SAN-Wavelet cycleGAN for anatomical structure reinforcement and tissue detail preservation in abdominal CT synthesis

Abstract

Generating a CT image from an existing but unpaired MR image has recently become a promising objective in the radiotherapy treatment planning (RTP), whereas the unsatisfactory structure and detail accuracy of the outcome is still a challenge for its clinical application. To address this issue, this work proposed an unsupervised model called structure-adaptive normalization wavelet (SAN-Wavelet) cycleGAN for unpaired MR-to-CT image synthesis. This method not only developed a module called SAN to ensure the anatomically structural integrity, but also introduced a neighborhood detail loss (ND loss) for detail consistency penalization between different modalities. Furthermore, a high-frequency discriminator and wavelet-trans skip connection were designed to keep with the high-frequency tissue detail. Experimental findings on abdominal area demonstrated the effectiveness of SAN-Wavelet cycleGAN for unpaired MR-to-CT synthesis, with mean squared error (MSE) of 66.38, root mean squared error (RMSE) of 8.07, peak signal-to-noise ratio (PSNR) of 25.944 dB, structural similarity index (SSlM) of 0.895 and mixture perceptual similarity index (MPSIM) of 0.723. Compared to other unsupervised approaches (i.e., cycleGAN, gc-, cc-, and sc-cycleGAN), the performances of SAN-cycleGAN improved by 5–15 % in terms of the metrics above. Moreover, the dosimetric distributions of the synthesized CT and real CT in planning target volume (PTV-45Gy) and organ at risk area (bowel) were statistically consistent (Mann-Whitney U test, P > 0.05) with a mean relative error being 0.09 % and 0.24 %, showcasing a promising applicability for clinical application.