Yang, ZhenyuYin, Fang-FangGuo, Dongyang2025-07-022025-07-022025https://hdl.handle.net/10161/32933<p>Introduction: Real-time tumor visualization during liver stereotactic body radiation therapy (SBRT) remains clinically challenging due to cone-beam CT's (CBCT) inherent limitations in soft-tissue contrast. Although MR-to-CBCT registration offers a promising solution, conventional approaches face critical limitations in addressing cross-modality deformation complexity, balancing computational efficiency with biomechanical accuracy, and managing large anatomical variations.Methods: We present a biomechanically guided deep learning framework validated through a comprehensive five-fold cross-validation study on 42 liver cases (27 clinical MR-CBCT pairs and 15 simulation cases with ground truth deformations). Our approach achieves both physical accuracy and computational efficiency through three key innovations: (1) a biomechanically-guided network architecture that explicitly encodes tissue mechanics, (2) a physics-constrained loss function combining image similarity and deformation plausibility, and (3) a robust training strategy incorporating multi-scale deformation augmentation. The framework's performance was rigorously evaluated using multiple metrics including DICE coefficient, target registration error, and Jacobian determinant analysis, ensuring both registration accuracy and physical plausibility. Results: Our proposed biomechanically guided deep learning framework significantly enhanced registration accuracy and anatomical plausibility. On simulated data, the method achieved an average Target Registration Error (TRE) of 7.72\pm5.16 mm, substantially outperforming purely deep learning-based methods (14.68\pm12.31 mm). In clinical evaluations, the framework demonstrated robust performance with an average TRE of approximately 2.22 mm, achieving clinical accuracy (≤3 mm) for 98.4% of evaluated landmarks. Directional analysis highlighted consistently lower registration errors, particularly along the Z-axis. Qualitative assessments and case-specific analyses further demonstrated superior robustness and anatomical consistency compared to state-of-the-art baseline methods. Conclusion: This breakthrough enables precise tumor targeting during daily treatment by providing real-time MRI-quality visualization in CBCT guidance. The framework's ability to combine physical accuracy with real-time performance represents a significant advancement toward liver SBRT. The improved targeting accuracy could potentially enable dose escalation strategies and better treatment outcomes without disrupting clinical workflow. </p>https://creativecommons.org/licenses/by-nc-nd/4.0/Medical imagingMaster's thesis