Aristophanous Michalis, Penney Bill C, Martel Mary K, Pelizzari Charles A
Department of Radiation and Cellular Oncology, The University of Chicago, Chicago, Illinois 60637, USA.
Med Phys. 2007 Nov;34(11):4223-35. doi: 10.1118/1.2791035.
The increased interest in 18F-fluorodeoxyglucose (FDG) positron emission tomography (PET) in radiation treatment planning in the past five years necessitated the independent and accurate segmentation of gross tumor volume (GTV) from FDG-PET scans. In some studies the radiation oncologist contours the GTV based on a computed tomography scan, while incorporating pertinent data from the PET images. Alternatively, a simple threshold, typically 40% of the maximum intensity, has been employed to differentiate tumor from normal tissue, while other researchers have developed algorithms to aid the PET based GTV definition. None of these methods, however, results in reliable PET tumor segmentation that can be used for more sophisticated treatment plans. For this reason, we developed a Gaussian mixture model (GMM) based segmentation technique on selected PET tumor regions from non-small cell lung cancer patients. The purpose of this study was to investigate the feasibility of using a GMM-based tumor volume definition in a robust, reliable and reproducible way. A GMM relies on the idea that any distribution, in our case a distribution of image intensities, can be expressed as a mixture of Gaussian densities representing different classes. According to our implementation, each class belongs to one of three regions in the image; the background (B), the uncertain (U) and the target (T), and from these regions we can obtain the tumor volume. User interaction in the implementation is required, but is limited to the initialization of the model parameters and the selection of an "analysis region" to which the modeling is restricted. The segmentation was developed on three and tested on another four clinical cases to ensure robustness against differences observed in the clinic. It also compared favorably with thresholding at 40% of the maximum intensity and a threshold determination function based on tumor to background image intensities proposed in a recent paper. The parts of the method that are user dependent were evaluated and resulted in initial estimates of the method's precision, which is in the order of +/-10% of the average tumor volume estimate. With this work we have established the applicability of the GMM-based segmentation on clinical studies and we have made an initial assessment of the method's precision with respect to tumor volume segmentation.
在过去五年中，人们对18F-氟脱氧葡萄糖（FDG）正电子发射断层扫描（PET）在放射治疗计划中的兴趣日益增加，这就需要从FDG-PET扫描中独立且准确地分割大体肿瘤体积（GTV）。在一些研究中，放射肿瘤学家基于计算机断层扫描勾勒出GTV轮廓，同时纳入PET图像的相关数据。或者，通常采用一个简单的阈值，即最大强度的40%，来区分肿瘤与正常组织，而其他研究人员则开发了算法来辅助基于PET的GTV定义。然而，这些方法都无法实现可靠的PET肿瘤分割，而这种分割可用于更复杂的治疗计划。因此，我们针对非小细胞肺癌患者选定的PET肿瘤区域开发了一种基于高斯混合模型（GMM）的分割技术。本研究的目的是以稳健、可靠且可重复的方式研究使用基于GMM的肿瘤体积定义的可行性。高斯混合模型基于这样一种理念，即任何分布，在我们的案例中是图像强度分布，可以表示为代表不同类别的高斯密度的混合。根据我们的实现方式，每个类别属于图像中的三个区域之一：背景（B）、不确定区域（U）和目标（T），从这些区域我们可以获得肿瘤体积。实现过程中需要用户交互，但仅限于模型参数的初始化以及选择建模所限制的“分析区域”。该分割方法在三个临床病例上进行了开发，并在另外四个临床病例上进行了测试，以确保对临床观察到的差异具有鲁棒性。它与最大强度的40%阈值分割以及最近一篇论文中提出的基于肿瘤与背景图像强度的阈值确定函数相比也具有优势。对该方法中依赖用户的部分进行了评估，并得出了该方法精度的初步估计值，其精度约为平均肿瘤体积估计值的±10%。通过这项工作，我们确立了基于GMM的分割在临床研究中的适用性，并对该方法在肿瘤体积分割方面的精度进行了初步评估。
