144 - Single-CT Simulation for Surface-Guided DIBH Radiotherapy Using Deep Learning-Synthesized 3D Body Contours from Scout Images

05:10pm - 05:20pm ET

Room 253

Presenter(s)

Inbum Lee, PhD - Seoul National University Hospital, Seoul, SEO

Y. Huh^1,2, J. Yang³, Y. S. Lee², Y. K. Kwak², S. Hwang², J. Jegal^1,2, I. Lee^1,2, J. H. Chang^2,4, C. H. Choi^1,2, S. Kang^1,2, and J. I. Kim^1,2; ¹Institute of Radiation Medicine, Seoul National University Medical Research Center, Seoul, Korea, Republic of (South), ²Department of Radiation Oncology, Seoul National University Hospital, Seoul, Korea, Republic of (South), ³3Biomedical Research Institute, Department of Radiology, UT Southwestern Medical Center, Dallas, TX, ⁴Department of Radiation Oncology, Seoul National University College of Medicine, Seoul, Korea, Republic of (South)

Purpose/Objective(s): Deep inspiration breath hold (DIBH) combined with surface-guided radiation therapy (SGRT) reduces cardiac and pulmonary dose in breast radiotherapy. However, many clinical workflows require both free-breathing (FB) and DIBH CT acquisitions, increasing scan time and imaging dose. We hypothesized that accurate 3D FB body contours could be synthesized from routine orthogonal CT scout images with geometric fidelity within clinically acceptable SGRT setup tolerances, potentially enabling single-CT simulation workflows.

Materials/Methods:

Following IRB approval, 176 thoracic CT simulation sessions with paired coronal and sagittal CT scout images and an FB-CT volume were retrospectively collected. Reference body contours from the RT Structure Set were rasterized on FB-CT and converted into binary filled-body masks. A dual-view 2D-to-3D deep learning network was trained using the two scout images as input and the filled-body mask as ground truth (80/20 train–validation split). Geometric accuracy was evaluated on an independent test cohort (n = 34) using Dice coefficient (DC), 95th percentile Hausdorff distance (HD95), and mean surface distance (MSD).

Results: In the independent test cohort, predicted contours achieved DC = 0.979 ± 0.005, HD95 = 4.01 ± 1.00 mm, and MSD = 1.52 ± 0.34 mm (mean ± SD across patients). The average surface discrepancy of 1.52 mm falls within commonly reported 3–5 mm clinical action levels for SGRT-based positioning, indicating geometric compatibility with surface-guided setup. Performance was consistent across patients without evidence of instability or outlier-driven degradation.

Conclusion: Accurate 3D body contours can be synthesized from two orthogonal CT scout images with sub–2 mm mean surface discrepancy. This approach may eliminate the need for a separate FB-CT acquisition in SGRT-based DIBH workflows, reducing imaging dose and simulation complexity while preserving surface registration accuracy. Prospective surface-registration validation and end-to-end workflow testing are ongoing.