Nearest neighbor spatial statistics as a framework for analyzing spatial patterns from image data.

Abstract

Quantitatively describing spatial patterns in biological substrates is a difficult task that requires careful consideration of labeling methods, imaging modality, and the motivating hypothesis. This problem is commonly addressed using intensity correlation paradigms which correlate the brightness between fluorescence images of co-labeled biomolecules as a measure of their colocalization. These approaches ultimately characterize the superposition of fluorescence signals and thus are limited to binary assessment of spatial relationships at the scale of the image resolution. We’ve previously demonstrated an image analysis framework, which leverages nearest neighbor (NN) distribution functions (F, G, and L) for richer, more informative quantification of the spatial distribution of immunofluorescently imaged biomolecules on all spatial scales captured within the image. Briefly, biomolecule signals and relevant biological structures are segmented, biomolecules are localized relative to landmarks, which are either structures or other biomolecules, by measuring NN distances between biomolecule and landmark voxels, and finally their global spatial association is assessed by comparing the resulting distance distribution to a null hypothesis distribution, in our case a homogenous Poisson distribution which represents complete spatial randomness. The sign of the difference between the distributions indicates repulsion or attraction of the biomolecule to the landmark and the magnitude indicates the relationship's strength. Ongoing efforts aim to further validate this framework by characterizing its sensitivity to image segmentation parameters and signal concentration, and the impact on measured spatial relationships of our methods for generating the null distribution and measuring distances within an image of anisotropic resolution. Preliminary results indicate our framework is robust over a wide range of input parameters and data quality, and our custom methods introduce errors that are smaller than the image's precision. Thus, this approach has wide applicability for scientific image analysis.