Claudia Lang, Claudio Neidhöfer, Peter M Keller
In a low-incidence hospital laboratory, AI-assisted digital microscopy was technically feasible when deployed in a supervised configuration. Diagnostic performance improved only modestly compared with manual microscopy, whereas reproducibility and hands-on time were favorably affected. Implementation required substantial technical adaptation and resource investment, indicating that digital microscopy offers incremental standardization and workflow benefits rather than transformative gains in diagnostic accuracy.
BACKGROUND: AI-assisted digital microscopy has been proposed to augment acid-fast bacilli smear interpretation, but its performance in low-incidence laboratory settings is uncertain. We evaluated diagnostic accuracy, grading concordance, reproducibility, and workflow implications of AI-assisted fluorescence microscopy in routine mycobacteriology of such a setting.
METHODS: In a prospective single-centre implementation study (November-December 2023), 284 consecutive routine specimens were analyzed in parallel by manual microscopy and an AI-enabled digital system. Fully automated output and machine-assisted interpretation were compared with mycobacterial culture as reference. Digital grading thresholds were recalibrated using whole-slide quantitative ranges and optimized probability cutoffs.
RESULTS: Of 283 evaluable specimens, 15 (5.3%) were culture positive. Manual microscopy showed 40.0% sensitivity and 98.9% specificity. Fully automated output achieved 93.0% sensitivity but only 26.5% specificity. Machine-assisted interpretation yielded 53.3% sensitivity and 96.6% specificity. Application of a ≥92% probability cutoff increased exact grading agreement from 24% to 59% and eliminated major category deviations. Digital analysis showed high repeatability (96% identical results across repeated runs), whereas manual grading demonstrated inter-reader variability. Digital microscopy reduced hands-on time but increased total time-to-result due to computational processing.
CONCLUSION: In a low-incidence hospital laboratory, AI-assisted digital microscopy was technically feasible when deployed in a supervised configuration. Diagnostic performance improved only modestly compared with manual microscopy, whereas reproducibility and hands-on time were favorably affected. Implementation required substantial technical adaptation and resource investment, indicating that digital microscopy offers incremental standardization and workflow benefits rather than transformative gains in diagnostic accuracy.