Automatic Analysis of Cephalogram for Orthodontics

(1995)对作者在1992年提出的一种新模糊神经元进行了推广和改进，我们称之为广义模糊学习机。该机器能够很好地把握各输入输出之间的非线性相关性。与传统的神经网络相比，它具有很高的非线性映射能力，并且保证了全局最小。此外，它的学习速度和精度都有了很大的提高，并成功地应用于X线片上标志点位置的自动检测，用于正畸治疗。(1996)提出了一种基于神经网络和模糊模板匹配的X线片标志点提取方法。针对模糊模板匹配问题，提出了两种加权相似性度量方法。首先通过神经网络找出地标应该位于的粗略区域。然后对该区域执行模糊模板匹配，以找到其地标的准确位置。在实际X线片中成功地找到了典型的标志点，误差在允许的范围内。(1997)提出了用径向基函数神经网络预测颅面复合体生长的方法。颅面复合体的发育预测在正畸领域是非常重要的，因为如果不能很好地预测它，就需要再次手术，这会给患者带来生理和/或精神上的痛苦。首先通过模糊聚类将一组学习数据划分为三个骨架类，然后为每个类构建RBFN。预测采用各径向基神经网络输出的加权和，预测结果令人满意。

(1995)A neo-fuzzy-neuron, presented by the authors in 1992, was generalized and modified, which we call a generalized fuzzy learning machine. This machine can well grasp the nonlinear correlation of each input and output. It has a very high nonlinear mapping ability compared with the conventional neural network, and it guaranteesa global minimum. Furthermore, the learning speed and its accuracy are improved drastically, It was successfully applied to the automatic detection of landmark positions in the roentgenographic cephalogram for an orthodontic treatment.(1996)An extraction of landmarks in a roentgenographic cephalogram by using a neural network and a fuzzy template matching was proposed. Two kinds of weighted similarity measures are newly proposed for a fuzzy template matching. The rough region where a landmark is supposed to be located is first found out by a neural network. The fuzzy template matching is then performed over this region to find the exact location of its landmark. Typical landmarks were successfully found in the actual roentgenographic cephalogram within a permissible error for a practical use.(1997)Growth prediction of craniofacial complex by using an RBFN(Radial Basis Function Network) was proposed. The growth prediction of craniofacial complex is very important in the field of orthodontics, because if it is not well predicted re-operation would be necessary, which causes physical and/or mental pain to a patient. A set of learning data was first divided into three skeletal groups by Fuzzy clustering, and then RBFN was constructed for each cluster. The prediction was performed by taking the weighted sum of the outputs of each RBFN.The prediction results were promising.