Author: Hilary P Bagshaw, Mark K Buyyounouski, Cynthia Fu-Yu Chuang, Yu Gao, Dimitre Hristov, Lianli Liu, Lawrie Skinner, Lei Xing 👨🔬
Affiliation: Department of Radiation Oncology, Department of Radiation Oncology, Stanford University 🌍
Purpose:
MR-guided radiation therapy has introduced a significant leap in cancer treatment by allowing adaptive treatment. The low-field MR-guided system predominantly uses the TrueFISP sequence, which is effective for anatomical alignment but lacks the detailed tissue characteristics provided by multiparametric MRI. Additionally, low-field MR imaging suffers from increased noise due to the limitations of reduced magnetic field strength. This study seeks to synthesize high quality T2-weighted (T2-w) MR images from the TrueFISP scans of a low-field MR-Linac system to better support MR-guided workflows.
Methods:
Data from 15 prostate cancer patients treated with the ViewRay MRIdian system were analyzed. For each patient, T2-weighted reduced field of view SPACE MRIs were acquired using a 3T diagnostic scanner, while TrueFISP MRIs were obtained from the onboard 0.35T MRI scanner. A CycleGAN network was trained to perform unpaired image-to-image translation between the two MRI scans. The training set consisted of 796 slices from 12 patients, and the testing set consisted of 192 slices from additional 3 patients. The model was trained over 200 epochs using an NVIDIA GeForce RTX™ 4090 GPU, with training completed in approximately 3.5 hours.
Results:
The network demonstrated successful convergence on the training dataset. The synthetic images on the testing patients had pronounced T2-weighted contrast that drastically enhanced the visualization of tissue characteristics compared to the low-field TrueFISP MRI. While mild breathing motion artifacts were noted in two testing patients, the network demonstrated robustness to motion, generating high-quality, artifact-free T2-weighted images despite the presence of motion artifacts in the original TrueFISP scans.
Conclusion:
This study highlights the potential of deep learning models to synthesize high-quality T2-weighted MRIs from low-field TrueFISP scans without registered image pairs. The enhanced image quality could improve the precision of tumor and organ-at-risk delineation, advancing the precision and utility of low-field MR-Linac systems.