Multivariate classifications of transitional cell tumors of the bladder: nuclear abnormality index and pattern recognition analysis.

The authors report their experiences in applying two types of computer-aided multivariate classification systems in transitional cell tumors of the bladder. The systems are based on nuclear changes in urothelial papillary lesions. 19 out of 54 parameters, selected on the basis of the overlapping area between contiguous grades and the monotonic function, were combined mathematically to obtain the nuclear abnormality index. This index expresses the progressive nuclear abnormalities as a single number on a continuous scale from 0.5 to 2.6. The standard deviation (SD) of the 10 largest nuclear area values in the lower half of epithelial thickness (L), the mean perimeter (L), the SD of the roundness factor in the upper half of epithelial thickness (U), the SD of the logarithm of area (L) and the percentage of round nuclei (U) represent the smallest, the most discriminant and the least correlated set of features for the pattern recognition analysis. The latter shows an agreement of 92% between computer and histologic classifications-estimated by applying Bayes theorem. A good correlation exists between nuclear abnormality index and computer grading by pattern recognition analysis.