Abstract
ABSTRACT
Automated micronucleus (MN) scoring is increasingly proposed as a remedy for the subjectivity and labour burden of manual scoring yet reported performance has not been appraised systematically. This critical systematic review with quantitative synthesis follows PRISMA 2020. PubMed was searched for studies applying computer‐vision, machine‐learning, or deep‐learning methods to MN determination against expert or expert‐derived reference standards. Twenty‐four records were screened and eight (2020–2025) met eligibility, spanning fluorescence microscopy, brightfield/Giemsa preparations, imaging flow cytometry, and whole‐slide imaging. Because no study reported a reconstructable 2 × 2 table, bivariate diagnostic‐accuracy meta‐analysis was not possible; reported scalar metrics were instead logit‐transformed and combined by DerSimonian–Laird random effects as explicitly descriptive summaries, with correlations combined on the Fisher‐z scale. Summarized recall was 0.924 (95% CI 0.845–0.965) and precision 0.874 (0.790–0.927), both with near‐total heterogeneity (I
2
≈ 97%), whereas agreement between automated and reference counts was high and homogeneous (Pearson r = 0.920, 0.911–0.929; I
2
= 0%). Metric definition was the dominant moderator of heterogeneity; residual variation in precision tracked imaging modality rather than network architecture. Although there are bias regarding choice of staining or systems used for analysis as discussed here, current evidence supports a future of automated MN frequency estimation for regulatory and biomonitoring endpoints; but not object‐level equivalence with expert scoring.