MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

MR-Based Synthetic-CT Generation Using Generative Adversarial Network for Head and Neck MR-Only Radiotherapy

M Qi1*, Y Li2 , A Wu1 , F Guo1 , Q Jia1 , L Zhou1 , T Song1 , (1) Department of Biomedical Engineering,Southern Medical University, Guangzhou (2) Department of Radiation Oncology, Sun Yat-sen University Cancer Center, Guangzhou, Guangdong

Presentations

(Sunday, 7/14/2019) 1:00 PM - 2:00 PM

Room: 303

Purpose: This work investigated the effects of 3 sequences of Magnetic resonance Imaging (MRI) images as input to the conditional generative adversarial network (cGAN) model on head-and-neck synthetic CT (sCT) generation.

Methods: This work collected data of 45 patients with nasopharyngeal carcinoma, including MRI images with 3 sequences (T1, T2 and Dixon) and CT images. A cGAN network including generator and discriminator was used to convert MR to sCT. Four models were created in this work, of which three models were input by 3 single-channels images (each sequence of MR images were used as inputs separately), the remaining one was input by 3-channels (T1+T2+Dixon), and the output of the four models was the ground truth CT. 30 patients were randomly selected as the training set, and the remaining 15 patients were used as the testing set. In order to verify the performance of cGAN, mean absolute error (MAE), structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) metrics were evaluated between the actual CT images and sCTs generated by the four cGAN models.

Results: Through the comparison of the predicted sCTs of the proposed four cGAN models, the model with 3-channels as input has a smallest MAE value 72.11±11.71HU, and the sequence of MAE values for the remaining 3 models is: T1< Dixon< T2. The SSIM and PSNR values of sCT predicted by 3-channels model are the largest of the four models. The comparison of actual CT and sCT randomly selected from the transverse plane and the coronal plane also indicates that the 3-channel sCT retains more image detail.

Conclusion: We proposed a cGAN model with 3-channels (T1+T2+Dixon) as input to predict sCT and achieved higher accuracy than a single MR sequence image as input. The accuracy achieved by our model is of great significance for dose calculation of sCT.

Funding Support, Disclosures, and Conflict of Interest: 1) National Key R&D Program of China (NO.2017YFC0113203); 2) National Natural Science Foundation of China (NO.81571771 and 81601577); 3) Post-doctoral Science Foundation of China (NO.2016M592510). 4) Public Welfare Research and Capacity Building Special Foundation of Guangdong, China (2015B020214002)

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email