BACKGROUND AND PURPOSE: To evaluate whether style transfer using generative adversarial networks (GANs) can synthesize T2 FLAIR-like MR images from non-contrast CT to enhance lesion conspicuity of brain metastases. MATERIALS AND METHODS: This retrospective study analyzed 321 patients (235 with brain metastases, 86 without) whose non-contrast CT head and T2 FLAIR MR images were paired based on lesion stability and temporal proximity. Following pre-processing, CT-MR pairs were used to train GANs for CT-to-synthetic-FLAIR style transfer with 5-fold cross-validation. PatchGAN discriminator was used. 3 UNet-family generators (Attention-UNet, SUNet, UNet++) were compared using mean-absolute-error (MAE), mean-squared-error (MSE), and structural similarity index measure (SSIM). 20 synthetic MR images generated by the best performing generator were shown to 12 neuroradiologists, who rated image quality and lesion conspicuity for CT, real MR, and synthetic MR (scale 1-5★), and realism preference between real and synthetic MR. RESULTS: UNet++ produced the best image quality, with SSIM = 0.9119 ± 0.0008, MAE = 0.1232 ± 0.0022, and MSE = 0.0201 ± 0.0004 (95% CIs, all p4.28 ± 0.15★ for real MR and 4.10 ± 0.14★ for synthetic MR; mean lesion conspicuity ratings were 4.67 ± 0.13★ for real MR, 4.34 ± 0.18★ for synthetic MR, and 3.16 ± 0.44★ for CT (all 95% CIs). Synthetic MR significantly improved lesion conspicuity over CT (pd=2.2; while synthetic MR showed slightly reduced conspicuity compared to real MR (pd=1.6 (95% CIs). Readers selected real MR as more realistic than synthetic MR in 73.3% of comparisons (pp=0.078). CONCLUSION: GAN-generated synthetic T2 FLAIR images from CT improve edema conspicuity of brain metastases relative to CT.
Hsu et al. (Wed,) studied this question.