Pathology

Prospective genomic profiling defines the contemporary molecular spectrum of meningioma

This single-center study prospectively profiled 1104 consecutive meningioma samples, mapping their genomic landscape in routine practice conditions. The data provide a large reference framework for integrating histology with recurrent molecular alterations in meningioma reporting. Findings support risk-adapted prognostication and may help identify candidates for targeted or trial-based therapies. Pathologists can use these results to contextualize local sequencing panels and refine integrated diagnostic classifications for meningioma. 1

Label-free autofluorescence plus deep learning virtually subtypes NSCLC and simulates key IHC stains

This study used autofluorescence imaging of unstained NSCLC tissue with deep learning to distinguish benign tissue, adenocarcinoma, squamous carcinoma, and other subtypes. Virtual immunohistochemistry for TTF-1 and p40 was generated directly from label-free images, approximating routine marker panels. Binary and multiclass classification achieved AUCs above 0.981 and 0.996, suggesting near–clinical-grade discrimination. Three thoracic pathologists blindly reviewed virtual IHC, supporting the potential to reduce turnaround and IHC consumption without compromising decisions. If validated locally, such tools could accelerate small-biopsy triage and preserve tissue for predictive testing. 2

HER2-(ultra)low breast cancer scoring shows substantial variability, especially for 0 versus ultralow

In this national Italian project, 121 pathologists scored HER2-(ultra)low breast cancer whole-slide images against an expert panel reference. Overall agreement with experts was substantial at 69%, but dropped to 48% for HER2 0 versus ultralow in challenging cases. Survey responses highlighted pre-analytic and analytic issues, including cold ischemia reporting, decalcification choice, and adequate HER2 controls. The authors advocate standardized criteria, targeted training, and continuous quality assurance for HER2-low and ultralow assessment. Pathologists should treat 0 versus ultralow calls with caution and document pre-analytic variables when reporting in the ADC era. 3

Histopathology foundation transformers outperform CNNs for endometrial molecular subtyping from WSIs

This multicenter study benchmarked histopathology foundation encoders plus attention-based multiple-instance learning against conventional CNNs for endometrial cancer molecular subtyping. The best foundation model configuration achieved macro-AUC 0.860 in cross-validation and maintained a macro-AUC of 0.780 on external validation. CNNs showed greater performance degradation on the external cohort, whereas foundation models preserved higher discrimination and balanced accuracy. Subtype performance was highest for p53-aberrant tumors, aligning with their distinct morphology. These results support using foundation-model pipelines to infer molecular class from H&E slides as an adjunct to targeted sequencing. 4