Comparison of manual counting with two rule-based image analysis programs, ImageJ and CellProfiler, for quantitative analysis of c-Fos immunohistochemical positivity in rat brainstem sections.

Abstract

Quantitative and semiquantitative assessment of biomarker expression using immunohistochemistry (IHC) can be performed using several analytical approaches. Manual counting remains widely used but is inherently subjective and susceptible to observer variability. Rule-based image-analysis platforms may provide accessible alternatives, highlighting the need for direct method-comparison and agreement studies. c-Fos is a well-established marker of neuronal activation commonly quantified by counting IHC-positive cells. This study compared manual counting, ImageJ, and CellProfiler for quantifying c-Fos-positive cells in rat brainstem sections. Of 125 images initially evaluated, one was excluded as an extreme outlier, leaving 124 images for final analysis. Each image was independently analyzed by two trained observers using all three methods. ImageJ and CellProfiler were applied through their standard graphical interfaces and built-in functions using predefined user-guided threshold ranges, without additional plugins, scripts, machine-learning models, or custom code. Agreement was assessed using intraclass correlation coefficients (ICCs) and Bland-Altman analysis. No statistically significant difference in c-Fos-positive cell counts was observed among the three methods (Friedman test, P = 0.07). Median (IQR) counts were 6.75 (0.50-13.50) for manual counting, 6.25 (1.00-14.38) for ImageJ, and 6.00 (0.63-14.50) for CellProfiler. Inter-observer agreement was excellent for manual counting (ICC = 0.983), ImageJ (ICC = 0.964), and CellProfiler (ICC = 0.962). Inter-method agreement was also excellent (single-measures ICC = 0.971; 95% CI = 0.961-0.979). Bland-Altman analysis showed small mean biases of 0.33, 0.27, and -0.07 cells, with no significant proportional bias. Manual counting, ImageJ, and CellProfiler produced statistically comparable c-Fos-positive cell counts with excellent agreement. Under the predefined user-guided threshold ranges used, both software platforms closely matched manual counting; however, image-dependent threshold adjustment remains a potential source of observer variability.