A Taguchi-optimized Pix2pix generative adversarial network for internal dosimetry in 18F-FDG PET/CT

Fast patient-specific internal dosimetry using a voxel-based Monte Carlo (MC) simulation is challenging in personalized medicine. While the MC technique is considered the gold standard for dosimetric calculation, it suffers from lengthy execution times. In this study, we exploited the original image-translation generative adversarial network (pix-to-pix GAN) to address these limitations. The training dataset consists of fused PET/CT images of the patients (domain A) paired with corresponding MC-calculated dose maps (domain B) as the ground-truth. The performance of the network was further improved using the Taguchi optimization method. A total of 27 Taguchi experimental designs were performed to search for optimal hyperparameters. The final choice was based on the corresponding signal-to-noise ratio (SNR) and mean square error (MSE) of each run. The images generated by the optimal p2p-U-Net-GAN arrangement were then compared with the MC target ones using the structural similarity index (SSIM), peak SNR (PSNR), and MSE evaluation metrics for 350 test slices and varied between 0.91 and 0.99, 32-38, and 1.5-3.6, respectively. The findings show that all evaluation metrics are in their appropriate ranges indicating the encouraging performance of the p2p-GAN architecture. The reproduced dose distribution using p2p-U-Net-GAN exhibits good agreement with those of the MC simulation as the reference while is much less time-demanding. In conclusion, the p2p-U-Net-GAN dose calculator offers not only comparable accuracy with the direct MC simulation but also a much faster dose calculation speed. Moreover, Taguchi can be the method of choice for optimizing hyperparameters of deep networks in internal dosimetry tasks.