Synthetic MRI from radiotherapy planning CT images for Head-and-Neck Cancer.
- 1Department of Radiation Oncology, Princess Alexandra Hospital, Woolloongabba, Queensland, Australia.
- 2Australian Institute for Bioengineering and Nanotechnology, The University of Queensland, St Lucia, Queensland, Australia.
- 3School of Chemistry and Physics, Queensland University of Technology, Brisbane, Queensland, Australia.
Abstract
Background
Computed Tomography (CT) provides geometric accuracy and electron density information essential for radiotherapy planning; however, it offers limited soft-tissue contrast. Magnetic Resonance Imaging (MRI) improves tumor visualization but introduces registration uncertainties when acquired separately from the radiotherapy planning CT (pCT). Cross-modality image synthesis enables the generation of synthetic MRI (sMRI) directly from pCT and may combine the strengths of both modalities, reducing reliance on MRI access.Purpose
We propose a two-stage deep-learning framework for generating sMRI from pCT image-sets specifically in head-and-neck (H&N) cancers, providing soft-tissue contrast comparable to ground-truth MR images (gtMRI).Methods
This study included 60 patient cases with H&N cancer treated with radical radiotherapy between March 2022 and April 2024 for model development (50 for training, 10 for validation), plus 10 additional cases, independent of the development cohort for testing. All patients underwent a standard, non-contrast pCT scan followed by an MRI acquired on a 3T Siemens MR simulator. Images were deformably-registered, yielding 10,000 matched CT-MRI slice-pairs for model training. The proposed model utilizes a two-stage architecture, a Pix2Pix conditional generative adversarial network (GAN) for coarse sMRI synthesis, followed by a conditional diffusion model for refinement. To test the model's robustness, 10 patient cases with both pCT and contrast-enhanced CT (ceCT) were used to generate sMRI. Model performance was assessed using structural similarity index (SSIM), multi-scale structural similarity index (MS-SSIM), mean absolute error (MAE), and peak signal-to-noise ratio (PSNR). In addition, the geometric agreement of the primary (GTVp_70) and nodal (GTVn_70) gross tumor volumes delineated on sMRI vs. gtMRI was independently assessed by two radiation oncologists (ROs) using the Dice similarity coefficient (DSC), 95th-percentile Hausdorff distance (HD95), and average symmetric surface distance (ASSD).Results
The average MS-SSIM, SSIM, MAE and PSNR for pCT-derived sMRI were 0.82 ± 0.02, 0.70 ± 0.03, 0.05 ± 0.01, and 20.42 ± 0.66 dB, respectively. The mean MS-SSIM and SSIM values for ceCT-derived sMRI were 0.02 lower than those for the corresponding pCT-to-sMRI metrics. For the contour-based evaluation, sMRI-derived GTV contours achieved a mean DSC of 0.56 and 0.60 for GTVp_70, and 0.72 and 0.74 for GTVn_70 across the two ROs, relative to gtMRI contours.Conclusion
Our proof-of-concept study demonstrates the feasibility of pCT-derived sMRI for H&N cancers using a deep-learning framework. The model produces grayscale patterns and soft-tissue contrast that closely resemble the gtMRI.
Related
- Automated Radiotherapy Planning for Head-and-neck Cancer: Multi-institutional Development, Validation, and Clinical Implementation of Monaco Scripting
- Comparison of different radiotherapy planning techniques in breast cancer using active breathing coordinator
- OpenGray: a benchmark for evaluating the decisions of radiotherapy planning agents
- Retrospective intrafraction motion verification using triggered kilovoltage images during bone metastases stereotactic body radiotherapy
- Cone-beam CT-based synthetic CT generation for adaptive radiotherapy: methodological paradigms, clinical reliability assessment, and translation
BibTeX
@article{ramachandran2026synthetic,
title = {{Synthetic MRI from radiotherapy planning CT images for Head-and-Neck Cancer.}},
author = {Prabhakar Ramachandran and Akshay C Dhargave and Catherine Jones and Dionee Liefman and Lachlan McDowell and Howard Liu},
journal = {Journal of applied clinical medical physics},
year = {2026},
month = oct,
doi = {10.1002/acm2.70849},
url = {https://europepmc.org/article/MED/42900554},
note = {CC BY},
}
RIS
TY - JOUR
TI - Synthetic MRI from radiotherapy planning CT images for Head-and-Neck Cancer.
AU - Prabhakar Ramachandran
AU - Akshay C Dhargave
AU - Catherine Jones
AU - Dionee Liefman
AU - Lachlan McDowell
AU - Howard Liu
PY - 2026
DA - 2026/10/30
JO - Journal of applied clinical medical physics
DO - 10.1002/acm2.70849
UR - https://europepmc.org/article/MED/42900554
AB - Background Computed Tomography (CT) provides geometric accuracy and electron density information essential for radiotherapy planning; however, it offers limited soft-tissue contrast. Magnetic Resonance Imaging (MRI) improves tumor visualization but introduces registration uncertainties when acquired separately from the radiotherapy planning CT (pCT). Cross-modality image synthesis enables the generation of synthetic MRI (sMRI) directly from pCT and may combine the strengths of both modalities, reducing reliance on MRI access.Purpose We propose a two-stage deep-learning framework for generating sMRI from pCT image-sets specifically in head-and-neck (H&N) cancers, providing soft-tissue contrast comparable to ground-truth MR images (gtMRI).Methods This study included 60 patient cases with H&N cancer treated with radical radiotherapy between March 2022 and April 2024 for model development (50 for training, 10 for validation), plus 10 additional cases, independent of the development cohort for testing. All patients underwent a standard, non-contrast pCT scan followed by an MRI acquired on a 3T Siemens MR simulator. Images were deformably-registered, yielding 10,000 matched CT-MRI slice-pairs for model training. The proposed model utilizes a two-stage architecture, a Pix2Pix conditional generative adversarial network (GAN) for coarse sMRI synthesis, followed by a conditional diffusion model for refinement. To test the model's robustness, 10 patient cases with both pCT and contrast-enhanced CT (ceCT) were used to generate sMRI. Model performance was assessed using structural similarity index (SSIM), multi-scale structural similarity index (MS-SSIM), mean absolute error (MAE), and peak signal-to-noise ratio (PSNR). In addition, the geometric agreement of the primary (GTVp_70) and nodal (GTVn_70) gross tumor volumes delineated on sMRI vs. gtMRI was independently assessed by two radiation oncologists (ROs) using the Dice similarity coefficient (DSC), 95th-percentile Hausdorff distance (HD95), and average symmetric surface distance (ASSD).Results The average MS-SSIM, SSIM, MAE and PSNR for pCT-derived sMRI were 0.82 ± 0.02, 0.70 ± 0.03, 0.05 ± 0.01, and 20.42 ± 0.66 dB, respectively. The mean MS-SSIM and SSIM values for ceCT-derived sMRI were 0.02 lower than those for the corresponding pCT-to-sMRI metrics. For the contour-based evaluation, sMRI-derived GTV contours achieved a mean DSC of 0.56 and 0.60 for GTVp_70, and 0.72 and 0.74 for GTVn_70 across the two ROs, relative to gtMRI contours.Conclusion Our proof-of-concept study demonstrates the feasibility of pCT-derived sMRI for H&N cancers using a deep-learning framework. The model produces grayscale patterns and soft-tissue contrast that closely resemble the gtMRI.
SN - 1526-9914
ER -