基于多种探测算法结果分级融合的PET探测器晶体查找表建立算法

高分辨率是现代正电子发射型断层显像仪(Positron Emission Tomography,PET)设备最重要的技术指标之一,高分辨率PET探测器通常由海量闪烁晶体组成,这使得探测器校准时,晶体查找表建立的工作量大幅增加,从而对相关自动化算法的鲁棒性提出了更高的要求。目前,晶体阵列的矫正主要依赖固定放射源对探测器晶体阵列成像得到的光子二维位置直方图。高分辨率PET探测器的晶体阵列中晶体个数增多,尺寸变小,导致二维位置直方图信噪比下降,且非线性形变更加复杂,使已有的晶体查找表建立算法都无法得到理想的自动化效率。在测试了多种前沿的晶体查找表建立算法后发现,某些算法的结果之间有很强的互补性,可以通过将多种算法结果相融合的方法得到优于单一探测算法的结果。因此提出了一种基于多种探测算法结果分级融合的晶体查找表建立算法,在实现过程中选取了三种互补性较强的晶体探测方法,分级融合其结果,大幅度降低了二维位置直方图上晶体分割的出错概率,获得了鲁棒的结果。

Automatic crystal lookup table generation for PET detector calibration by fusion of information from multiple approaches

Background: High resolution is of key importance for positron emission tomography (PET) devices. Modern PET scanners use a large number of crystals in detector blocks to improve their resolution. The extra crystals lead to a significant increase in workload of crystal identification in PET detector calibration. Hence the robustness of crystal identification algorithms becomes a challenge. Crystal identification is typically based on flood histograms obtained by irradiating the detector blocks with an annihilation photonflood source. More crystals of smaller size degrade the quality of flood histogram and generate more complex spatial distortions. Existing crystal identification methods are not robust enough to handle these cases.Purpose: Our study aims to improve the robustness and success rate of crystal identification on PET detectors of different-level quality and reduce the manual interaction in PET detectors calibration.Methods: In this study, we developed a hierarchical fusion approach to overcome the limitations of the existing methods for accurate PET detector calibration. We first implemented three complementary state-of-the-art crystal identification methods and then assessed the consistency of the three identification results based on the quality of detected crystal position and row/column label information. The results of higher consistency are considered more reliable. The hierarchical fusion approach integrated first the identification results of the highest consistency and then those of gradually lower consistencies.Results: The performance of the crystal identification method was evaluated with PET detectors of three different levels (normal, heavily distorted, and low signal-to-noise ratio (SNR)). Although the three crystal identification methods alone achieved over 95% of success rate on normal quality detectors, they only achieved around 80% of success rate on detectors of heavy distortion and low SNR. Our hierarchical fusion approach consistently improved the success rate to 99.8%, 92.2% and 94.3% for detectors of normal, heavily distorted, and low SNR.Conclusion: Compared to three individual PET detector identification methods, the fusion-based approach consistently and significantly improved the success rate of PET detector identification, especially for heavily distorted and low SNR detectors.