CT image generation from ultrashort echo time MRI for dosimetric assessment in lung cancer radiotherapy.

BACKGROUND AND PURPOSE: To facilitate magnetic resonance imaging (MRI)-only radiotherapy workflows and overcome the lack of electron density information in MRI for accurate radiotherapy dose calculation, we proposed a cycle-consistent generative adversarial network-based framework, Fast Fourier Transform Spatial-Channel Net (FFTSC-Net). This study aimed to synthesize high-quality synthetic computed tomography (sCT) from lung ultrashort echo time-MRI (UTE-MRI) and assess its potential clinical dosimetric viability.

MATERIALS AND

METHODS

UTE-MRI and CT data from 60 patients with lung cancer were retrospectively collected and spatially aligned through deformable registration. FFTSC-Net was developed by integrating spatial and channel reconstruction convolution, squeeze-and-excitation, and residual fast Fourier transform modules, alongside a dual contrastive loss function. The synthesis quality was evaluated using structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and mean absolute error (MAE). The clinical feasibility was assessed through Hounsfield unit (HU) consistency analysis in organs at risk (OARs) and planning target volume (PTV), dose volume histogram parameters, and gamma passing rates.

RESULTS

FFTSC-Net achieved an SSIM of 0.80, PSNR of 20.07 dB, and MAE (HU) of 59.31. No statistically significant differences were observed in HU distributions for lungs, spinal cord, and PTV compared with planning CT (p > 0.05). Dosimetric analysis revealed no significant differences across all PTV and OAR indices (p > 0.05). Gamma passing rates reached 95.39% (2%/2 mm) and 98.08% (3%/3 mm).

CONCLUSION

FFTSC-Net demonstrates high image fidelity and dosimetric accuracy, supporting its potential for MRI-only radiotherapy planning in patients with lung cancer.