An automated pattern recognition system for the quantification of inflammatory cells in hepatitis-C-infected liver biopsies

This paper presents an automated system for the quantification of inflammatory cells in hepatitis-C-infected liver biopsies. Initially, features are extracted from colour-corrected biopsy images at positions of interest identified by adaptive thresholding and clump decomposition. A sequential floating search method and principal component analysis are used to reduce dimensionality. Manually annotated training images allow supervised training. The performance of Gaussian parametric and mixture models is compared when used to classify regions as either inflammatory or healthy. The system is optimized using a response surface method that maximises the area under the receiver operating characteristic curve. This system is then tested on images previously ranked by a number of observers with varying levels of expertise. These results are compared to the automated system using Spearman rank correlation. Results show that this system can rank 15 test images, with varying degrees of inflammation, in strong agreement with five expert pathologists.