Author: Zachary Buchwald, Chih-Wei Chang, Zach Eidex, Hui Mao, Richard L.J. Qiu, Justin R. Roper, Mojtaba Safari, Hui-Kuo Shu, Xiaofeng Yang, David Yu 👨🔬
Affiliation: Emory University and Winship Cancer Institute, Emory University, Department of Radiation Oncology and Winship Cancer Institute, Emory University, Emory University School of Medicine 🌍
Purpose: MRI-guided radiation therapy (MRgRT) benefits significantly from enhanced soft-tissue contrast and spatial resolution, which aid in accurately delineating tumors and organs at risk. Although 7T T1-mapping offers superior tissue contrast and resolution compared to lower field strengths, its use is limited by high costs, potential patient discomfort, and limited availability. This study presents a deep-learning framework designed to synthesize high-quality 7T T1 maps from 1.5T or 3T T1-weighted MRI.
Methods: We developed the Adaptive Norm Vision Transformer (ViT), a model capable of synthesizing high-resolution MR images from 1.5T to 7T and 3T to 7T by integrating adaptive layer norms that adjust to inputs from either 1.5T or 3T MRIs within a single framework. The model was trained and validated on an institutional dataset of 139 patients, split into training (n=107), validation (n=16), and test (n=16) groups, using 1.5T and 3T T1-weighted MRIs alongside corresponding 7T MP2RAGE T1 maps. Model performance was evaluated using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), normalized mean squared error (NMSE), and normalized cross-correlation (NCC).
Results: Synthesized 7T T1 maps demonstrated high visual fidelity and quantitative agreement with actual 7T data. Metrics for Adaptive Norm ViT were: for 1.5T to 7T conversion - PSNR: 24.4 ± 4.1, SSIM: 0.840 ± 0.085, NMSE: 2.58 ± 1.49E-2, NCC: 0.967 ± 0.038; for 3T to 7T conversion - PSNR: 23.1 ± 5.6, SSIM: 0.809 ± 0.121E-2, NMSE: 5.45 ± 7.22E-2, NCC: 0.941 ± 0.051.
Conclusion: The proposed framework effectively enables the synthesis of 7T MP2RAGE T1 maps from standard 1.5T or 3T MRI scans, potentially enhancing tumor and normal tissue delineation in MRgRT. This generalizable approach could significantly advance clinical workflows by improving contrast and spatial resolution without the need for high-field MRI scanners.