Gardner Stephen J, Wen Ning, Kim Jinkoo, Liu Chang, Pradhan Deepak, Aref Ibrahim, Cattaneo Richard, Vance Sean, Movsas Benjamin, Chetty Indrin J, Elshaikh Mohamed A
Department of Radiation Oncology, Josephine Ford Cancer Institute, Henry Ford Health System, Detroit, MI 48202, USA.
Phys Med Biol. 2015 Jun 7;60(11):4429-47. doi: 10.1088/0031-9155/60/11/4429. Epub 2015 May 19.
This study was designed to evaluate contouring variability of human-and deformable-generated contours on planning CT (PCT) and CBCT for ten patients with low-or intermediate-risk prostate cancer. For each patient in this study, five radiation oncologists contoured the prostate, bladder, and rectum, on one PCT dataset and five CBCT datasets. Consensus contours were generated using the STAPLE method in the CERR software package. Observer contours were compared to consensus contour, and contour metrics (Dice coefficient, Hausdorff distance, Contour Distance, Center-of-Mass [COM] Deviation) were calculated. In addition, the first day CBCT was registered to subsequent CBCT fractions (CBCTn: CBCT2-CBCT5) via B-spline Deformable Image Registration (DIR). Contours were transferred from CBCT1 to CBCTn via the deformation field, and contour metrics were calculated through comparison with consensus contours generated from human contour set. The average contour metrics for prostate contours on PCT and CBCT were as follows: Dice coefficient-0.892 (PCT), 0.872 (CBCT-Human), 0.824 (CBCT-Deformed); Hausdorff distance-4.75 mm (PCT), 5.22 mm (CBCT-Human), 5.94 mm (CBCT-Deformed); Contour Distance (overall contour)-1.41 mm (PCT), 1.66 mm (CBCT-Human), 2.30 mm (CBCT-Deformed); COM Deviation-2.01 mm (PCT), 2.78 mm (CBCT-Human), 3.45 mm (CBCT-Deformed). For human contours on PCT and CBCT, the difference in average Dice coefficient between PCT and CBCT (approx. 2%) and Hausdorff distance (approx. 0.5 mm) was small compared to the variation between observers for each patient (standard deviation in Dice coefficient of 5% and Hausdorff distance of 2.0 mm). However, additional contouring variation was found for the deformable-generated contours (approximately 5.0% decrease in Dice coefficient and 0.7 mm increase in Hausdorff distance relative to human-generated contours on CBCT). Though deformable contours provide a reasonable starting point for contouring on CBCT, we conclude that contours generated with B-Spline DIR require physician review and editing if they are to be used in the clinic.
本研究旨在评估10例低危或中危前列腺癌患者在计划CT（PCT）和CBCT上由人工绘制及可变形模型生成的轮廓的变异性。对于本研究中的每位患者，5名放射肿瘤学家在1个PCT数据集和5个CBCT数据集上绘制前列腺、膀胱和直肠的轮廓。使用CERR软件包中的STAPLE方法生成共识轮廓。将观察者绘制的轮廓与共识轮廓进行比较，并计算轮廓指标（骰子系数、豪斯多夫距离、轮廓距离、质心[COM]偏差）。此外，通过B样条可变形图像配准（DIR）将第一天的CBCT与后续的CBCT分次（CBCTn：CBCT2 - CBCT5）进行配准。轮廓通过变形场从CBCT1转移到CBCTn，并通过与从人工轮廓集生成的共识轮廓进行比较来计算轮廓指标。PCT和CBCT上前列腺轮廓的平均轮廓指标如下：骰子系数 - 0.892（PCT），0.872（CBCT - 人工），0.824（CBCT - 变形）；豪斯多夫距离 - 4.75毫米（PCT），5.22毫米（CBCT - 人工），5.94毫米（CBCT - 变形）；轮廓距离（整体轮廓） - 1.41毫米（PCT），1.66毫米（CBCT - 人工），2.30毫米（CBCT - 变形）；COM偏差 - 2.01毫米（PCT），2.78毫米（CBCT - 人工），3.45毫米（CBCT - 变形）。对于PCT和CBCT上的人工轮廓，与每位患者观察者之间的差异（骰子系数标准差为5%，豪斯多夫距离标准差为2.0毫米）相比，PCT和CBCT之间平均骰子系数的差异（约2%）和豪斯多夫距离的差异（约0.5毫米）较小。然而，对于可变形模型生成的轮廓发现了额外的轮廓变异性（相对于CBCT上人工生成的轮廓，骰子系数约降低5.0%，豪斯多夫距离增加0.7毫米）。尽管可变形轮廓为CBCT上的轮廓绘制提供了合理的起点，但我们得出结论，如果要在临床中使用，通过B样条DIR生成的轮廓需要医生进行审查和编辑。
