AI-assisted digital pathology improves diagnostic performance and reduces experience-related performance gaps in pathology residents interpreting prostate biopsies

Artificial intelligence (AI) is increasingly integrated into digital pathology, yet evidence regarding its value as a decision-support tool for pathology residents remains limited. We evaluated whether AI-assisted digital pathology improves diagnostic performance in prostate biopsy interpretation and whether its benefit varies according to resident experience and case difficulty. In this prospective within-subject study, 15 pathology residents (PGY-1 to PGY-3) independently reviewed 186 prostate core biopsy whole-slide images (104 adenocarcinomas, 75 benign, and 7 atypical small acinar proliferations) in a fixed sequence (unaided review followed by AI-assisted review of the same slides) using a digital pathology platform, which provided Gleason pattern annotations and tumor region highlighting. Residents were stratified by prior genitourinary (GU) pathology experience. Bayesian hierarchical mixed-effects models fitted by Markov chain Monte Carlo (MCMC) were used to evaluate the effects of AI assistance and resident experience on overall diagnostic accuracy, cancer detection sensitivity, and benign-slide specificity. In exploratory analyses, an ability-weighted slide difficulty metric derived from unaided performance was added as a covariate. AI assistance improved diagnostic performance in both experience groups, with the greatest benefit observed among less experienced residents. Observed accuracy among less experienced residents rose from 62.3% to 75.1% (+ 12.9% points [pp]; 95% CI 7.8 to 18.7) and among experienced residents from 77.9% to 83.0% (+ 5.1 pp; 95% CI 3.2 to 7.5). AI assistance increased the odds of a correct diagnosis (odds ratio [OR] 2.39 (95% CrI: 1.94 to 2.96)), of cancer detection (OR 2.55 (95% CrI: 1.83 to 3.62)), and of correct benign classification (OR 2.76 (95% CrI: 2.04 to 3.74)). The AI-by-experience interaction indicated a smaller relative benefit for experienced residents for accuracy (OR 0.65 (95% CrI: 0.48 to 0.87)) and specificity (OR 0.49 (95% CrI: 0.32 to 0.74)) but not for sensitivity (OR 0.82 (95% CrI: 0.51 to 1.37)). AI display converted 307 of 786 unaided errors (39%) into correct diagnoses and 75 of 1913 correct unaided diagnoses (3.9%) into errors. In exploratory analyses, AI benefit was larger on more difficult slides, consistent with ceiling effects on easy slides. AI-assisted digital pathology improved prostate biopsy interpretation among pathology residents during assisted review, particularly for less experienced trainees and for diagnostically challenging cases. Because the fixed reading order does not separate AI assistance from second-look effects, and because learning and retention were not measured, these findings support AI as a diagnostic decision-support tool for trainees; its educational value remains a hypothesis for longitudinal, counterbalanced studies.