ESTRO 2021 Abstract Book

S583

ESTRO 2021

only obvious large errors need to be adjusted. A quick sanity check of the contours is sufficient to warrant clinical use. When the target volume is close to the spinal cord or heart, more attention should be paid to the automatic contour of this OAR, because a small error could possibly lead to major differences in maximum dose.

Poster discussions: Poster Discussion 5: Optimising and generating CT images

PD-0752 Synthetic CT generation from cone-beam CT using deep-learning for breast adaptive radiotherapy X. Wang 1 , W. Jian 2 , X. Zhou 2 , H. Meng 1 , Y. Chen 1 , G. Yang 1 , S. Zhang 1 , Z. Wang 2 , X. Tan 1 , Z. Dai 1 1 The Second Affiliated Hospital, Guangzhou University of Chinese Medicine, Department of Radiation Therapy, Guangzhou, China; 2 Guangzhou University of Chinese Medicine, School of Medical Information Engineering, Guangzhou, China Purpose or Objective Cone-beam computed tomography (CBCT) is widely used for patient setup during fractional radiotherapy. However, there are still two remaining challenges for CBCT-based adaptive radiotherapy treatment, including image artifacts and the Hounsfield Unit (HU) inaccuracy compared with planning CT (pCT). To this end, this study investigated the feasibility of generation of synthetic CT (sCT) from CBCT images with deep learning for CBCT-guided breast cancer adaptive radiotherapy. Materials and Methods A total of sixty-three patients receiving radiotherapy after breast-conserving surgery were retrospectively included in this study. The CBCT scanning was performed under breath-hold (BH) guided by optical surface monitoring system (OSMS) to improve the image quality. We developed a U-Net model to generate sCT from CBCT. Eighteen patients were randomly chosen as the test sets. The U-Net model was trained with the CBCT- CT image pairs through deformable registration. The mean error (ME), mean absolute error (MAE), peak signal to noise ratio (PSNR), and structural similarity index (SSIM) between sCT and pCT were calculated for image quality evaluation of the sCT. Subsequently, the original planning CT (pCT) treatment plan was transferred to sCT keeping the same parameters. The dosimetric evaluation was performed by a quick dose recalculation on sCT relying on gamma analysis. The performances of U-Net model were compared with cycle-GAN and evaluated for image similarity and voxel-based dosimetric accuracy for breast cancer clinical plans.

Figure 1. The architecture of the U-net model.

Results Experimental results of image similarity and dosimetric accuracy of synthetic CT (sCT) demonstrate that the U-Net model outperforms the cycle-GAN model. The ME, MAE, PSNR, and SSIM within the body between pCT and sCT were 12.90±9.99HU, 62.53±9.14HU, 17.74±1.74, 0.81±0.04 for U-Net, whereas 5.95±16.66HU, 73.87±7.96HU, 23.00±3.73,0.92±0.01 for cycle-GAN. The gamma pass rates under 3mm/3% and 2mm/2% criteria were 91.40±3.52% and 85.95±4.75% for U-net, 91.00±3.90% and 83.96±4.90% for cycle-GAN respectively.

Figure 2. The top row shows an axial view of CBCT, CT and sCT generated by U-net on a breast cancer patient. The second row shows the HU profile of the dashed lines in the sCT image. CBCT: Cone-beam CT; sCT: synthetic CT Conclusion The U-net model can generate sCT with high image similarity and dosimetric accuracy. The approach could be