应用深度学习方法减少冠状动脉 CT 血管造影图像中的运动伪影。
Motion artefact reduction in coronary CT angiography images with a deep learning method.
机构信息
Department of Radiology, Beijing Friendship Hospital, Capital Medical University, No. 95 Yongan Road, Xicheng District, Beijing, People's Republic of China.
Philips Healthcare, Beijing, China.
出版信息
BMC Med Imaging. 2022 Oct 28;22(1):184. doi: 10.1186/s12880-022-00914-2.
BACKGROUND
The aim of this study was to investigate the ability of a pixel-to-pixel generative adversarial network (GAN) to remove motion artefacts in coronary CT angiography (CCTA) images.
METHODS
Ninety-seven patients who underwent single-cardiac-cycle multiphase CCTA were retrospectively included in the study, and raw CCTA images and SnapShot Freeze (SSF) CCTA images were acquired. The right coronary artery (RCA) was investigated because its motion artefacts are the most prominent among the artefacts of all coronary arteries. The acquired data were divided into a training dataset of 40 patients, a verification dataset of 30 patients and a test dataset of 27 patients. A pixel-to-pixel GAN was trained to generate improved CCTA images from the raw CCTA imaging data using SSF CCTA images as targets. The GAN's ability to remove motion artefacts was evaluated by the structural similarity (SSIM), Dice similarity coefficient (DSC) and circularity index. Furthermore, the image quality was visually assessed by two radiologists.
RESULTS
The circularity was significantly higher for the GAN-generated images than for the raw images of the RCA (0.82 ± 0.07 vs. 0.74 ± 0.11, p < 0.001), and there was no significant difference between the GAN-generated images and SSF images (0.82 ± 0.07 vs. 0.82 ± 0.06, p = 0.96). Furthermore, the GAN-generated images achieved the SSIM of 0.87 ± 0.06, significantly better than those of the raw images 0.83 ± 0.08 (p < 0.001). The results for the DSC showed that the overlap between the GAN-generated and SSF images was significantly higher than the overlap between the GAN-generated and raw images (0.84 ± 0.08 vs. 0.78 ± 0.11, p < 0.001). The motion artefact scores of the GAN-generated CCTA images of the pRCA and mRCA were significantly higher than those of the raw CCTA images (3 [4-3] vs 4 [5-4], p = 0.022; 3 [3-2] vs 5[5-4], p < 0.001).
CONCLUSIONS
A GAN can significantly reduce the motion artefacts in CCTA images of the middle segment of the RCA and has the potential to act as a new method to remove motion artefacts in coronary CCTA images.
背景
本研究旨在探讨像素到像素生成对抗网络（GAN）去除冠状动脉 CT 血管造影（CCTA）图像中运动伪影的能力。
方法
回顾性纳入 97 例行单心动周期多期 CCTA 的患者，采集原始 CCTA 图像和 SnapShot Freeze（SSF）CCTA 图像。选择 RCA 进行研究，因为它的运动伪影是所有冠状动脉中最明显的。采集的数据分为 40 例患者的训练数据集、30 例患者的验证数据集和 27 例患者的测试数据集。使用 SSF CCTA 图像作为目标，通过训练像素到像素 GAN 从原始 CCTA 成像数据生成改进的 CCTA 图像。通过结构相似性（SSIM）、Dice 相似系数（DSC）和圆形度指数评估 GAN 去除运动伪影的能力。此外，两位放射科医生对图像质量进行了视觉评估。
结果
RCA 的 GAN 生成图像的圆形度明显高于原始图像（0.82±0.07 与 0.74±0.11，p<0.001），且 GAN 生成图像与 SSF 图像之间无显著差异（0.82±0.07 与 0.82±0.06，p=0.96）。此外，GAN 生成图像的 SSIM 为 0.87±0.06，明显优于原始图像的 0.83±0.08（p<0.001）。DSC 结果显示，GAN 生成图像与 SSF 图像之间的重叠明显高于 GAN 生成图像与原始图像之间的重叠（0.84±0.08 与 0.78±0.11，p<0.001）。RCA 的 GAN 生成 CCTA 图像的运动伪影评分明显高于原始 CCTA 图像（3[4-3] 与 4[5-4]，p=0.022；3[3-2] 与 5[5-4]，p<0.001）。
结论
GAN 可显著减少 RCA 中段 CCTA 图像的运动伪影，有望成为一种去除冠状动脉 CCTA 图像运动伪影的新方法。
Zotero 插件