Cross-modality image-to-image translation from MR to synthetic F-FDOPA PET/MR fusion images using conditional GAN in brain cancer.

OBJECTIVE

This study aims to identify the possibility of cross-modality image-to-image translation from magnetic resonance (MR) to synthetic positron emission tomography (PET)/MR fusion images using conditional generative adversarial networks (CGAN).

METHODS

Retrospective study was conducted involving 32 simultaneous 6-[F]-fluoro-L-3,4-dihydroxyphenylalanine (F-FDOPA) PET/MR imaging examinations from 27 patients diagnosed with brain cancer. We applied paired axial T1-weighted contrast MR (T1C) and PET/T1C fusion images to translate from T1C to synthetic PET/T1C fusion images using the Pix2Pix algorithm of CGAN. To access the image similarity between real and synthetic PET/T1C fusion images, we calculated correlation coefficients for the maximum/mean tumor-to-background ratio (TBR) and quantitative analyses were performed using peak signal-to-noise ratio (PSNR), mean squared error (MSE), structural similarity index (SSIM), and feature similarity index measure (FSIM).

RESULTS

Total 2167 pairs of T1C and PET/T1C fusion images were obtained, which were randomly assigned to training and test datasets in 9:1 ratio (1950 and 217 pairs), and training data were further divided into training and validation datasets in 4:1 ratio (1560 and 390 pairs). The correlation coefficients were 0.706 (CI:0.533-0.822) for TBR (p < 0.001) and 0.901 (CI:0.831-0.943) for TBR (p < 0.001). The quantitative analyses were PSNR of 31.075 ± 3.976, MSE of 0.001 ± 0.001, SSIM of 0.868 ± 0.079, and FSIM of 0.922 ± 0.044, respectively.

CONCLUSION

CGAN based on simultaneous F-FDOPA PET/MR imaging data demonstrated the potential for cross-modality image-to-image translation from T1C to PET/T1C fusion images, though limitations in small dataset and lack of external validation requiring further research.