多断层融合的肺CT肿瘤靶区超分辨率重建

Super-resolution reconstruction of pulmonary nodules based on CT multi-section fusion

针对由曝光不均、噪声等因素引起的病灶区CT数据漏检、边界模糊等问题,设计了一种多方向神经网络(NN)插值算法。通过融合各断层层内和层间信息,对病灶区进行精确超分辨率重建。首先,将预测网络拓展为多方向三维空间;然后,根据肿瘤特殊灰度分布特征,设计最优初始权值;最后,预测漏检数据,提高病灶区分辨率。将本文算法与当前具有代表性的3种超分辨率重建算法PCGLS法、180°线性插值、单方向神经网络方法进行比较,结果表明:本文方法实时性更好,迭代次数平均减少25.9%,重建图像病灶区定位更精确,空间分辨率更高,质心偏离度平均降低27.1%,中心偏离度平均降低23.0%,病灶面积平均减少21.5%,平均PSNR提高了1.59 dB。本算法不但适用于肺部CT图像,也可以根据具体图像特征推广到其他生物信号和遥感图像等领域中。

An interpolation algorithm based on multi-direction Neural Networks（NN） is presented to solve the problems on lost data and fuzzy boundary in CT images caused by the unevenness exposure and noise.The information in every section and between different sections is integrated for the super-resolution reconstruction of focal zones.Firstly,a forecast net is extended to a 3D space,then optimal initial weights are designed according to the special gray feature distribution of pulmonary nodules.Finally,lost data are forecasted to improve the resolution.The results of simulation experiments indicate that this approach can improve performance in several respects such as location,real-time and PSNRs as compared with the present representative three methods,PCGLS,180° linear interpolation and one-way neural network.It is shown that the deviations of centre and centroid are averagely reduced by 27.1% and 23.0% respectively,and the target area and the iterations are averagely reduced by 21.5% and 25.9%,respectively.Moreover,the average PSNR is increased by 1.59 dB.The proposed method can be used in not only pulmonary CT images but also biological and remote sensing images.