Deep learning algorithm for lung CT synthesis based on iterative registration and perceptual loss
10.3969/j.issn.1005-202X.2025.01.009
- VernacularTitle:基于迭代配准和感知损失的肺部伪CT合成深度学习算法
- Author:
Tao YANG
1
;
Miao HUANG
;
Cong LIU
;
Zhihua HU
;
Lili TAO
;
Shuping ZHANG
Author Information
1. 上海第二工业大学智能制造与控制工程学院,上海201209;上海第二工业大学计算机与信息工程学院,上海201209
- Publication Type:Journal Article
- Keywords:
cone beam computed tomography;
CycleGAN;
perceptual loss;
Elastix;
image synthesis
- From:
Chinese Journal of Medical Physics
2025;42(1):59-66
- CountryChina
- Language:Chinese
-
Abstract:
Objective To synthesize high-quality synthetic CT (sCT) images from cone beam CT (CBCT) by learning lung CT domain image features with a deep learning algorithm. Methods A sCT generation algorithm which employs perceptual loss-based cyclic generative adversarial network model (CycleGAN) and iterative registration was presented. CycleGAN model was trained to generate high-quality sCT images by combining perceptual loss and cycle consistency loss;and Elastix was used to register the generated sCT image and the planned CT (pCT) image,and iterate CycleGAN generator model. Results Experiments were conducted on the obtained pCT and CBCT data of 70 patients with lung tumors. From a quantitative perspective,the SSIM between sCT generated by the proposed algorithm and pCT was improved by 11.9％ as compared with that between CBCT and pCT,increasing from 0.825 to 0.923;additionally,RMSE dropped from 110.97 HU to 78.62 HU,PSNR increased from 32.21 dB to 34.74 dB,and mutual information increased from 1.187 to 1.418. The visual evaluation revealed that the proposed algorithm greatly eliminated the scattering artifacts of CBCT slices,highlighted the bone structure,and repaired the soft tissue structure. The comparisons with U-CycleGAN,R-CycleGAN and CUT models confirmed the effectiveness of the proposed algorithm. Conclusion Using the proposed algorithm for sCT images generation can effectively reduce the dose error and structural error between CBCT and pCT,making it possible to apply the proposed algorithm to accurate dose calculations and assist doctors in clinical diagnosis.
