Low-dose CT image denoising based on edge prior and high-frequency sensitive feature fusion network
Low-dose CT image denoising based on edge prior and high-frequency sensitive feature fusion network
复制标题
DOI：
10.1007/s11760-023-02560-9
复制
发表时间：
2023-05
期刊：
影响因子：
--
通讯作者：
Xueying Cui;Yingting Guo;Wenqiang Hao;H. Shangguan;Xiong Zhang;B. Liu;Anhong Wang;Lizhong Jin
中科院分区：
文献类型：
--
作者：
Xueying Cui;Yingting Guo;Wenqiang Hao;H. Shangguan;Xiong Zhang;B. Liu;Anhong Wang;Lizhong Jin
低剂量CT（LDCT）是降低患者辐射剂量的可行方法。然而，伪影和噪声都出现在重建图像中，降低了图像的清晰度。为了去除伪影和噪声，提高图像质量，提出了一种基于边缘先验和高频敏感融合网络（EHSFN）的深度学习方法。EHSFN首先将LDCT图像分解为高频（HF）和低频（LF）部分，用于将伪影噪声与LF部分解耦。然后引入可训练的边缘先验，丰富组织边缘信息，指导高频部分组织结构特征的提取。为了充分提取和融合边缘和高频信息，设计了多级特征提取模块和多级融合模块。此外，边缘损失，高频损失和重建损失被用来指导网络的训练。实验结果表明，与现有的网络相比，该网络不仅能有效地去除噪声和伪影，而且能保留更多的纹理和组织细节。
Low-dose CT (LDCT) is a feasible method to reduce the radiation dose to the patient. However, both artifacts and noise appear in the reconstructed images, reducing the clarity of the images. To remove artifacts and noise and improve image quality, a deep learning method based on edge prior and high-frequency sensitive fusion network (EHSFN) is proposed. The EHSFN first decomposes the LDCT image into high-frequency (HF) and low-frequency (LF) parts for decoupling the artifact noise from the LF part. And then a trainable edge prior is introduced to enrich tissue edge information and guide the extraction of tissue structure feature in HF part. To fully extract and fuse edge and HF information, a multi-stage feature extraction module and a multi-level fusion module are designed. Besides, edge loss, HF loss and reconstruction loss are applied to guide the training of the network. Experimental results show that the proposed network can not only effectively remove noise and artifacts but also retain more texture and tissue details, compared with the existing network.