Application of automatic thresholding in image analysis scoring of cells in human solid tumors labeled for proliferation markers

Abstract

We previously reported an image analysis program that uses four investigator-defined parameters including two thresholds, i.e., gray-level threshold (GLT) and hue threshold (HT), to determine the number of cells (TC) and the proliferating cell nuclear antigen (PCNA) or bromodeoxyuridine (BrdUrd) labeling indices (LI) in human solid tumors. The present study investigated if the accuracy and reproducibility of image analysis results can be improved by using computer-defined GLT and HT. Three investigators evaluated 142 images on 3 days, using visual analysis and four image analysis routines, which used different combinations of computer- and investigator- selected GLT and HT. The data show that image analysis using computer- selected GLT and HT yielded (i) LI of PCNA and BrdUrd that were indistinguishable from visual analysis, (ii) equal (BrdUrd LI) or better (TC and PCNA LI) inter-day reproducibility relative to visual analysis results, and (iii) results that were equally (TC) or more accurate (LI of PCNA and BrdUrd) with higher inter-day reproducibility (TC and LI of PCNA and BrdUrd) than image analysis obtained using investigator-defined thresholds. We conclude that the use of computer-defined GLT and HT improved the accuracy and reproducibility of image analysis results.