapy of lung cancer on a 1.5T MR-Linac. This application is challenging due to 1) respiratory motion, 2) rapidly decaying MRI signal in lung tissue, 3) complex tissue-density structure in the thorax, and 4) the need for fast acquisition and reconstruction methods.
Objectives: Develop a synCT technique which 1) is respiratory resolved (4D), 2) employs ultrashort echo time (UTE) imaging to minimise signal loss related to lung’s fast MRI relaxation parameters, 3) uses a machine-learning based approach to generate synCT, and 4) performs image acquisition and reconstruction in a clinically feasible time frame.
Methods: A UTE-Dixon sequence was developed on an Elekta Unity MR-Linac and used to acquire data for 13 lung-cancer patient volunteers who underwent 4D-CT as part of their workup. To allow for offline reconstruction of UTE images, the gradient system transfer function (GSTF) of the MR-Linac was measured with a phantom-based approach and used to correct k-space trajectories. Self-navigation, and two iterative, compressed sensing algorithms were used to reconstruct 4D UTE-Dixon, fat, and water images. For each patient, 4D-CT images were collected and used with 4D UTE/fat/water images to train a cycle-GAN to generate synCT. One synCT image volume was evaluated by comparison of Hounsfield units and dose metrics for a 4D-CT of the same patient. To accelerate a motion-compensated 4D-MRI reconstruction technique, a convolutional neural network for fast image registration was trained using 4D abdominal data acquired on the MR-Linac and compared with a Demons registration algorithm.
Results: Acquisition, reconstruction, and generation methods for 4D synCT of the thorax were demonstrated using MR-Linac data. Measuring Hounsfield units in the cardiac blood yielded 40.3+/-33.5 vs 38.0+/-11.5 (mean+/-stdv) The machine-learning based registration model performed well and was 48× faster than a standard Demons registration algorithm. The methods developed in this work could be used to further develop the cycle-GAN approach and validate dose metrics across a range of patients.
Citation
2024
DOI
Source Title
Publisher
Institute of Cancer Research (University Of London)
ISSN
eISSN
Collections
Research Team
Magnet Resonance Imaging