Sep 28

QP 14 - Modernizing Definitive Therapy in Cervical Cancer: From Systemic Intensification to Adaptive Radiation

1082 - Contrast-Free Virtual Enhancement: Multi-Modal Deep Learning Synthesis of CECT from pCT and MRI

05:25pm - 05:30pm ET

Room 204

Presenter(s)

Haonan Xiao, PhD - Shandong First Medical University Affiliated Tumor Hospital, Jinan City, Shandong

J. Chen¹, H. Xiao², and Y. Yin³; ¹Cancer Hospital of Shandong First Medical University, Shandong, China, ²Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, Shandong, China, ³Department 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):

Contrast-enhanced CT (CECT) is routinely used in radiotherapy planning to improve visualization of tumors, lymph nodes, and vascular anatomy for accurate target delineation. However, contrast administration is contraindicated in a subset of patients (e.g., iodine allergy, renal insufficiency), and reliance on non-contrast planning CT (pCT) alone may compromise soft-tissue contrast and introduce contouring uncertainty. We propose a Multi-Modal Generative Synthesis Network (MMGSN) to generate synthetic CECT (sCECT) from pCT and T2-weighted MRI, aiming to provide a non-invasive “virtual enhancement” tool for contouring support without physical contrast injection.

Materials/Methods:

We retrospectively collected 47 pelvic tumor patients treated with radiotherapy, each with paired non-contrast pCT, T2-weighted MRI, and ground-truth CECT. For each patient, the CECT volume used in this study consisted of 40 axial slices (used as the reference for training/evaluation). MMGSN adopts a dual-stream encoder to fuse the geometric fidelity of pCT with the soft-tissue characterization of MRI, followed by a decoder to reconstruct high-fidelity sCECT. Training employed a compound objective combining adversarial loss, weighted L1 loss, and SSIM loss. Performance was evaluated on a held-out test set using PSNR, SSIM, and MSE, and compared against the baseline similarity between pCT and CECT.

Results:

MMGSN generated sCECT with high quantitative fidelity to ground-truth CECT, achieving a mean PSNR of 43.36 dB and mean SSIM of 0.979 on the test set. Compared with the baseline (pCT vs CECT), MMGSN demonstrated a clear improvement in perceptual/structural similarity (e.g., representative case: SSIM 0.978 vs 0.966 and PSNR 43.088 vs 39.757, respectively). As illustrated in the tri-planar views (axial/coronal/sagittal), sCECT better reproduced contrast-related appearance and improved delineation cues for pelvic vasculature and soft-tissue boundaries relative to pCT, while maintaining geometric consistency required for radiotherapy planning.

Conclusion:

MMGSN enables high-quality generation of contrast-free sCECT from pCT and MRI in pelvic radiotherapy patients. This approach has the potential to provide clinically useful “virtual enhancement” for target/OAR contouring in patients with contrast contraindications, improving visualization without exposing patients to contrast-related risks and potentially streamlining radiotherapy workflows.