Retrospective Validation of a Resource-Aware Assistive AI Tool for Screening Prostate Needle Biopsies and Classifying...

Introduction: Prostate needle biopsy reporting is time-consuming because each case typically includes 12-18 cores that require detailed assessment before a final case-level diagnosis is issued. Reporting is also influenced by pathologist expertise, particularly for tumour detection and ISUP Grade Group assignment. This study evaluated an artificial intelligence (AI)-based system designed to identify tumour and provide segmentation-based visual outputs highlighting Gleason patterns and suggesting an ISUP Grade Group. Methods: A strongly supervised approach was used to develop the algorithm. Twenty-seven pathologists annotated 2115 prostate needle biopsy whole-slide images, followed by two levels of senior pathologist reviews. An independent external test dataset of 150 prostate needle biopsy whole-slide images was then evaluated. Ground truth (GT) was established by two genitourinary pathologists, with disagreements resolved by consensus. The same slides were independently reviewed by 11 pathologists without AI assistance, while the AI system analysed the slides in parallel. AI performance and pathologist consensus were compared to GT. Results: For benign versus malignant classification, the AI identified all 109 malignant slides, with no false-negative predictions, while the pathologist consensus missed two malignant slides. The AI correctly classified 38/41 benign slides, compared to 39/41 for the pathologist consensus. For ISUP Grade Group assignment, the AI showed exact agreement with GT in 57/109 malignant cases. The AI more frequently assigned a higher Grade Group than GT. Conclusions: In this retrospective evaluation, the AI system showed high sensitivity for tumour detection and promising ordinal agreement for ISUP Grade Group assignment. These findings support further evaluation of the system within supervised prostate biopsy workflows.