MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Synthetic CT Generation Using Structure-Aware Generative Adversarial Networks (SA-GANs) for Pelvis Magnetic Resonance-Only Radiotherapy

H Emami1*, M Dong2, S Nejad-Davarani3, C Glide-Hurst4, (1) Wayne State Univ, Detroit, MI, (2) ,,,(3) ,Ypsilanti, MI, (4) Henry Ford Health System, Detroit, MI

Presentations

(Sunday, 7/12/2020) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Room: Track 2

Purpose:
While MR-only treatment planning using synthetic CTs (synCTs) offers potential for streamlining workflow, varying bladder/rectal status introduce challenges in pelvis synCT generation. This work introduces a novel and efficient end-to-end method of pelvis synCT and contour generation using a structure-aware generative adversarial network (SA-GAN) to facilitate real-time MR-only planning.

Methods:
T2-Weighted MRI and CT-SIM images from fifteen prostate cancer patients were evaluated. A novel GAN generator was implemented in a modified ResNet combined with segmentation network to segment shape-changing organs. The modified residual network generator includes several skip connections from the input MRI to the next layers for organ preservation. A CNN discriminator classified the input image as real/synthetic. To evaluate performance, 5 repeat leave-one-out cross-validations were performed. Mean absolute error (MAE), Dice similarity coefficient (DSC), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR) metrics between synCT and CT were computed for SA-GAN and compared to GAN, CycleGAN, and CNN. SA-GAN organ segmentations were compared to MRI delineations using DSC.

Results:
SA-GAN generated synCTs and organ segmentations in ~12 seconds. SA-GAN MAEs between synCT and CT-SIM were 49.61±4.37, 478.84±70.48, and 277.70±62.56 HU across the field-of-view, bone, and rectal gas. GAN had higher MAEs of 53.04±6.60 HU, 504.03 ±91.04 HU and 644.24±95.20 HU across the field-of-view, bone, and rectal gas, respectively. Similarly, CycleGAN yielded an MAE of 66.50±5.23 HU, 618.65 ±49.76 HU and 699.60±62.21 HU across the field-of-view, bone, and rectal gas. For SA-GAN, mean PSNR was 29.04±1.09 and SSIM was 0.92±0.008. SA-GAN synCTs preserved bladder, rectum, and rectal gas status with DSCs of 0.94, 0.80, and 0.91, respectively and maintained better details than CNN.

Conclusion:
We developed a novel SA-GAN synCT model using a single MRI input that rapidly generates high quality synCTs preserving air and organ shapes in seconds, offering potential for real-time MR-only planning.

Funding Support, Disclosures, and Conflict of Interest: The submitting institution holds research agreements with Philips Healthcare, ViewRay, Inc., and Modus Medical. Research partially supported by the NCI/NIH, Award R01CA204189. The PI is on the Philips Healthcare Advisory Board.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email