Understanding spatial inhomogeneity and clustering in point patterns arises in many contexts, ranging from disease outbreak monitoring to analyzing radiologically-based emphysema in biomedical images. This can often involve classifying individual points as being part of a feature/cluster or as being part of a background noise process. Existing methods for this task can struggle when there are differences in the size and/or density of individual clusters. In this work, we propose employing kernel density estimates of the underlying point process density function, using an existing data-driven approach to bandwidth selection, to separate feature points from noise. This is achieved by constructing a null distribution, either through asymptotic properties or Monte Carlo simulation, and comparing kernel density estimates to a given quantile of this distribution. We demonstrate that our method, termed Kernel Density and Simulation based Filtering (KDS-Filt), showed superior performance to existing alternative approaches, especially when there is inhomogeneity in cluster sizes and density. We also show the utility of KDS-Filt for identifying clinically relevant information about the spatial distribution of emphysema in lung computed tomography scans. The KDS-Filt methodology is available as part of the sncp R package, which can be downloaded at https://github.com/stop-pre16/sncp.

了解点模式的空间不均匀性和聚类在许多情况下都会出现，从疾病爆发监测到分析生物医学图像中基于放射学的肺气肿。这通常涉及将各个点分类为特征/簇的一部分或背景噪声过程的一部分。当各个簇的大小和/或密度存在差异时，用于此任务的现有方法可能会很困难。在这项工作中，我们建议采用底层点过程密度函数的核密度估计，使用现有的数据驱动的带宽选择方法，将特征点与噪声分开。这是通过渐近性质或蒙特卡罗模拟构造零分布，并将核密度估计值与该分布的给定分位数进行比较来实现的。我们证明，我们的方法（称为基于核密度和模拟的过滤（KDS-Filt））显示出优于现有替代方法的性能，特别是当簇大小和密度不均匀时。我们还展示了 KDS-Filt 在识别肺部计算机断层扫描中肺气肿空间分布的临床相关信息方面的实用性。KDS-Filt 方法作为sncp R 包的一部分提供，可以在 https://github.com/stop-pre16/sncp 下载。