Yang Jinzhong, Beadle Beth M, Garden Adam S, Schwartz David L, Aristophanous Michalis
Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, Texas 77030.
Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, Texas 77030.
Med Phys. 2015 Sep;42(9):5310-20. doi: 10.1118/1.4928485.
To develop an automatic segmentation algorithm integrating imaging information from computed tomography (CT), positron emission tomography (PET), and magnetic resonance imaging (MRI) to delineate target volume in head and neck cancer radiotherapy.
Eleven patients with unresectable disease at the tonsil or base of tongue who underwent MRI, CT, and PET/CT within two months before the start of radiotherapy or chemoradiotherapy were recruited for the study. For each patient, PET/CT and T1-weighted contrast MRI scans were first registered to the planning CT using deformable and rigid registration, respectively, to resample the PET and magnetic resonance (MR) images to the planning CT space. A binary mask was manually defined to identify the tumor area. The resampled PET and MR images, the planning CT image, and the binary mask were fed into the automatic segmentation algorithm for target delineation. The algorithm was based on a multichannel Gaussian mixture model and solved using an expectation-maximization algorithm with Markov random fields. To evaluate the algorithm, we compared the multichannel autosegmentation with an autosegmentation method using only PET images. The physician-defined gross tumor volume (GTV) was used as the "ground truth" for quantitative evaluation.
The median multichannel segmented GTV of the primary tumor was 15.7 cm(3) (range, 6.6-44.3 cm(3)), while the PET segmented GTV was 10.2 cm(3) (range, 2.8-45.1 cm(3)). The median physician-defined GTV was 22.1 cm(3) (range, 4.2-38.4 cm(3)). The median difference between the multichannel segmented and physician-defined GTVs was -10.7%, not showing a statistically significant difference (p-value = 0.43). However, the median difference between the PET segmented and physician-defined GTVs was -19.2%, showing a statistically significant difference (p-value =0.0037). The median Dice similarity coefficient between the multichannel segmented and physician-defined GTVs was 0.75 (range, 0.55-0.84), and the median sensitivity and positive predictive value between them were 0.76 and 0.81, respectively.
The authors developed an automated multimodality segmentation algorithm for tumor volume delineation and validated this algorithm for head and neck cancer radiotherapy. The multichannel segmented GTV agreed well with the physician-defined GTV. The authors expect that their algorithm will improve the accuracy and consistency in target definition for radiotherapy.
开发一种自动分割算法，整合计算机断层扫描（CT）、正电子发射断层扫描（PET）和磁共振成像（MRI）的成像信息，以在头颈癌放疗中勾勒靶区体积。
招募了11例扁桃体或舌根不可切除疾病患者，这些患者在放疗或放化疗开始前两个月内接受了MRI、CT和PET/CT检查。对于每位患者，首先分别使用可变形配准和刚性配准将PET/CT和T1加权对比增强MRI扫描配准到计划CT上，以将PET和磁共振（MR）图像重采样到计划CT空间。手动定义一个二值掩码以识别肿瘤区域。将重采样后的PET和MR图像、计划CT图像以及二值掩码输入到自动分割算法中进行靶区勾勒。该算法基于多通道高斯混合模型，并使用带有马尔可夫随机场的期望最大化算法求解。为了评估该算法，我们将多通道自动分割与仅使用PET图像的自动分割方法进行了比较。将医生定义的大体肿瘤体积（GTV）用作定量评估的“金标准”。
原发肿瘤多通道分割的GTV中位数为15.7 cm³（范围为6.6 - 44.3 cm³），而PET分割的GTV为10.2 cm³（范围为2.8 - 45.1 cm³）。医生定义的GTV中位数为22.1 cm³（范围为4.2 - 38.4 cm³）。多通道分割的GTV与医生定义的GTV之间的中位数差异为 - 10.7%，无统计学显著差异（p值 = 0.43）。然而，PET分割的GTV与医生定义的GTV之间的中位数差异为 - 19.2%，有统计学显著差异（p值 = 0.0037）。多通道分割的GTV与医生定义的GTV之间的中位数Dice相似系数为0.75（范围为0.55 - 0.84），它们之间的中位数敏感性和阳性预测值分别为0.76和0.81。
作者开发了一种用于肿瘤体积勾勒的自动化多模态分割算法，并对头颈癌放疗中的该算法进行了验证。多通道分割的GTV与医生定义的GTV吻合良好。作者期望他们的算法将提高放疗靶区定义的准确性和一致性。
