Model-based spatiotemporal synthetic data generation framework and deep-learning reconstruction for real-time MRI oxygen extraction fraction mapping.
Authors
Affiliations (8)
Affiliations (8)
- Xiamen University, Xiang'an Campus, Xiamen University, Xiang'an South Road, Xiamen, Xiamen, 361102, CHINA.
- Department of Electronic Science, Xiamen University, Xiang'an Campus, Xiamen University, Xiang'an South Road, Xiamen, Xiamen, Fujian, 361102, CHINA.
- School of Ocean Information Engineering, Jimei University, No. 185, Yinjiang Road, Jimei District, Xiamen, Xiamen, Xiamen, None Selected, 361021, CHINA.
- Department of Radiology, Zhongshan Hospital Fudan University Xiamen Branch, 668 Jinhu Road, Huli District, Xiamen, Xiamen, Fujian, 361015, CHINA.
- Department of Magnetic Resonance Imaging, The First Affiliated Hospital of Zhengzhou University, No. 1 Jianshe Dong Road, Erqi District, Zhengzhou, Henan, 450052, CHINA.
- School of Intelligent Medicine and Information Engineering, Jiangxi University of Chinese Medicine, No. 1688, Meiling Avenue, Bayi District, Nanchang, Nanchang, 330004, CHINA.
- Department of Electronic Science, Xiamen University, Xiang'an Campus, Xiamen University, Xiang'an South Road, Xiamen, Xiamen, 361102, CHINA.
- Department of Imaging Sciences, University of Rochester, 601 Elmwood Avenue, Box 648, Rochester, Rochester, New York, 14627, UNITED STATES.
Abstract
Synthetic data has emerged as a highly efficient solution to address the scarcity of training data in deep learning-based quantitative magnetic resonance imaging (qMRI) reconstruction. However, current applications of synthetic data predominantly focus on two-dimensional spatial reconstruction, with limited capability to leverage the spatiotemporal correlation inherent in real-time dynamic imaging data to further enhance reconstruction quality. 
Approach: A model-based spatiotemporal synthetic data generation framework tailored for real-time dynamic scenarios in supervised learning-based reconstruction was proposed. This framework is complemented by an ultra-fast multiple overlapping-echo detachment (MOLED) imaging technique, which enables a three-dimensional spatiotemporal reconstruction method designed to track dynamic changes in the oxygen extraction fraction (OEF). The proposed method imposes constraints that align the predicted parameters with the ground truth while ensuring consistency in their singular-value subspaces.
Main results: The accuracy and effectiveness of the proposed method for T2 and T2* mapping were validated through numerical brain, water phantom and human brain experiments, demonstrating superior performance compared to both traditional two-dimensional spatial and three-dimensional spatiotemporal reconstruction methods. This framework reliably enabled dynamic OEF tracking during breath-hold and oxygen inhalation cycles, highlighting its robustness and applicability in real-time scenarios.
Significance: The model-based spatiotemporal synthetic data generation framework and the spatiotemporal reconstruction method offer an effective and robust solution for real-time qMRI, with potential applications in other similar real-time dynamic quantitative reconstruction tasks. Moreover, the MOLED sequence underscores the potential for precise and dynamic measurement of brain oxygen metabolism, providing valuable insights into cerebral physiology and metabolic changes.