2465 - Unet-Attention Guided Cycle Generative Adversarial Networks Based Synthetic 4DCT from 4D-CBCT Generation for Lung Cancer Stereotactic Body Radiotherapy

X. Zhang¹, J. Li¹, W. Wei¹, G. Li², and S. Bai³; ¹West china hospital, Sichuan university, chengdu, China, ²Radiation Physics Center, State Key Laboratory of Biotherapy and Cancer Center, West China Hospital, Sichuan University, Chengdu, China, ³Division of Radiation Physics, Cancer Center, West China Hospital, Sichuan University, Chengdu, Sichuan, China

Purpose/Objective(s): Obtaining four dimensions cone-beam computed tomography (4D-CBCT) with the similar image quality as plan CT (pCT) is helpful for online adaptive radiotherapy in lung cancer patients treated with stereotactic body radiotherapy. This study presented an Unet-attention guided cycle generative adversarial networks for generating high quality synthetic 4DCT (s4DCT) from 4D-CBCT and compared with other deep learning methods to find the most reliable algorithm. Materials/

Methods: For 16 lung cancer patients, the 4DCTs were deformably registered to the 4D-CBCTs to generated the ground truths for training of fast contrastive unpaired translation (FastCUT), contrastive unpaired translation (CUT), Unet, pix2pix, cyclegan, attention guided cyclegan (AGGAN-1) and attention guided cyclegan with NCC loss (AGGAN-2). The attention mechanisms in AGGAN-1 and AGGAN-2 were conducted by pretrained Unets for generating externals. Normalized cross-correlation between 4D-CBCT and 4DCT was added in loss function in AGGAN-2. Thirteen patients were used for training, and 3 patients were used for testing. The mean absolute error (MAE), structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) were used to compare hounsfiled unit of background-removed anatomical regions in both s4DCT and ground truth. Algorithms above the median of performance and stability were identified as highly accurate and reliable. The performance of different models was compared using paired t-test.
Results: The MAEs of FastCUT, CUT, Unet, pix2pix, cyclegan, AGGAN-1 and AGGAN-2 were 108.63 ± 9.31, 103.97 ± 15.73, 69.39 ± 1.60, 64.22 ± 1.93, 61.45 ± 1.61, 58.96 ± 1.80 and 57.68 ± 1.66, respectively. The PSNRs of FastCUT, CUT, Unet, pix2pix, cyclegan, AGGAN-1 and AGGAN-2 were 116.81 ± 1.86, 118.55 ± 1.63, 124.43 ± 0.46, 125.31 ± 0.37, 110.85 ± 0.86, 118.02 ± 1.44 and 123.97 ± 0.36, respectively. The SSIMs of FastCUT, CUT, Unet, pix2pix, cyclegan, AGGAN-1 and AGGAN-2 were 0.9948 ± 0.0016, 0.9946 ± 0.0014, 0.9977 ± 0.0006, 0.9977 ± 0.0006, 0.9946 ± 0.0004, 0.9963 ± 0.0005 and 0.9976 ± 0.0004, respectively. AGGAN-2 was the most reliable algorithm, combining performance and stability.
Conclusion: Comparable quality to pCT and stable synthetic 4DCT was generated from 4D-CBCT using Unet-attention guided cyclegan with NCC loss.

Metrics	FastCUT	CUT	Unet	pix2pix	cyclegan	AGGAN-1	AGGAN-2
MAE(HU)	108.63 ± 9.31	103.97 ± 15.73	69.39 ± 1.60	64.22 ± 1.93	61.45 ± 1.61	58.96 ± 1.80	57.68 ± 1.66
PSNR(dB)	116.81±1.86	118.55 ± 1.63	124.43 ± 0.46	125.31 ± 0.37	110.85 ± 0.86	118.02 ± 1.44	123.97 ± 0.36
SSIM	0.9948 ± 0.0016	0.9946 ± 0.0014	0.9977 ± 0.0006	0.9977 ± 0.0006	0.9946 ± 0.0004	0.9963 ± 0.0005	0.9976 ± 0.0004