CHO数字观察器在PET图像滤波方法评估中的应用

在临床应用中需要限制扫描时间和药物剂量,这往往会使正电子发射断层扫描(PET)的图像的分辨率变低,噪声变多。为提供可供临床诊断的图像,去噪是一个必须的手段,而在重建后增加一个滤波器是目前最常用的去噪方法。因此对不同滤波器滤波效果的比较是PET图像重建中的重要环节,其中最关键的是滤波参数的选取。目前采用的信噪比(SNR)以及恢复系数(RC)等评估方法可以用来非定量地选取参数,研究者们只能凭经验选取最优参数。而通道化霍特林观察器(CHO)作为一个比较通用的数字观察器,已被用于与PET图像质量相关的各种参数的选择,如重建算法参数、系统设计参数、临床协议参数等,然而其在评估不同滤波方法对图像重建质量的影响中的应用研究还比较少。通过比较CHO计算得到的ROC(receiver operating characteristic)曲线下面积(area under the ROC curve,AUC),选择两种常用的滤波器(即高斯滤波器和非局部均值(Non-Local Mean, NLM)滤波器)的最优参数,并评估它们在PET中的滤波效果。结果表明,对于13 mm球体,σ为1.1~1.4的高斯滤波器和f为0.5~0.9的NLM滤波器可以达到最大的检测能力值,而对于10 mm球体,σ为1.4~2.0的高斯滤波器和f为0.5~0.9的NLM滤波器可以达到最大的检测能力值。虽然两个滤波器所对应的AUC值都能高达0.9,但是NLM滤波器的AUC值高于高斯滤波器。通过IEC图像和病人图像也能发现,NLM滤波后的PET图像中的亮点比高斯滤波的更加清晰,噪声更少。该结论和传统滤波器评估方法得到的结论一致,这说明在PET的病灶检测任务中,CHO能够准确地比较这两种滤波器的性能。

Application of Numerical Observer CHO in Evaluation of Filtering Method in PET

In clinical applications, it is necessary to limit the scan time and dose, which tends to lower the resolution of the positron emission tomography (PET) image and increase the noisein PET. A denoise method is required to achieve the clinically acceptable images, and a post filter after reconstruction is the most widely used method. Therefore, the comparison of smoothing effect of different filters, for instance the selection of filter parameters, is an important step in PET image reconstruction. Generally, the signal-to-noise ratio (SNR), recovery coefficient (RC) or similar methods are used in the parameter selections. But researchers still rely on their experience since those methods cannot be used quantitatively. As a generalized numerical observer, channelized hotelling observer (CHO) has been used in selection of various parameters in PET, such as reconstruction algorithm parameters, system design parameters, clinical protocol parameters and so on. However its application in the assessment of different filtering methods of image reconstruction is not widely studied. The purpose of this paper is to select the optimal parameters of two widely used filters, i.e. Gauss filter and Non-Local Mean (NLM) filter, and evaluate their smoothing effect in PET by comparing the area under the receiver operating characteristic(ROC)curve (AUC) values calculated by CHO. Experimental results showed that for the 13 mm sphere, Gauss filter with σ of1.1~1.4 and NLM filter with f of 0.5-0.9 achieved the maximum detectability, and for the 10 mm sphere, Gauss filter with σ of 1.4~2.0 and NLM filter with f of 0.5~0.9 achieved the maximum detectability. Though AUC values of both filters were as high as 0.9, the AUC value of NLM filter was larger than that of Gauss filter. It was also found out that bright spots had better contrast and lower noise in IEC images and patient images with the NLM filter than that with the Gauss filter. This conclusion was consistent with results obtained by traditional evaluation methods of the filter, which indicated that CHO accurately compared the performance of these two filters in the lesion detection task of PET.