Knowledge-guided segmentation and morphometric analysis of colorectal dysplasia.

The histopathologic grading of colorectal adenomatous dysplasia is a subjective process. In this study, automated image analysis using knowledge-guided software was used to quantitatively assess colorectal glandular characteristics. Cases of histologically normal mucosa and of low, moderate and high grade dysplasia were examined using the technique. A total of 19 morphometric and densitometric features were measured on each gland. These included gland shape, epithelial area, nuclear stratification and nuclear optical density. Discriminant analysis of the data revealed those morphometric features which provided the best discrimination between the various histologic groups. The area of epithelium occupied by nuclei was the strongest discriminating variable, and this, in combination with a discriminant function derived from the remaining variables, was used to plot cases in bivariate sample space. The plots revealed that the data for normal glands were consistently well separated from dysplastic gland data. Data sets belonging to the various grades of dysplasia showed varying degrees of separation, depending which two histologic groups the discriminant function was based on. This study showed that automated image analysis of complex histologic scenes is possible using knowledge-guided segmentation and that it can provide useful data for the objective classification of colorectal dysplasia.