KEY POINTS
- This retrospective single-center study from Princess Alexandra Hospital, Australia, evaluated whether synthetic MRI (sMRI) could be generated directly from radiotherapy planning CT in patients with head and neck cancer. The dataset included 70 patients treated with curative-intent RT between March 2022 and April 2024, divided into 50 training, 10 validation, and 10 independent internal test cases. All patients underwent noncontrast planning CT and gadolinium-enhanced T1-weighted MRI acquired on a 3-T Siemens MR simulator.
- The proposed framework combined Pix2Pix conditional GAN-based image generation followed by diffusion-based refinement. A two-stage model first generated coarse MRI-like images from CT, then refined anatomical detail using a conditional diffusion model. Development used approximately 10,000 matched CT–MRI slice pairs, created through deformable image registration. The complete synthesis pipeline required approximately 3–5 minutes per CT volume on an NVIDIA RTX A4500 GPU.
- Image registration introduced a measurable source of uncertainty. The mean landmark-based target registration error between planning CT and MRI was 2.15 ± 1.19 mm, with a median of 1.95 mm. Registration accuracy was assessed using anatomical landmarks in the upper cervical spine and skull base. Although independent test patients were excluded from model training, the images used for evaluation had been deformably registered to the MRI reference, and performance on native, unwarped planning CT was not directly tested.
- Across the 10-patient test cohort, generated sMRI demonstrated mean multiscale structural similarity (MS-SSIM) of 0.82 ± 0.02, SSIM of 0.70 ± 0.03, normalized mean absolute error of 0.05 ± 0.01, and PSNR of 20.42 ± 0.66 dB. Diffusion refinement improved MAE by 0.85%, SSIM by 0.73%, and PSNR by 0.32% compared with the initial GAN output, with statistically significant slice-level improvements (p<0.001). However, these numerical gains were small and their clinical relevance was not independently demonstrated.
- Agreement with true MRI-based tumor contours was substantially lower than global image similarity might suggest. Two radiation oncologists independently delineated target volumes on sMRI, obtaining primary tumor GTV Dice scores of 0.56 and 0.60, corresponding to a pooled mean of 0.58 ± 0.13. For involved nodal targets, Dice scores were 0.74 and 0.72, with a pooled mean of 0.73 ± 0.08. Overall HD95 was 12.81 mm for primary GTV and 11.25 mm for nodal GTV, illustrating potentially important geometric discrepancies in target definition.
- The model was also tested using contrast-enhanced CT as an alternative input, despite being trained exclusively on noncontrast planning CT. Mean MS-SSIM decreased from 0.82 for planning CT-derived sMRI to 0.80 for contrast-enhanced CT-derived sMRI, while SSIM decreased from 0.70 to 0.68. An exploratory comparison between two CT scanner vendors demonstrated nearly identical mean MS-SSIM values, but only six Philips and four Toshiba test cases were available, preventing reliable conclusions about cross-vendor generalizability.
- The fundamental limitation is that noncontrast CT does not contain the patient-specific physiological information represented by gadolinium enhancement on MRI. Generated enhancement patterns therefore reflect learned statistical associations rather than directly measured tumor vascularity or contrast uptake. The study lacked a CT-only contouring comparator, independent multicenter validation, and assessment of downstream treatment-plan dosimetry or clinical outcomes. In particular, the moderate primary-tumor Dice score and relatively large surface discrepancies mean that MRI-like appearance cannot be assumed to provide clinically reliable target delineation.
CLINICAL TAKEAWAY
Synthetic MRI generated from planning CT may eventually provide supplementary soft-tissue information for head and neck radiotherapy planning, particularly where MRI access is limited. However, the modest agreement with actual MRI-based tumor contours, uncertainty in generated enhancement patterns, and absence of prospective dosimetric validation mean that this approach remains experimental and cannot currently replace acquired MRI for clinical target definition.