基于改进Wasserstein生成对抗网络的出血性脑卒中CT图像去噪研究
Research on denoising of hemorrhagic stroke CT images based on improved Wasserstein generative adversarial networks
符炜浩 1范应威 1唐晓英1
作者信息
- 1. 北京理工大学医学技术学院(北京 100081)
- 折叠
摘要
目的 为提高无配对参考图像的出血性脑卒中 CT 的图像质量,提出一种基于改进Wasserstein生成对抗网络(Wasserstein generative adversarial network,W-GAN)的CT图像去噪算法.方法 以W-GAN网络为框架,在生成器部分引入视觉几何组(visual geometry group,VGG)网络计算感知损失模块,并在鉴别器部分加入自注意力机制和谱归一化卷积对模型进行改进,对输入的低剂量CT数据进行去噪,得到接近标准剂量的图像.随后对无配对参考图像的出血性脑卒中数据用训练完成的模型进行迁移学习,并对最终得到的图像分别使用全变分(total variation,TV)、无参考图像空间域质量评估(blind/referenceless image spatial quality evaluator,BRISQUE)和对比语言-图像预训练模型图像质量评估(contrastive language-image pre-training image quality assessment,CLIP-IQA)3 种无参考图像质量评估方式进行评估.结果 在TV、BRISQUE和CLIP-IQA 3 种无参考图像质量评估指标上相对于输入提升分别为0.0165、0.1272、0.007.结论 本文提出的改进W-GAN网络模型可以用于出血性脑卒中低剂量CT图像去噪的迁移学习任务,并取得良好的性能提升,为辅助医师诊断出血性脑卒中提供了一种可能的工具.
Abstract
Objective To enhance the image quality of unpaired-reference hemorrhagic stroke CT images,a denoising algorithm for CT images based on an improved Wasserstein generative adversarial network(W-GAN)is proposed.Methods Using the W-GAN network as the framework,the visual geometry group(VGG)network is introduced in the generator part to calculate the perceptual loss module,and the discriminator part is improved by adding self-attention mechanisms and spectral normalization convolutions.The model is used to denoise low-dose CT data,obtaining images close to normal dose.Subsequently,transfer learning is performed on the unpaired-reference hemorrhagic stroke data using the trained model,and the final images obtained are evaluated using no-reference image quality assessment.The assessments are conducted using three no-reference image quality evaluation methods:total variation(TV),blind/referenceless image spatial quality evaluator(BRISQUE),and contrastive language-image pre-training image quality assessment(CLIP-IQA).Results The final results show improvements of 0.016 5,0.127 2,and 0.007 compared to the input on the three reference free image quality evaluation metrics of TV,BRISQUE,and CLIP-IQA,respectively.Conclusions The improved W-GAN network model proposed in this paper can be used for the transfer learning task of denoising low-dose CT images of hemorrhagic stroke,achieving good performance improvement,and providing a potential tool to assist physicians in diagnosing hemorrhagic stroke.
关键词
生成对抗网络模型/CT图像/出血性脑卒中/图像去噪/Wasserstein距离Key words
generative adversarial network model/CT image/hemorrhagic stroke/image denoising/Wasserstein distance引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据