ESTRO 2021 Abstract Book

S365

ESTRO 2021

tools are an important step towards a CBCT-guided adaptive RT workflow for photon and proton RT.

OC-0477 Self-attention Condition GAN for Synthetic CT Generation from CBCT for Head and Neck Radiotherapy S. Wu 1 , S. Liu 2 , A. Chen 3 , D. Qian 2 , Y. Lu 1 1 School of Computer Science and Engineering, Sun Yat-Sen University, Computational Medical Imaging Laboratory, Guangzhou, China; 2 Guangzhou Perception Vision Medical Technologies Co.,Ltd., R&D Department, Guangzhou, China; 3 Sun Yat-sen University Cancer Center , Department of Radiotherapy, Guangzhou, China Purpose or Objective Cone Beam CT(CBCT) is commonly used for accurate patient positioning and anatomic changes tracking during the IGRT, which performs poorly on organ segmentation and dose calculation in adaptive radiotherapy(ART) because of the strong artifacts and Hounsfield unit(HU) inaccuracy. To overcome the CBCT limitation, we proposed a self-attention Generative Adversarial Network(GAN) to generate Synthetic CT(sCT) from CBCT, improving the CT number accuracy while keeping the same anatomy. Materials and Methods 38 cases H&N patients(30 for training, 8 for testing) with planning-CT(pCT) and CBCT in the first treatment fraction were included. We proposed a self-attention condition GAN generate CBCT-based sCT, with non- local block for image artifacts denoising and feature representation enhancement. To demonstrate the model performance, a Res-Unet generation model was trained for the comparison. The mean absolute error(MAE) and peak signal-noise ratio(pSNR) between the pCTs and sCTs were calculated for evaluating synthesis performance. Results The proposed self-attention GAN is validated with registered CT and CBCT images. Figure 1 displays the typical slices of CBCT CT, sCTs generated from U-net and proposed method and their discrepancy images. Figure 1(a) and (b) are typical slices of CTand CBCT. Figure 1(c) and (d) are the CBCT-based synthetic CT images with U- net and the proposed model. Less artifacts and sharper organ boundaries were found in sCT compared to CBCT. Figure 1(d), (e) and (f) are the zoom images of CT and CBCT , sCT with U-net , sCT with the proposed model. zoom images indicate that the proposed method got less scatter artifact(red arrow) and smaller density error (blue arrow) than U-net, especially the soft tissue.

Figure 1. A typical slice of CBCT CT, sCTs generated from U-net & proposed method and their discrepancy images. Figure 2 shows the MAE and pSNR for the 8 tested patients with respect to the U-net and proposed model. Figure 2(a) and (b) and (c) are the MEAs within the body, soft tissue, and bone with different models. Figure 2(d), (e) and (f) are the MEAs within the body, soft tissue, and bone with different models. Significant differences of MAE and pSNR were observed in the whole body and soft tissue, which showed that ours sCTs could provide accurate and reliable anatomy information compared to the U-net model.