Pathology

Autonomous whole-slide edge tomography delivers triage-grade cervical cytology without manual review

A fully autonomous cytopathology pipeline combines real-time whole-slide optical tomography with edge AI for cervical liquid-based cytology analysis. The model achieved single-cell AUCs above 0.99 for detecting LSIL, HSIL, and adenocarcinoma, indicating highly accurate lesion recognition. In 1,124 samples from four centers, slide-level AUCs were 0.86–0.91 for LSIL+ and 0.89–0.97 for HSIL+, supporting robust triage performance. AI-derived LSIL counts correlated with HPV positivity and HSIL counts scaled with diagnostic severity, linking outputs to clinically meaningful biology. This platform enables objective, scalable autonomous triage cytology, potentially reserving human review for discordant, equivocal, or high-risk cases. 1

Pan-cancer cfDNA methylome and fragmentome atlas defines 14,202 regions for multicancer detection

Researchers assembled 1,074 cfMeDIP-seq methylation profiles from 11 cancer types, Li-Fraumeni carriers, and healthy controls into a harmonized resource. A uniform computational workflow reduced cross-study technical and biological confounding, enabling direct comparison across cohorts and tumor types. They identified 14,202 pancancer differentially methylated regions for detection plus cancer-specific markers suitable for subtype monitoring. Fragmentomic features, including 5′ end motifs, fragment lengths, and nucleosome footprints, further distinguished cancer from non-cancer samples. Integrating methylome and fragmentome signals improved detection and classification, and validation in 220 independent samples confirmed robustness. This compendium provides a foundational reference for designing and benchmarking future multicancer cfDNA assays in clinical development. 2

M-PACT classifies pediatric CNS tumors from ultra–low-input CSF cfDNA methylomes

M-PACT is a deep neural network that classifies CNS tumors using subnanogram CSF cfDNA methylome input. In an embryonal CNS tumor benchmarking cohort of 79 cases, classification accuracy was 92%, and 88% in an independent 58-case validation cohort. The workflow also handled nonembryonal CNS tumors, balanced tumor genomes, and nonmalignant CSF, suggesting broad diagnostic applicability. Beyond classification, it enables methylation-based cellular deconvolution and copy-number variation detection from the same low-input cfDNA. These data outline a practical CSF liquid-biopsy methylation strategy as a minimally invasive complement to pediatric neuropathologic diagnosis. 3

KEEP: a knowledge-enhanced vision–language foundation model improves cancer pathology, especially in rare tumors

KEEP integrates a comprehensive disease knowledge graph with millions of pathology image–text pairs to pretrain a vision–language model for cancer diagnosis. The graph spans 11,454 diseases and 139,143 attributes, organizing training data into 143,000 ontology-aligned, semantically structured groups. Knowledge-enhanced pretraining aligns visual and textual representations within hierarchical semantic spaces, improving modeling of disease relationships and morphology. Across 18 public benchmarks totaling over 14,000 whole-slide images, KEEP consistently outperformed existing pathology foundation models. On four institutional rare cancer datasets with 926 cases, performance gains were substantial, addressing a critical limitation of prior models. 4