School of Applied Sciences and Health Innovations Research Institute, RMIT University, Melbourne 3000, Australia.
Med Phys. 2013 Oct;40(10):101701. doi: 10.1118/1.4819945.
Deformable image registration (DIR) has become a key tool for adaptive radiotherapy to account for inter- and intrafraction organ deformation. Of contemporary interest, the application to deformable dose accumulation requires accurate deformation even in low contrast regions where dose gradients may exist within near-uniform tissues. One expects high-contrast features to generally be deformed more accurately by DIR algorithms. The authors systematically assess the accuracy of 12 DIR algorithms and quantitatively examine, in particular, low-contrast regions, where accuracy has not previously been established.
This work investigates DIR algorithms in three dimensions using deformable gel (DEFGEL) [U. J. Yeo, M. L. Taylor, L. Dunn, R. L. Smith, T. Kron, and R. D. Franich, "A novel methodology for 3D deformable dosimetry," Med. Phys. 39, 2203-2213 (2012)], for application to mass- and density-conserving deformations. CT images of DEFGEL phantoms with 16 fiducial markers (FMs) implanted were acquired in deformed and undeformed states for three different representative deformation geometries. Nonrigid image registration was performed using 12 common algorithms in the public domain. The optimum parameter setup was identified for each algorithm and each was tested for deformation accuracy in three scenarios: (I) original images of the DEFGEL with 16 FMs; (II) images with eight of the FMs mathematically erased; and (III) images with all FMs mathematically erased. The deformation vector fields obtained for scenarios II and III were then applied to the original images containing all 16 FMs. The locations of the FMs estimated by the algorithms were compared to actual locations determined by CT imaging. The accuracy of the algorithms was assessed by evaluation of three-dimensional vectors between true marker locations and predicted marker locations.
The mean magnitude of 16 error vectors per sample ranged from 0.3 to 3.7, 1.0 to 6.3, and 1.3 to 7.5 mm across algorithms for scenarios I to III, respectively. The greatest accuracy was exhibited by the original Horn and Schunck optical flow algorithm. In this case, for scenario III (erased FMs not contributing to driving the DIR calculation), the mean error was half that of the modified demons algorithm (which exhibited the greatest error), across all deformations. Some algorithms failed to reproduce the geometry at all, while others accurately deformed high contrast features but not low-contrast regions-indicating poor interpolation between landmarks.
The accuracy of DIR algorithms was quantitatively evaluated using a tissue equivalent, mass, and density conserving DEFGEL phantom. For the model studied, optical flow algorithms performed better than demons algorithms, with the original Horn and Schunck performing best. The degree of error is influenced more by the magnitude of displacement than the geometric complexity of the deformation. As might be expected, deformation is estimated less accurately for low-contrast regions than for high-contrast features, and the method presented here allows quantitative analysis of the differences. The evaluation of registration accuracy through observation of the same high contrast features that drive the DIR calculation is shown to be circular and hence misleading.
变形图像配准（DIR）已成为自适应放疗的关键工具，以考虑器官的内和分次间变形。目前的研究兴趣在于，对于变形剂量积累的应用，即使在近均匀组织内存在剂量梯度的低对比度区域，也需要准确的变形。人们期望 DIR 算法通常能更准确地变形高对比度特征。作者系统地评估了 12 种 DIR 算法的准确性，并特别定量检查了低对比度区域，此前尚未确定该区域的准确性。
本工作使用可变形凝胶（DEFGEL）[U. J. Yeo、M. L. Taylor、L. Dunn、R. L. Smith、T. Kron 和 R. D. Franich，“用于 3D 变形剂量测定的新方法”，《医学物理学》39，2203-2213（2012 年）]研究了三维中的 DIR 算法，适用于质量和密度守恒的变形。在三个不同的代表性变形几何形状中，对具有 16 个基准标记（FM）的 DEFGEL 体模的 CT 图像进行了变形和未变形状态的采集。使用公共领域的 12 种常用算法进行非刚性图像配准。为每个算法确定了最佳参数设置，并在三种情况下测试了每个算法的变形准确性：（I）具有 16 个 FM 的原始 DEFGEL 图像；（II）通过数学方式删除 8 个 FM 的图像；以及（III）通过数学方式删除所有 FM 的图像。然后将在场景 II 和 III 中获得的变形矢量场应用于包含所有 16 个 FM 的原始图像。通过 CT 成像确定的算法估计的 FM 位置与实际位置进行比较。通过评估真实标记位置和预测标记位置之间的三维向量来评估算法的准确性。
对于场景 I 到 III，每个样本的 16 个误差向量的平均幅度分别为 0.3 到 3.7、1.0 到 6.3 和 1.3 到 7.5mm，跨算法。原始 Horn 和 Schunck 光流算法表现出最大的准确性。在这种情况下，对于场景 III（不参与驱动 DIR 计算的删除 FM），在所有变形中，平均误差是修改后的 demons 算法（表现出最大误差）的一半。有些算法根本无法再现几何形状，而有些算法则准确地变形了高对比度特征，但不能变形低对比度区域，表明在地标之间插值不佳。
使用组织等效、质量和密度守恒的 DEFGEL 体模对 DIR 算法的准确性进行了定量评估。对于研究的模型，光流算法比 demons 算法表现更好，原始 Horn 和 Schunck 算法表现最好。误差的程度更多地受到位移幅度的影响，而不是变形的几何复杂性的影响。正如预期的那样，低对比度区域的变形比高对比度特征的变形估计精度低，并且本文提出的方法允许对差异进行定量分析。通过观察驱动 DIR 计算的相同高对比度特征来评估配准准确性被证明是循环的，因此具有误导性。
