AI-Assisted Prostate Cancer Grading Using Virtual Cytokeratin Segmentation: A Human-AI Collaborative Framework for Enhanced Diagnostic Consistency

Background: Variability in Gleason grading of prostate cancer remains a major challenge for diagnostic consistency. Existing AI-based approaches are limited by labor-intensive annotation processes and binary classification schemes that fail to address diagnostically ambiguouscases. We developed an integrated AI framework that combines virtual cytokeratin–guided epithelial segmentation with multi-class Gleason grading to improve annotation efficiency. Methods: To manage diagnostically uncertain cases, an additional “atypical” category was introduced. The model was trained and validated on 135 prostate specimens (461 slides) from Far Eastern Memorial Hospital. Clinical performance was evaluated in a crossover reader study involving three pathologists who reviewed 60 cases both with and without AI assistance. Results: The pixel-level model achieved an average recall of 0.8781 and a precision of 0.8665. Slide-level classification reached 95.5% accuracy for benign versus malignant discrimination and showed strong agreement with original diagnostic reports (QWK = 0.8605). With AI assistance, pathologists' accuracy and interobserver agreement improved markedly (QWK: 0.63–0.65 to 0.82–0.89). Diagnostic efficiency increased across all experience levels, with time reductions of 9.6–67.7 s per case. Importantly, optimal performance was achieved only through human-AI collaboration rather than AI alone. Conclusions: This study demonstrates a practical and effectiveAI-assisted framework for prostate cancer grading that prioritizes human-AI collaboration over replacement. The virtual cytokeratin approach and introduction of an atypical category address real-world deployment challenges while enhancing diagnostic consistency and efficiency in routine pathology workflows.