KEY POINTS
- The study included 60 patients with lung cancer who underwent planning CT and respiratory-gated ultrashort-echo-time MRI on the same day. Data were split 60:20:20% into training, validation and testing cohorts.
- Investigators developed FFTSC-Net, a CycleGAN-derived deep-learning framework integrating spatial/channel reconstruction, attention and fast-Fourier-transform modules. It was compared with Cytran, Pix2Pix and Utrans on the same test dataset.
- FFTSC-Net achieved the best image-generation performance, with SSIM 0.80, PSNR 20.07 dB and mean absolute error 59.31 HU. Lung, spinal-cord and PTV HU values did not differ significantly from planning CT.
- The heart was the notable exception, showing a significant HU discrepancy. The authors attribute this partly to cardiac motion because the UTE acquisition was respiratory-gated but not ECG-gated.
- Clinical IMRT plans prescribed 60 Gy in 30 fractions and were optimized on planning CT before identical beam geometry, fluence and monitor units were transferred to the synthetic CT for dose recalculation.
- None of the evaluated PTV or OAR dose metrics differed significantly between synthetic CT and planning CT. Mean relative differences were only −0.20% for PTV D95, −0.05% for D98, and approximately ≤0.6% for most OAR metrics.
- Mean gamma passing rates reached 95.39% at 2%/2 mm, 97.30% at 3%/2 mm and 98.08% at 3%/3 mm. External validation, larger datasets and better handling of cardiac motion are still required before an MRI-only thoracic workflow can be considered clinically established.
CLINICAL TAKEAWAY
Generating electron-density information from MRI remains one of the central barriers to MRI-only lung RT. This UTE-based model produced encouraging dosimetric agreement with planning CT, but the work is still a technical proof of feasibility rather than evidence that CT can yet be removed from routine thoracic planning.