Using computer vision libraries to streamline nuclei quantification: A validation study

Abstract

Valid analyses of cell function-based experiments depend on accuracy and reproducibility of the experimental protocol, often requiring normalization to cell count or other measures. The process of counting cells manually is laborious and susceptible to human error. Although cell counting programs exist, they are not preferred by all users, may be cost-prohibitive, and validation is minimal. Therefore, we aimed to validate an object visualization effciency tool that we developed to provide users with an alternative means of automated cell counting for the very specific, necessary, and widely performed task: quantifying nuclei to assess cell count in images. We hypothesized strong agreement between manual and automated counts. To ensure validation across cell densities, C2C12 myoblasts were seeded at 7.3x104 cells per cm2, proliferated for various timeframes (24, 48, and 72 hours; 6 wells across 2 plates per time point), and fixed with ice-cold methanol for 10 minutes. The nuclei were then stained using the fluorescent DNA stain, 4’,6-diamidino-2-phenylindole (DAPI), and 5 images per well were captured. The number of nuclei in each image were manually quantified by 2 separate raters (MC1 and MC2) using ImageJ, and with our automated program developed using computer vision libraries in Python (AI). To minimize risk for observer bias, AI counts and counts completed by each rater were kept separately until all 3 counts for each image were complete. To validate the program, we assessed interrater reliability between 1) MC1 and MC2 to ensure manual count agreement and 2) the average of MC1 and MC2 (MCavg) with AI to assess agreement between manual and AI counts. Spearman’s Rho (rs) was utilized to assess the strength of the association between counts across all images and separately for each quartile (Q) where Q1 had the fewest and Q4 had the most nuclei by manual counts. Nuclei per image ranged from 44 to 1166. Interrater reliability was nearly perfect between MC1 and MC2 (ICC=0.999) and between MCavg and AI (ICC=0.991). A significant, strong correlation was observed between AI and MCavg (rs=0.998, p<0.001). This strong correlation between persisted regardless of cell density (MCavg vs AI: Q1: rs=.998, p<0.001; Q2: rs=0.999, p<0.001; Q3: rs=1.000, p<0.001; Q4: rs=1.000, p<0.001). The results indicate strong agreement between nuclei counts obtained manually and using our automated object recognition program, which remained consistent regardless of cell quantity, demonstrating the validity of this program. Therefore, the automated object recognition program analyzed in the present study can be considered an accurate, rapid, user-friendly, and inexpensive resource for quantifying nuclei to assess cell count across a range of cell densities. Use of this software will save the user valuable time during the critically important but often time-consuming task of quantifying nuclei. Danielle Levitt's start-up funding through Texas Tech University. This is the full abstract presented at the American Physiology Summit 2024 meeting and is only available in HTML format. There are no additional versions or additional content available for this abstract. Physiology was not involved in the peer review process.