急性缺血性脑卒中 CT 生成 MRI 算法——基于影像组学的边缘感知扩散 GAN-《计算机技术与发展》

文章信息/Info

Title:: Acute Ischemic Stroke CT-to-MRI Conversion Algorithm——Radiomics-based Edge-aware DiffusionGAN

Author(s):: ZHANG Mei-mei; QIN Pin-le; CHAI Rui; ZENG Jian-chao; ZHAI Shuang-jiao; School of Data Science and Technology,North University of China,Taiyuan 030051,China

Keywords:: medical image synthesis; computed tomography ( CT ) ; magnetic resonance imaging ( MRI ) ; acute ischemic stroke; DiffusionGAN; Radiomics

摘要:: 相较于 CT,急性缺血性脑卒中病灶在 MRI 上可清晰显示。鉴于患者中存在一些不适用于 MRI 检查的情况,如体内金属或幽闭恐惧症等,容易耽误患者病情。通过结合影像组学和扩散生成对抗网络( DiffusionGAN) ,该文提出了基于影像组学的边缘感知扩散 GAN,实现了急性缺血性脑卒中患者的 CT 生成 MRI。该算法弥补了现有 CT 生成 MRI 中病灶信息不足和边缘模糊的问题。具体而言,通过在 CT 上应用影像组学定位病灶,借助特征图辅助 MRI 生成,并提出边缘感知扩散 GAN 以提升生成 MRI 的边缘感知。实验证实,生成的 MRI 在峰值信噪比( PSNR) 方面达到 69. 607,结构相似性(SSIM)为 0. 821,皮尔逊相关系数( PCC)为 0. 974,显著优于现有模型。医生评估显示生成的 MRI 不存在错误病灶,阳性 /阴性分类准确度高达 87. 91% ,表明这一创新性算法为解决在 MRI 不适用的特殊情况下的脑卒中诊断问题提供了新的途径。

Abstract:: Compared to CT, the lesions of acute ischemic stroke are more clearly visible on MRI. Given that some patients haveconditions unsuitable for MRI, such as the presence of?
metal in the body or claustrophobia, it can hinder the diagnosis of patients.Combined radiomics and Diffusion Generative Adversarial Network ( DiffusionGAN) ,we propose an Edge-Aware DiffusionGAN basedon radiomics,achieving the conversion of CT to MRI for patients with acute ischemic stroke. The algorithm addresses the issues ofinsufficient lesion information and blurry edges in existing CT-to-MRI conversion methods. Specifically,it employs radiomics to locatelesions on CT,assists in MRI generation with feature maps,and introduces the Edge-Aware DiffusionGAN to enhance edge perception inthe generated MRI. Experimental results demonstrate that the generated MRI achieves a peak signal-to-noise ratio ( PSNR) of 69. 607,astructural similarity index ( SSIM) of 0. 821,and a Pearson correlation coefficient ( PCC) of 0. 974,significantly outperforming existing　models. Medical evaluations reveal no erroneous lesions in the generated MRI,with an accuracy of 87. 91% in positive / negative classification,indicating that this innovative algorithm provides a new way to solve the problem of stroke diagnosis in special cases where MRI isnot applicable.