Presenter(s)
J. Ma1, J. Chen1, W. Gao2, X. LI3, H. Xiao4, and Y. Yin5; 1Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, Shandong, China, 2Cancer Hospital of Shandong First Medical University, Shandong, China, 3Cancer Hospital of Shandong First Medical University, jinan, China, 4Shandong Provincial Key Medical and Health Laboratory of Pediatric Cancer Precision Radiotherapy (Shandong Cancer Hospital), Jinan, Shandong, China, 5Department of Radiation Oncology Physics and Technology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China
Purpose/Objective(s): 4D-MRI offers superior soft-tissue contrast for abdominal motion management but is often limited by the trade-off between image quality and scan duration. Moreover, existing techniques are largely restricted to anatomical imaging, lacking the integration of functional data. This study proposes a deep learning framework to generate high-quality (HQ) 4D anatomical and functional MRI from low-quality (LQ) inputs via accurate motion estimation and reconstruction, aiming to significantly enhance tumor tracking reliability and precision in Image-Guided Radiotherapy (IGRT).
Materials/Methods: A total of 169 abdominal MRI datasets with complete liver coverage were partitioned into training and internal validation cohorts (8:2). An external test cohort (n=12) was acquired. The model employed a 3D U-Net-based architecture, utilizing HQ static 3D-MRI and LQ 4D-MRI inputs to synthesize HQ 4D anatomical and functional MRI, and deformation vector fields (DVFs). To ensure spatial consistency, DWI was aligned via a hierarchical cross-contrast registration (HCR) pipeline. Image quality was quantified using Full Width at Half Maximum (FWHM) and CNR, while motion consistency was assessed via liver centroid trajectories.
Results: In the validation and testing cohorts, quantitative motion analysis demonstrated consistent sub-voxel accuracy. For every patient and across all motion directions, the mean 3D trajectory errors were consistently <1 mm. Furthermore, the maximum error for each case remained below the original voxel dimensions (Training:1.56×1.56×3.0mm³; Testing: 2.68×2.68×2.7mm³). For a representative test case, Figure 1 presents the motion tracking curves and quality metrics (CNR, FWHM). Figure 2 displays the generated HQ images (T1, T2, and DWI). As shown in Figure 2, while the tumor was indistinguishable on anatomical sequences, the generated HQ-DWI delineated the lesion, validating the model's capability to recover functional information for target definition.
Conclusion: The proposed framework successfully reconstructs HQ 4D anatomical and functional MRI from LQ inputs while maintaining precise motion information. This personalized, multi-parametric 4D-MRI approach demonstrates feasibility for fast and reliable motion management, potentially enabling high-precision IGRT for abdominal cancers.