摘要
跨模态图像合成是医学图像处理研究的重要任务之一。但监督模型所需的医学图像标注信息获取困难、成本昂贵,导致现有的图像合成模型不能较好地保留输入图像的结构信息。针对该问题,提出一种新的融合边缘感知信息的无监督跨模态医学图像合成方法。算法以CycleGAN为基本框架,采用改进后的U-Net为生成器网络,在跳跃连接中加入残差路径,缓解编、解码器间的语义差异。对编、解码器的扩张路径和收缩路径中相邻的卷积块,以密集连接的方式融合,增加重用特征信息,提高网络的特征表达能力。然后通过在模型中增加边缘感知模块,使得网络能同时学习到医学图像的纹理信息和边缘信息,更好地反映异常区域,方便医生区分正常和病变组织。最后,在公开脑部数据集上进行实验验证,结果表明了提出的跨模态医学图像合成方法的有效性,并将本文提出的方法应用到其他场景中,进一步验证了提出方法的泛化性能。
Cross-modal image synthesis is one of the important tasks in medical image processing. The medical image annotation information required by the supervised model is difficult and expensive to obtain. Which make the existing image synthesis models cannot well preserve the structural information of the input image. Aiming at addressing this problem, a novel unsupervised crossmodal medical image synthesis method that fuses edge perception information was proposed. The algorithm takes CycleGAN as the basic framework, adopts an improved u-net as the generator network and incorporates residual paths in the skip connection to alleviate the semantic difference between encoder and decoder. The adjacent convolution blocks in the expansion path and contraction path of the encoder and decoder are fused in a densely connected way to increase the reuse of feature information and improve the feature expression ability of the network. Then by supplementing the edge perception module to the model, the network can learn the texture information and edge information of medical images simultaneously, which can better reflect the abnormal area,and facilitate doctors to distinguish normal and diseased tissues. Finally, experimental verification was carried out on public brain dataset. The results show the effectiveness of the proposed cross-modal medical image synthesis method, and the generalization performance of the proposed method is further validated by applying it to other scenarios.
作者
侯冰震
张桂梅
龚磊
符祥
HOU Bing-zhen;ZHANG Gui-mei;GONG Lei;FU Xiang(Institute of Computer Vision,Nanchang Hangkong University,Nanchang 330063,China)
出处
《南昌航空大学学报（自然科学版）》
CAS
2022年第4期28-35,共8页
Journal of Nanchang Hangkong University(Natural Sciences)
基金
国家自然科学基金(No.61763033)。
