基于新型多通道多路径条件生成对抗网络的多参数 MRI 伪 CT 生成用于鼻咽癌患者。

Pseudo-CT generation from multi-parametric MRI using a novel multi-channel multi-path conditional generative adversarial network for nasopharyngeal carcinoma patients.

机构信息

The Hong Kong Polytechnic University, Hong Kong SAR, China.

Nanjing University, Nanjing, China.

出版信息

Med Phys. 2020 Apr;47(4):1750-1762. doi: 10.1002/mp.14062. Epub 2020 Feb 21.

DOI:10.1002/mp.14062

PMID:32012292

Abstract

PURPOSE

To develop and evaluate a novel method for pseudo-CT generation from multi-parametric MR images using multi-channel multi-path generative adversarial network (MCMP-GAN).

METHODS

Pre- and post-contrast T1-weighted (T1-w), T2-weighted (T2-w) MRI, and treatment planning CT images of 32 nasopharyngeal carcinoma (NPC) patients were employed to train a pixel-to-pixel MCMP-GAN. The network was developed based on a 5-level Residual U-Net (ResU-Net) with the channel-based independent feature extraction network to generate pseudo-CT images from multi-parametric MR images. The discriminator with five convolutional layers was added to distinguish between the real CT and pseudo-CT images, improving the nonlinearity and prediction accuracy of the model. Eightfold cross validation was implemented to validate the proposed MCMP-GAN. The pseudo-CT images were evaluated against the corresponding planning CT images based on mean absolute error (MAE), peak signal-to-noise ratio (PSNR), Dice similarity coefficient (DSC), and Structural similarity index (SSIM). Similar comparisons were also performed against the multi-channel single-path GAN (MCSP-GAN), the single-channel single-path GAN (SCSP-GAN).

RESULTS

It took approximately 20 h to train the MCMP-GAN model on a Quadro P6000, and less than 10 s to generate all pseudo-CT images for the subjects in the test set. The average head MAE between pseudo-CT and planning CT was 75.7 ± 14.6 Hounsfield Units (HU) for MCMP-GAN, significantly (P-values < 0.05) lower than that for MCSP-GAN (79.2 ± 13.0 HU) and SCSP-GAN (85.8 ± 14.3 HU). For bone only, the MCMP-GAN yielded a smaller mean MAE (194.6 ± 38.9 HU) than MCSP-GAN (203.7 ± 33.1 HU), SCSP-GAN (227.0 ± 36.7 HU). The average PSNR of MCMP-GAN (29.1 ± 1.6) was found to be higher than that of MCSP-GAN (28.8 ± 1.2) and SCSP-GAN (28.2 ± 1.3). In terms of metrics for image similarity, MCMP-GAN achieved the highest SSIM (0.92 ± 0.02) but did not show significantly improved bone DSC results in comparison with MCSP-GAN.

CONCLUSIONS

We developed a novel multi-channel GAN approach for generating pseudo-CT from multi-parametric MR images. Our preliminary results in NPC patients showed that the MCMP-GAN method performed apparently superior to the U-Net-GAN and SCSP-GAN, and slightly better than MCSP-GAN.

摘要

目的

使用多通道多路径生成对抗网络（MCMP-GAN）从多参数磁共振图像中开发和评估生成伪 CT 的新方法。

方法

使用 32 名鼻咽癌（NPC）患者的对比前和对比后 T1 加权（T1-w）、T2 加权（T2-w）MRI 和治疗计划 CT 图像来训练像素到像素的 MCMP-GAN。该网络基于具有基于通道的独立特征提取网络的 5 级残差 U-Net（ResU-Net）构建，用于从多参数 MR 图像生成伪 CT 图像。添加了具有五个卷积层的鉴别器来区分真实 CT 和伪 CT 图像，从而提高了模型的非线性和预测准确性。实施了 8 倍交叉验证来验证所提出的 MCMP-GAN。根据平均绝对误差（MAE）、峰值信噪比（PSNR）、Dice 相似系数（DSC）和结构相似性指数（SSIM），将伪 CT 图像与相应的计划 CT 图像进行评估。还对多通道单路径 GAN（MCSP-GAN）和单通道单路径 GAN（SCSP-GAN）进行了类似的比较。

结果

在 Quadro P6000 上训练 MCMP-GAN 模型大约需要 20 小时，在测试集中为受试者生成所有伪 CT 图像不到 10 秒。MCMP-GAN 中伪 CT 和计划 CT 之间的头部平均 MAE 为 75.7±14.6 亨氏单位（HU），明显（P 值<0.05）低于 MCSP-GAN（79.2±13.0 HU）和 SCSP-GAN（85.8±14.3 HU）。对于仅骨骼，MCMP-GAN 的平均 MAE（194.6±38.9 HU）小于 MCSP-GAN（203.7±33.1 HU）和 SCSP-GAN（227.0±36.7 HU）。MCMP-GAN 的平均 PSNR（29.1±1.6）高于 MCSP-GAN（28.8±1.2）和 SCSP-GAN（28.2±1.3）。在图像相似性指标方面，MCMP-GAN 达到了最高的 SSIM（0.92±0.02），但与 MCSP-GAN 相比，在骨 DSC 结果方面并没有明显改善。