Zhang Xin, Su Ting, Zhang Yunxin, Cui Han, Tan Yuhang, Zhu Jiongtao, Xia Dongmei, Zheng Hairong, Liang Dong, Ge Yongshuai
Research Center for Medical Artificial Intelligence, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China.
University of Chinese Academy of Sciences, Beijing, China.
Quant Imaging Med Surg. 2024 Jan 3;14(1):640-652. doi: 10.21037/qims-23-768. Epub 2024 Jan 2.
Recently, deep learning techniques have been widely used in low-dose computed tomography (LDCT) imaging applications for quickly generating high quality computed tomography (CT) images at lower radiation dose levels. The purpose of this study is to validate the reproducibility of the denoising performance of a given network that has been trained in advance across varied LDCT image datasets that are acquired from different imaging systems with different spatial resolutions.
Specifically, LDCT images with comparable noise levels but having different spatial resolutions were prepared to train the U-Net. The number of CT images used for the network training, validation and test was 2,400, 300 and 300, respectively. Afterwards, self- and cross-validations among six selected spatial resolutions (62.5, 125, 250, 375, 500, 625 µm) were studied and compared side by side. The residual variance, peak signal to noise ratio (PSNR), normalized root mean square error (NRMSE) and structural similarity (SSIM) were measured and compared. In addition, network retraining on a small number of image set was performed to fine tune the performance of transfer learning among LDCT tasks with varied spatial resolutions.
Results demonstrated that the U-Net trained upon LDCT images having a certain spatial resolution can effectively reduce the noise of the other LDCT images having different spatial resolutions. Regardless, results showed that image artifacts would be generated during the above cross validations. For instance, noticeable residual artifacts were presented at the margin and central areas of the object as the resolution inconsistency increased. The retraining results showed that the artifacts caused by the resolution mismatch can be greatly reduced by utilizing about only 20% of the original training data size. This quantitative improvement led to a reduction in the NRMSE from 0.1898 to 0.1263 and an increase in the SSIM from 0.7558 to 0.8036.
In conclusion, artifacts would be generated when transferring the U-Net to a LDCT denoising task with different spatial resolution. To maintain the denoising performance, it is recommended to retrain the U-Net with a small amount of datasets having the same target spatial resolution.
最近，深度学习技术已广泛应用于低剂量计算机断层扫描（LDCT）成像应用中，以在较低辐射剂量水平下快速生成高质量的计算机断层扫描（CT）图像。本研究的目的是验证一个预先训练好的给定网络在从具有不同空间分辨率的不同成像系统获取的各种LDCT图像数据集上的去噪性能的可重复性。
具体而言，准备了具有可比噪声水平但具有不同空间分辨率的LDCT图像来训练U-Net。用于网络训练、验证和测试的CT图像数量分别为2400、300和300。之后，研究了六种选定空间分辨率（62.5、125、250、375、500、625µm）之间的自验证和交叉验证，并进行了并排比较。测量并比较了残余方差、峰值信噪比（PSNR）、归一化均方根误差（NRMSE）和结构相似性（SSIM）。此外，对少量图像集进行网络再训练，以微调具有不同空间分辨率的LDCT任务之间的迁移学习性能。
结果表明，在具有一定空间分辨率的LDCT图像上训练的U-Net可以有效降低其他具有不同空间分辨率的LDCT图像的噪声。无论如何，结果表明在上述交叉验证期间会产生图像伪影。例如，随着分辨率不一致性增加，在物体的边缘和中心区域出现明显的残余伪影。再训练结果表明，通过仅使用约20%的原始训练数据大小，可以大大减少由分辨率不匹配引起的伪影。这种定量改进导致NRMSE从0.1898降至0.1263，SSIM从0.7558增至0.8036。
总之，将U-Net转移到具有不同空间分辨率的LDCT去噪任务时会产生伪影。为保持去噪性能，建议使用少量具有相同目标空间分辨率的数据集对U-Net进行再训练。
