Methods: An automated digital imaging method was used to segment the epithelial compartment in images from 9 human normal mucosa samples and 5 well differentiated squamous cell carcinomas into theoretical cell extents belonging to exclusive areas of influence of the epithelial cell nuclei. The spatial arrangement of these cells in circular neighbourhoods of 37.5 (small) and 75 (large) micrometres diameter was characterised by 29 statistical properties derived from four constrained graph networks constructed over the cell centroids. In total, 85,704 small neighbourhoods (38,573 normal and 47,131 neoplastic) and 55,577 large neighbourhoods (24,175 normal and 31,402 neoplastic) were analysed.
Results: Stepwise hierarchical discriminant analysis revealed that 68.2% of small and 74.5% of the large neighbourhoods could be classified correctly in the original diagnostic classes. Principal component analysis showed that 98% of the variance in the data could be explained by 7 components. Fisher discriminant coefficients were then used to produce a data mapping procedure to label and pinpoint in the original images the epithelial regions typical of each diagnostic class. Case-wise the classification rate using means and standard deviations of all architectural parameters from large neighbourhoods was 100% correct for both diagnostic classes.
Conclusion: Although the spatial arrangement of cells in the epithelial compartment of the oral mucosa cannot be consistently assessed by visual observation alone, it can be objectively quantified using the methods described. This approach could prove useful in the development of automated diagnostic systems, as quantitative markers of disease and treatment progression.