Cone Beam CT (CBCT) Based Synthetic CT Generation Using Deep Learning Methods for Dose Calculation of Nasopharyngeal Carcinoma Radiotherapy

Abstract

Objective: To generate synthetic CT (sCT) images with high quality from CBCT and planning CT (pCT) for dose calculation by using deep learning methods. Methods: 169 NPC patients with a total of 20926 slices of CBCT and pCT images were included. In this study the CycleGAN, Pix2pix and U-Net models were used to generate the sCT images. The Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Peak Signal to Noise Ratio (PSNR), and Structural Similarity Index (SSIM) were used to quantify the accuracy of the proposed models in a testing cohort of 34 patients. Radiation dose were calculated on pCT and sCT following the same protocol. Dose distributions were evaluated for 4 patients by comparing the dose-volume-histogram (DVH) and 2D gamma index analysis. Results: The average MAE and RMSE values between sCT by three models and pCT reduced by 15.4 HU and 26.8 HU at least, while the mean PSNR and SSIM metrics between sCT by different models and pCT added by 10.6 and 0.05 at most, respectively. There were only slight differences for DVH of selected contours between different plans. The passing rates of 2D gamma index analysis under 3 mm/3% 3 mm/2%, 2 mm/3%and 2 mm/2% criteria were all higher than 95%. Conclusions: All the sCT had achieved better evaluation metrics than those of original CBCT, while the performance of CycleGAN model was proved to be best among three methods. The dosimetric agreement confirmed the HU accuracy and consistent anatomical structures of sCT by deep learning methods.

Deep learning methods for enhancing cone‐beam CT image quality toward adaptive radiation therapy: A systematic review

“…The most objective comparisons of image quality can be found in studies that compare multiple architectures or correction techniques for generating sCT images using the same datasets. 18,19,27,35,39,41,44,47,49,57,60,61,63 Where DL methods were compared to classical CBCT correction methods, Barateau et al 61 demonstrated that their GAN sCT achieved a lower MAE than DIR of the CT (82.4 ± 10.6 vs. 95.5 ± 21.2 HU), which was found to be consistent with the results in Thummerer et al (36.3 ± 6.2 vs. 44.3 ± 6.1 HU). 57 Similarly in Liang et al, 39 cycle-GAN showed improved image qual-ity metrics over DIR of the CT when a saline-adjustable phantom was used in a controlled experiment.…”

“…The greatest disparity was observed between cycle-GAN and U-Net (51.0% vs. 36.8% respectively, p = 0.37), 18,19,41 whereas a smaller difference was noted between cycle-GAN and GAN networks (58.3% vs. 51.5% respectively, p = 0.24). 18,19,39,49…”

“…For studies utilizing cycle-GAN, the mean percentage improvement in MAE for 2D approaches 19,37,39,42,45,46,[48][49][50]52 was slightly higher than 3D-patch based approaches 35,38,41 (52.08% vs. 49.83%, p = 0.41). When controlling for training set sizes within ± 4 patients, a similar trend was observed with 2D networks 37,42 still slightly outperforming patched 3D networks 35,38,41 (51.51% vs. 49.83%, p = 0.46).…”

“…Tables 3a-c show that a total of nineteen studies investigated sCT generation primarily using cycle-GAN for their translation task, 19,[35][36][37][38][39][40][41][42][43][44][45][46][47][48][49][50][51][52] compared to eight utilizing U-Net, [53][54][55][56][57][58][59][60] three implementing GANs, 15,18,61 three exploring deep-CNNs, [62][63][64] and one study utilizing a novel architecture called artifact disentanglement network (ADN). 27 Deep-CNNs maintain the input image dimensions as the feature maps flow through the network.…”

Patient‐specific synthetic magnetic resonance imaging generation from cone beam computed tomography for image guidance in liver stereotactic body radiation therapy

“…Xue et al. implemented three deep learning models, CycleGAN, Pix2pix, and UNet, for synthetic CT generation from onboard CBCT images 11 . All of the models achieved better evaluation metrics based on the DVH and 2D gamma index analysis.…”