Variability in visual segmentation of digitized prostate tissue microarray cores

To examine bias associated with human-interactive semi-automated systems key components with machine vision used in quantitative histometry.

A standard image set of 20 images was created using 5 nuclei sampled from hematoxylin-eosin-stained sections of benign tissue within a prostate tissue microarray that were rotated through the cardinal directions. Four trained technicians performed segmentation of these images at the start, then at the end, of 3 daily sessions, creating a total analytic set of 480 observations. Measurements of nuclear area (NA), nuclear roundness factor (NRF), and mean optical density (MOD) were compared by segmenter, time, and rotational orientation.

NA varied significantly among sessions (p < 0.0009) and session variance differed within segmenter (p < 0.0001). NRF was significant among segmenters (p < 0.001) and sessions (p < 0.0001), and in session (p < 0.0001) and intra-session differences (p = 0.026). Differences in MOD varied among sessions (p < 0.0001) and within sessions (p < 0.049).

Imaging systems remain vulnerable to statistical inter-segmenter variation, in spite of extensive efforts to eliminate variation among individual segmenters. As statistical significance often guides decision-making in morphometric analysis, statistically significant effects potentially produce bias. Current practices and quality assurance methods require review to eliminate individual operator effects in semiautomated machine systems.