Pataky Todd C, Goulermas John Y, Crompton Robin H
HACB, School of Biomedical Sciences, University of Liverpool, Sherrington Buildings, Ashton Street, Liverpool L69 3GE, UK.
J Biomech. 2008 Oct 20;41(14):3085-9. doi: 10.1016/j.jbiomech.2008.08.001. Epub 2008 Sep 13.
Image registration, the process of transforming images such that homologous structures optimally overlap, provides the pre-processing foundation for pixel-level functional image analysis. The purpose of this study was to compare the performances of seven methods of within-subjects pedobarographic image registration: (1) manual, (2) principal axes, (3) centre of pressure trajectory, (4) mean squared error, (5) probability-weighted variance, (6) mutual information, and (7) exclusive OR. We assumed that foot-contact geometry changes were negligibly small trial-to-trial and thus that a rigid-body transformation could yield optimum registration performance. Thirty image pairs were randomly selected from our laboratory database and were registered using each method. To compensate for inter-rater variability, the mean registration parameters across 10 raters were taken as representative of manual registration. Registration performance was assessed using four dissimilarity metrics (#4-7 above). One-way MANOVA found significant differences between the methods (p<0.001). Bonferroni post-hoc tests revealed that the centre of pressure method performed the poorest (p<0.001) and that the principal axes method tended to perform more poorly than remaining methods (p<0.070). Average manual registration was not different from the remaining methods (p=1.000). The results suggest that a variety of linear registration methods are appropriate for within-subjects pedobarographic images, and that manual image registration is a viable alternative to algorithmic registration when parameters are averaged across raters. The latter finding, in particular, may be useful for cases of image peculiarities resulting from outlier trials or from experimental manipulations that induce substantial changes in contact area or pressure profile geometry.
图像配准是将图像进行变换以使同源结构实现最佳重叠的过程，它为像素级功能图像分析提供了预处理基础。本研究的目的是比较七种受试者内足压图像配准方法的性能：（1）手动配准；（2）主轴配准；（3）压力中心轨迹配准；（4）均方误差配准；（5）概率加权方差配准；（6）互信息配准；（7）异或配准。我们假设每次试验之间足部接触几何形状的变化极小，可以忽略不计，因此刚体变换能够产生最佳的配准性能。从我们的实验室数据库中随机选择了30对图像，并使用每种方法进行配准。为了补偿不同评分者之间的差异，将10名评分者的平均配准参数作为手动配准的代表。使用四种差异度量（上述方法4 - 7）评估配准性能。单因素多元方差分析发现不同方法之间存在显著差异（p<0.001）。Bonferroni事后检验表明，压力中心法的性能最差（p<0.001），主轴法的性能往往比其他方法更差（p<0.070）。平均手动配准与其他方法没有差异（p = 1.000）。结果表明，多种线性配准方法适用于受试者内足压图像，并且当对评分者的参数进行平均时，手动图像配准是算法配准的可行替代方法。特别是后一个发现，对于因异常试验或导致接触面积或压力分布几何形状发生重大变化的实验操作而产生图像特殊性的情况可能有用。
