Pathology

LLMs extract breast cancer pathology fields with high study-specific accuracy but variable validation

This systematic review pooled nine studies covering about 14,161 breast pathology reports and multiple LLM architectures. Reported study-specific accuracies ranged from 87.7% to 97.4%, but metrics and target elements differed across studies. Quality assessment found just over half the studies had low overall risk, with the Outcome domain most variable. Key limitations include inconsistent reference-standard construction, sparse external validation, and poor fairness reporting. 1

Prospective multicenter evaluation shows AI aids prostate biopsy reporting and speeds workflow

In three NHS centres, 1,613 prostate biopsy cases were evaluated, with 1,049 reported using AI assistance. Staged AI review altered diagnosis or Grade Group in 5.4% of reviewed cases and potentially affected management in 1.3%. Concurrent AI-assisted workflows reduced mean turnaround time by 30.1 hours at one site and lowered immunohistochemistry use across sites. Findings support clinical benefit for accuracy and efficiency, though impact on patient outcomes and broader generalizability remain unproven. 2

Self-paced pathology course achieved complete module knowledge-check completion in pilot cohort

A 24-module, self-paced online pathology course enrolled 29 biomedical researchers and scholars in a pilot cohort. All participants (100%) completed objective knowledge checks across every module, and most reported meeting learning objectives. Ninety-four percent reported improved understanding of basic pathology, and 79% reported applying learned material to current roles. This pilot suggests scalable training for nonpathologist scientists, but sample size and single-center setting limit generalizability. 3