摘要
CT图像中的金属伪影是由金属物体引起的,它们会严重降低图像质量并影响诊断。在这项工作中,本文提出了一种基于生成器-鉴别器结构的生成对抗网络(GAN)模型,同时引入线性插值方法来约束生成器提取更多关键特征以生成更可信的图像。作者还将注意力机制应用于残差网络中,以增加模型特征提取的能力。在Deeplesion数据集上验证了模型的有效性。实验结果表明,文章提出的模型能有效减少金属伪影,图像细节保真度也优于其他方法。
Metal artifacts in CT images are caused by metallic objects,and they could severely decrease image quality and affect diagnosis.In this work,we propose a generative adversarial network(GAN)model based on the generator-discriminator structure,and at the same time we introduce linear interpolation(LI)to constraint the generator to extract more key features to generate more credible images.We also apply the attention mechanism to the residual network to increase the feature extraction ability of the model.The validity of the model was verified on the Deeplesion dataset.Experimental results show that the proposed model can effectively reduce metal artifacts,and the image detail fidelity is better than other methods.
作者
嵇龙寅
李光
杨姣
Ji Longyin;Li Guang;Yang Jiao(Biological Sciences and Medical Engineering Department of Southeast University,Nanjing 210096,China;Children's Hospital Affiliated to Nanjing Medical University,Nanjing 210008,China)
出处
《信息化研究》
2022年第5期10-15,38,共7页
INFORMATIZATION RESEARCH
基金
国家自然科学基金项目(No.62001112)
江苏省重点研发计划项目(No.BE2021609)
作者简介
嵇龙寅(1998—),男,硕士研究生,主要研究方向为医学图像处理等。
