正常成年人前角膜表面形状的聚类分析
Clustering of anterior corneal surface shapes in normal adults
浏览论文内容
中文总结 AI 辅助
本研究对正常成年人角膜地形图进行聚类分析，用Zernike多项式降维后采用k-means算法确定3类前角膜形状，为生物合成角膜植入物的分类设计提供依据。
中文摘要 AI 辅助
本研究调查了一个大型正常成年人角膜地形图数据集，旨在探索性地识别其三维形状的自然分组，实际临床目标是对新一代生物合成角膜植入物的形状进行分类，以匹配患者自身的角膜形状类别。由于正常角膜的形状范围广泛，我们需要确定有限数量的正常角膜形状类别，以指导不同植入物的设计，使其匹配患者的角膜形状类别。在本研究阶段，我们聚焦于前表面，其在正常人群中表现出更大的变异，且承担了眼睛的大部分屈光力。所用的角膜表面三维数据是取自角膜地形图的前表面高度图。对角膜地形图高度图进行聚类以揭示正常角膜形状类别，为便于分组，通过Zernike多项式建模降低角膜数据的维度。使用不同的聚类分数、表示方法和统计聚类比较对所得聚类进行评估。尽管以这种方式测试了多种硬聚类、软聚类、线性和非线性聚类方法，但k-means（k均值）算法被证明足以完成此任务。最佳聚类数为3个，它们主要根据角膜曲率（Zernike离焦系数）进行区分。这些聚类与已确立的临床参数相关，证实该算法仅基于正常角膜的三维形状提取了相关信息。
英文摘要
In the present study, we investigated a large dataset of normal adult corneal topographies in an exploratory attempt to identify the natural groupings of their 3D shapes, with the practical clinical aim of categorizing the shape of a future generation of biosynthetic corneal implants. Because of the wide range of shapes among normal corneas, we needed to identify a limited number of normal corneal shape categories to guide the conception of different implants to match the patient's own corneal shape category. At this stage of our research, we focused on the anterior surface, which exhibits greater variation in a normal population and is responsible for most of the refractive power of the eye. The corneal surface 3D data used were the anterior elevation maps taken from corneal topographies. Clustering corneal topography elevation maps was performed to reveal the normal corneal shape categories. To facilitate groupings, corneal data had their dimensionality reduced by Zernike polynomial modeling. The resulting clusters were evaluated using different clustering scores, representations and statistical cluster comparisons. While several hard and soft linear and nonlinear clustering methods were tested in this way, k-means proved to be sufficient for this task. The best number of clusters was three and they were primarily differentiated according to corneal curvature (Zernike defocus coefficient). These clusters were related to established clinical parameters, confirming that the algorithm extracted relevant information based solely on the normal 3D shape of corneas.
