稀疏Transformer联合细节还原网络的CT金属伪影去除算法

CT Metal Artifact Reduction Based on Sparse Transformer Combined with Detail Restoration Network

摘要:

在CT成像中，若病人体内存在金属植入物时，CT重建图像中会出现严重的金属伪影，降低图像质量并影响医生诊断结果。对现有金属伪影去除算法进行了研究，分析了简单的神经网络模型去除金属伪影时存在伪影残留及组织细节模糊的问题，提出了稀疏Transform er联合细节还原网络。该网络由伪影去除网络和细节还原网络两个独立的子网络构成，伪影去除网络将标准Transformer中的自注意力替换为稀疏注意力机制，并且引入混合尺度前馈网络提取多尺度信息，以产生更好地图像去噪特征；细节还原网络同时提取全局和局部信息，在不损失图像分辨率的前提下清晰地恢复原始图像细节，然后通过加法运算将其整合到上述去除伪影的图像中。在Deeplesion数据集上验证了模型的有效性，实验结果表明，该方法在金属伪影去除效果上优于目前已有方法，在PSNR、SSIM指标上表现更优，能有效去除金属伪影，恢复出大部分精细的结构细节。

Abstract:

In CT imaging, if metal implants are present in the patient"s body, serious metal artifacts will appear in the CT reconstructed images, which lower the image quality and have an impact on the doctor"s diagnosis. A study was conducted on the existing metal artifact reduction methods, analyzing the problems of artifact residue and fuzzy organizational details when reducing metal artifacts in a simple neural network model. A combined sparse Transformer and detail restoration network is proposed, which include two separate sub-networks. Specially, the two separate sub-networks are the artifact reduction network and the detail restoration network. The artifact reduction network replaces the self-attention mechanism in standard Transformer with sparse attention mechanism, and to generate better image denoising features, it utilizes a mixed-scale feed-forward network to extract multi-scale information. Additionally, the detail restoration network extracts both global and local information to clearly recovers the ori