Unusual structures inherent in point pattern data predict colon cancer patient survival

Cancer patient diagnosis and prognosis is informed by assessment of morphological properties observed in patient tissue. Pathologists normally carry out this assessment, yet advances in computational image analysis provide opportunities for quantitative assessment of tissue. A key aspect of that quantitative assessment is the development of algorithms able to link image data to patient survival. Here, we develop a point process methodology able to describe patterns in cell distribution within cancerous tissue samples. In particular, we consider the Palm intensities of two Neyman Scott point processes, and a void process developed herein to reflect the spatial patterning of the cells. An approximate-likelihood technique is taken in order to fit point process models to patient data and the predictive performance of each model is determined. We demonstrate that based solely on the spatial arrangement of cells we are able to predict patient survival.
Comment: 20 pages including Appendix