Pathology

LLMs show high but heterogeneous accuracy extracting breast cancer pathology elements

This systematic review pooled nine studies covering ~14,161 breast pathology reports and over 30 LLM architectures. Reported study-specific accuracies ranged from 87.7% to 97.4%, with best models approaching human-level extraction. Methodological quality varied and only 55.6% of studies had low risk across PROBAST+AI domains. Key limitations included limited external validation, variable reference-standard development, and gaps in fairness and open-science reporting. 1

Prospective multicentre AI assistance altered diagnoses modestly and sped prostate biopsy workflows

In 1,613 prostate biopsy cases across three NHS centres, 1,049 used a commercial AI assistance workflow. Second-read AI prompted review and changed diagnosis or Grade Group in 5.4% (21/386) of reviewed cases, with 1.3% potentially affecting management. Concurrent-read AI reduced mean turnaround time by 30.1 hours at one site and reduced immunohistochemistry use significantly across sites. This prospective evaluation shows AI can modestly improve diagnostic accuracy and workflow efficiency, though clinical impact on outcomes requires further study. 2

Review: biomarkers and ctDNA central to modern metastatic breast cancer care

This review synthesizes current and emerging biomarkers for metastatic breast cancer, including ER/PR, HER2, PIK3CA, ESR1, BRCA1/2, and PD-L1. It emphasizes routine re-testing at metastasis because of tumor heterogeneity and advocates ctDNA for noninvasive, real-time monitoring. The authors project integration of AI and spatial transcriptomics to shift pathology from static diagnosis to dynamic predictive care. 3