Friday, May 6 2022
9:00am - 11:00am
Masters Presentation
An Application of Spatial Filtering to Populations of Lung CT Images

Abstract: A defining characteristic of classical linear models is the assumption of independent errors. That is, dependent values do not relate to one another outside of a shared relationship with a set of explanatory variables. However, spatially referenced data is defined by an opposing characteristic: spatial autocorrelation. In short, observations in a spatially referenced data set are more similar when they are in closer physical proximity to one another, and inference that fails to account for this relationship can be invalid. An example of spatially autocorrelated data are the voxels of 3-dimensional lung density images produced by computed tomography (CT). These images are a powerful tool for clinicians to detect and track lung diseases, however they are typically analyzed through individual visual inspection. Large-scale statistical analysis of CT images may provide valuable insight into the phenotypes and progression of lung diseases, but doing so requires that analysts overcome the hurdle of modeling spatially autocorrelated data. Eigenvector spatial filtering (ESF) attempts to account for spatial autocorrelation by adapting linear models to use spatially varying coefficients through the inclusion of the eigenvectors of distance matrices. In this analysis, we have applied spVBM, an ESF method created to analyze populations of CT lung images, to a sample of lung CT images from individuals with sarcoidosis, a multisystem inflammatory disease that often manifests in non-caseating granulomas in the lung. Using spVBM, we evaluate spatially varying associations between several features of lung density and measures of lung function. We will discuss the pre-processing steps required to analyze populations of lung CT data, the spVBM method, the associated findings of our application of spVBM, and the degree to which spatial autocorrelation was accounted for in this application.
Speaker:Max McGrath
Affiliation:
Location:Zoom link in email


Download as iCalendar

Done