Pathology

DNA/RNA NGS of ERCP specimens markedly improves diagnosis and management of neoplastic bile duct strictures

Over six years, 2865 ERCP brushings, biopsies, and bile samples from 2080 patients underwent BiliSeqV2/V3 DNA/RNA NGS testing. Among 1979 patients with follow-up, NGS showed 82% sensitivity and 98% specificity for neoplastic strictures, versus 44% and 99% for pathology. Combining NGS with cytology/biopsy increased sensitivity to 88% while maintaining high specificity at 97%. In high-risk groups, including PSC, NGS sensitivity reached 74–86% compared with 26–50% for pathology alone. Actionable alterations were found in 20% of BiliSeqV3-positive neoplasms and changed management in 30% of these cases. 1

Additional molecular alterations define intermediate-risk IDH-mutant astrocytomas beyond CDKN2A/B loss

Two cohorts totaling 1207 WHO grade 2–4 IDH-mutant astrocytomas were analyzed for prognostic molecular features. CDK4 amplification, CCND2 amplification, PDGFRA alteration, PIK3R1 mutation, MYCN amplification, and EGFR alteration each associated with shorter overall survival. About 18% of histologic grade 2/3 tumors carried one or more of these markers and had intermediate survival between low-grade and grade 4 tumors. Median survival was 67–82 months for this intermediate group, versus 135–141 months for low-risk and 35–45 months for grade 4. These events flag histologically low-grade tumors with higher risk and merit consideration in future grading frameworks. 2

CHROMA foundation model enables broad, cell-level cytogenetic abnormality detection with less annotation

CHROMA is a single-chromosome foundation model trained on over 4 million images from more than 84,000 cytogenetic specimens. It detects numerical and structural abnormalities, including ultra-rare patterns, in one unified, cell-level framework. The model maintains state-of-the-art accuracy under imbalanced data and challenging imaging conditions while reducing expert annotation workload by about 40%. A built-in risk-control strategy flags uncertain or rare cases for human review, supporting safe use as a screening and triage tool. These capabilities support scalable, more consistent karyotype interpretation, particularly where cytogenetic expertise is limited. 3

Multimodal AI improves distinction of early HCC from high-grade dysplastic nodules

A two-stage multiscale deep-learning model (TMC-net) was developed to separate early hepatocellular carcinoma from high-grade dysplastic nodules on H&E. TMC-net outperformed a pathology foundation model and handcrafted histologic features, capturing subtle morphology consistent with current diagnostic criteria. Heatmaps highlighted relevant regions on virtual slides and improved diagnostic accuracy for junior pathologists. Four marker genes identified by transcriptome analysis were combined with image features in a multimodal classifier. This multimodal model reached AUROCs of 0.8875 internally and 0.95 externally for distinguishing early HCC from high-grade dysplastic nodules. 4