COMPARATIVE DIAGNOSTIC CONCORDANCE AND TURNAROUND TIME OF ARTIFICIAL INTELLIGENCEASSISTED AND CONVENTIONAL HISTOPATHOLOGY: A PROSPECTIVE OBSERVATIONAL STUDY

Abstract

Background: Whole-slide imaging has created a practical route for artificial intelligence (AI) to support histopathological interpretation. Its value in routine diagnostic practice depends not only on algorithmic accuracy but also on whether it improves pathologist concordance and reduces reporting time across clinically varied specimens. Objectives: To compare case-level diagnostic concordance with a final integrated reference diagnosis and the time required for AI-only output, conventional pathologist diagnosis, and pathologist diagnosis augmented by AI. Materials and Methods: This prospective observational comparative study included 50 consecutive histopathology cases evaluated in the Department of Pathology, Shri Sathya Sai Medical College and Research Institute, Tiruporur, during 2025–2026. Routine slides were digitally scanned. An AI-assisted diagnostic system generated a top-ranked diagnosis and confidence score. Two pathologists initially interpreted the cases without AI and, after a washout period, re-evaluated the digital slides with AI support to record a consensus diagnosis. The final integrated histopathological diagnosis, incorporating clinical correlation and ancillary investigations where required, served as the reference standard. Exact concordance was compared using McNemar’s test; paired diagnostic times were analysed using the Friedman and Wilcoxon signed-rank tests. Results: The mean age was 52.08±13.19 years, and 27 (54.0%) patients were female. The reference diagnoses were malignant in 25 (50.0%), benign in 15 (30.0%), inflammatory in 6 (12.0%), and premalignant in 4 (8.0%) cases. Exact concordance with the reference standard was 100% for the AI-only output and AI-augmented pathologist diagnosis, compared with 68.0% for conventional pathologist diagnosis (34/50; exact McNemar p<0.001). All 16 conventional discordances were concordant after AI augmentation. Mean diagnostic time was 0.51±0.10 minutes for AI output, 12.58±2.84 minutes for conventional interpretation, and 6.50±1.50 minutes for AI-augmented interpretation. AI assistance reduced mean pathologist review time by 48.3% (p<0.001). As no discordance occurred in either AI-related arm, AI error patterns and calibration of confidence scores against diagnostic error could not be evaluated. Conclusion: Within this small, heterogeneous single-centre cohort, AI-supported interpretation was associated with higher exact diagnostic concordance and substantially shorter review time than conventional interpretation alone. The perfect concordance in both AI-related arms should be regarded as hypothesis-generating because selection effects, reference-standard dependence, and the absence of external validation may have influenced the observed estimates.