基于卷积神经网络的能谱CT射束硬化伪影抑制方法

Beam Hardening Artifacts Reduction in Spectral CT Based on Convolutional Neural Network

X射线能谱CT探测器可以通过一次扫描,对两个不同的能量区间进行图像重建,由于重建算法基于单色X射线源的假设,低能量重建图像会存在明显的射束硬化伪影.为此提出了一种用于校正射束硬化伪影的卷积神经网络结构,使用大规模的体模样本进行训练:利用低能和高能重建图像共同作为网络输入以提取伪影特征,通过最小化输出和无伪影图像之间的均方误差来对低能图像进行校正.在两种不同能谱组合中的测试结果证明了该方法可以在保留物质结构特征的基础上抑制射束硬化伪影,校正后低能图像的峰值信噪比大约提高了20 dB.

The X-ray spectral computed tomography(CT) detector can finish the reconstruction task of two different energy bins by one scan. Since the reconstruction algorithm based on the assumption that Xray source is monochromatic, there will be severe beam hardening artifacts in the low-energy reconstruction image. An architecture of convolutional neural network(CNN) for correcting hardening artifacts is proposed, which is trained using large-scale phantom samples. The low energy and high energy reconstruction images are collectively used as a network input to extract artifact features, and the low energy image is corrected by minimizing the mean square error(MSE) between the output and the artifact free image. The test results with two different spectrum combinations prove that the method can suppress the beam hardening artifact while preserving the structural features of materials. The peak signal to noise ratio(PSNR) value of the corrected image reconstructed by low energy bin is increased by about 20 dB on average.