3721 - Realistic High-Fidelity CT Simulation for AI-Aided Radiation Oncology: Bridging Data Gaps and Improving Clinical Decision
Presenter(s)
Z. Shang1, J. Dong2, and W. Zhao1,3; 1School of Physics, Beihang University, Beijing, China, 2Department of Radiotherapy, Beijing Hospital, Beijing, China, 3Hangzhou International Innovation Institute, Beihang University, Zhejiang, China
Purpose/Objective(s): Artificial intelligence (AI) has been extensively investigated in radiotherapy to improve precision, efficiency, and the personalization of treatment. However, AI development is constrained by limited access to large-scale, clinically representative imaging data due to ethical and logistical barriers. Simulation methods have been proposed to mitigate the challenge, but Traditional simulation methods often fail to replicate real-world CT scan complexities, creating a "domain gap" that hinders AI performance. This study presents a high-fidelity CT simulation framework to generate synthetic images that closely mimic clinical data, addressing data scarcity and enabling robust AI training for radiation oncology applications.
Materials/Methods: We introduced a numerical procedure to construct realistic phantoms from routine CT scans by decomposing them into elementary-based voxels, each assigned a voxel-wise mass attenuation coefficient. A spectrum ray-tracing platform, augmented by parallel computing, was employed to simulate CT scans under diverse conditions (e.g., different spectra, detectors, and scan parameters) that consistent with routine scanning protocols. To validate the effectiveness of the simulated data for clinical use, deep learning models were trained on the simulated dataset for two critical tasks: image denoising and Z-axis super-resolution. These tasks were designed to assess the ability of the simulated images to enhance AI model performance in a clinical context.
Results: Simulated CT images demonstrated high fidelity to real-world data, with Wasserstein distances (116.76 ± 1.23 vs. 114.90 ± 1.15) and texture similarity indices confirming minimal domain discrepancies. In denoising tasks, models trained on simulated data achieved superior performance, yielding a peak signal-to-noise ratio (PSNR) improvement from 45.02 dB to 53.52 dB. Super-resolution results exhibited enhanced spatial resolution (0.1-mm reconstruction accuracy) compared to conventional methods, enabling better visualization of anatomical details critical for target delineation. These findings highlight the potential of synthetic data to augment AI training pipelines without compromising clinical relevance.
Conclusion: This study presents a realistic CT simulation method that effectively bridges the data gap between simulations and real-world imaging. By generating high-fidelity simulated CT images, we demonstrated improvements in the performance of AI models for denoising and Z-axis super-resolution tasks. Our results suggest that the proposed simulation method holds significant promise for supporting AI-driven tools in radiotherapy, ultimately improving treatment accuracy and patient outcomes.