基于DCPAN的低剂量能谱CT图像去噪方法

Low Dose Spectral Computed Tomography Image-Based Denoising Method via DCPAN

能谱式计算机断层扫描(CT)成像技术具备良好的能量分辨率,能够精确地鉴别人体组织成分,从而为后续诊断提供更准确的检测信息.随着辐射剂量的降低,能谱CT图像中噪声水平显著提高,对成像质量产生严重影响,进而降低了组织成分的解析精度.基于卷积神经网络(CNN)的去噪模型虽然可以显著降低图像中的噪声含量,但随着卷积层数的增加,深层神经网络通常会丢失高频信息.为了解决这一问题,并实现在低剂量条件下重建出高质量能谱CT图像,本文提出了一种结合通道注意力机制(CA)和持续自注意力机制(PSA)的密集连接持续注意力网络(DCPAN).两种注意力机制分别建立特征图像在通道和全局维度的联系以提高网络对图像高频分量的敏感程度,进而抑制高频细节信息的丢失.该模型所采用的密集连接结构通过特征复用的方式可以在前馈传播中保留高频信息,使用复合损失函数来监督网络的训练可以使该模型对边缘特征和组织细节信息更加敏感.实验结果表明,经该模型处理的腹部切片CT图像峰值信噪比、结构相似性指数和特征相似性指数分别达到了38.25 dB、0.9937和0.9732以上.相比于目前先进的CT噪声去除方法,该方法具有更好的噪声抑制能力,处理得到的重建图像组织结构清晰、噪声含量更低,为后续诊断和其他处理工作提供更精确的检测信息.

Spectral computed tomography(CT)has good energy resolution,which makes it more accurate to identify tissue components and provides more precise detection information for subsequent diagnosis.As low dose requirements increase,the noise in spectral CT images improves significantly,affecting imaging quality and reducing resolution accuracy.Although the denoising model based on a convolutional neural network(CNN)can reduce the noise level,with the increase of layers,the denoised images from deeper neural networks often have less high-frequency information.To solve this problem,the densely connected persistent attention network(DCPAN)is proposed in this paper,which adopts the channel attention mechanism(CA)and the persistent self-attention mechanism(PSA),establishing the correlation of feature maps in the channel and global dimensions to improve the sensitivity of the network to the high-frequency components and thus suppress the loss of high-frequency information.The adopted densely connected structure realizes the retention of high-frequency information in feedforward propagation through feature multiplexing.During the period of training,a hybrid loss function is deployed to make the model more sensitive to edge and detailed information.The experimental results show that the peak signal to noise ratio,structural similarity,and feature similarity of abdomen slices are over 38.25 dB,0.9937,and 0.9732.Compared with the current advanced methods,the method has a better denoising effect,and the tissue structure in the denoised images is clearer,giving a more accurate reference for the following diagnosis.