3204 - An Information-Augmentation Driven Paradigm for Intra-Fractional Respiratory Modeling-Based Anatomical Deformation Reconstruction
Presenter(s)
G. Zhang1, Z. Jiang2, H. Shu3, and L. Wang4; 1Artificial Intelligence Laboratory, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Ji'nan, China, 2Department of Radiation Oncology Physics and Technology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China, 3Southeast University, Laboratory of Image Science and Technology, The Key Laboratory of Computer Network and Information Integration, Ministry of Education, Jiangsu Provincial Joint International Research Laboratory of Medical Information Processing, Cent, Nanjing, China, 4Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Ji'nan, China
Purpose/Objective(s): To advance sparse-information constrained respiratory modeling toward an information-augmentation driven paradigm by incorporating data augmentation, multimodality guidance, prior-informed representation, and improved optimization pipelines.
Materials/Methods:
We introduce a patient-specific motion-modeling framework for internal deformation reconstruction. To enhance respiratory-phase diversity and variability characterization, a multi-stage data augmentation pipeline integrates PCA-based motion characterization, static velocity field(SVF) interpolation, setup error simulation, and variational autoencoder(VAE)-based generation, which i) effectively expands temporal-phases, ii) incorporates irregular respiratory variabilities, and iii) preserves diffeomorphism and physical plausibility. Instead of direct image synthesizing/structural tracking, we estimate intermediate DVFs to mitigate image blurring, which benefits from prior-enhancements of respiratory characteristics captured by PCA-derived DVF-eigenvectors. We also propose cross-domain collaborative optimization to integrate advantages of supervised-and-unsupervised learning by contributing an implicit optimization-loop with cross-domain informing and regularization, which i) enhances modeling adaptabilities to unknown phases through enforcing surrogate consistency, and ii) eliminates impacts of surrogate variations on modeling robustness. We assemble a multi-center, multi-site 4D database enrolling 35 patients—10 prospective and 25 retrospective—undergoing IMRT/IMPT.Results: We achieved RMSE of 1.337m-1, PSNR of 31.510dB, and SSIM of 0.958 for CT reconstruction, while deviation amplitude of centroid (DCAM) of 0.385/0.310 mm, DSC of 0.958/0.967 and relative volume changes of 20.2%/14.6% for tumor/whole-lungs tracking. Within a IMRT planning scheme (prescribed dose: 60Gy; beam configuration: 0°,40°,80°,120°,160°,200°), no significant differences were observed in voxel-wise dose differences(P=0.04, two-side T-tests), DVH indices (P=0.03 for V20 of whole-lung; P<0.01 for D95 of tumor), global Gamma-passing-rates(P<0.01), dose HI(P=0.02), and CI(P=0.02). No dramatic fluctuations exhibited across full-respiratory-cycle, with standard deviations of 0.122m-1, 0.687dB, and 0.019 for RMSE, PSNR, and SSIM, respectively. High correlation were demonstrated between reconstructed images and ground-truths, with Pearson correlation coefficients of 0.959±0.029(SI), 0.955±0.036(LR), and 0.923±0.039(AP).
Conclusion: Comprehensive multi-center evaluations have substantiated methodological sophistication of our framework, which represents a promising advance toward adaptive respiratory tracking-enhanced precise radiotherapy.