Pseudo-CT generation from multi-parametric MRI using a novel multi-channel multi-path conditional generative adversarial network for nasopharyngeal carcinoma patients.

Abstract

To develop and evaluate a novel method for pseudo-CT generation from multi-parametric MR images using multi-channel multi-path generative adversarial network (MCMP-GAN). Pre- and post-contrast T1-weighted (T1-w), T2-weighted (T2-w) MRI, and treatment planning CT images of 32 nasopharyngeal carcinoma (NPC) patients were employed to train a pixel-to-pixel MCMP-GAN. The network was developed based on a 5-level Residual U-Net (ResU-Net) with the channel-based independent feature extraction network to generate pseudo-CT images from multi-parametric MR images. The discriminator with five convolutional layers was added to distinguish between the real CT and pseudo-CT images, improving the nonlinearity and prediction accuracy of the model. Eightfold cross validation was implemented to validate the proposed MCMP-GAN. The pseudo-CT images were evaluated against the corresponding planning CT images based on mean absolute error (MAE), peak signal-to-noise ratio (PSNR), Dice similarity coefficient (DSC), and Structural similarity index (SSIM). Similar comparisons were also performed against the multi-channel single-path GAN (MCSP-GAN), the single-channel single-path GAN (SCSP-GAN). It took approximately 20h to train the MCMP-GAN model on a Quadro P6000, and less than 10s to generate all pseudo-CT images for the subjects in the test set. The average head MAE between pseudo-CT and planning CT was 75.7±14.6Hounsfield Units (HU) for MCMP-GAN, significantly (P-values<0.05) lower than that for MCSP-GAN (79.2±13.0HU) and SCSP-GAN (85.8±14.3HU). For bone only, the MCMP-GAN yielded a smaller mean MAE (194.6±38.9HU) than MCSP-GAN (203.7±33.1HU), SCSP-GAN (227.0±36.7HU). The average PSNR of MCMP-GAN (29.1±1.6) was found to be higher than that of MCSP-GAN (28.8±1.2) and SCSP-GAN (28.2±1.3). In terms of metrics for image similarity, MCMP-GAN achieved the highest SSIM (0.92±0.02) but did not show significantly improved bone DSC results in comparison with MCSP-GAN. We developed a novel multi-channel GAN approach for generating pseudo-CT from multi-parametric MR images. Our preliminary results in NPC patients showed that the MCMP-GAN method performed apparently superior to the U-Net-GAN and SCSP-GAN, and slightly better than MCSP-GAN.