AI pathology model stratifies glioma patient risk across 2,552 cases

Researchers have built an artificial intelligence framework that reads whole-slide pathology images of gliomas and converts them into a risk score strongly linked to patient survival, while also pointing to the genes and biological pathways behind that risk. The multicenter study, published in Acta Neuropathologica Communications as an early-access article, was developed and evaluated in a total of 2552 cases, drawing on tissue from three independent sources.

These large pretrained models extract powerful features from digitized tissue slides, but those features typically lack biological interpretability and clear links to disease mechanisms, which limits their value in the clinic. The new framework was designed to keep the predictive strength of foundation-model features while making the output traceable back to genes and pathways.

Technically, the method begins by slicing each whole slide image into tens of thousands of small patches. These patches are then clustered into 100 groups, so that each group represents a recurring pattern of tissue appearance. The foundation-model features of the 100 cluster centroids are combined with the proportional representation of each cluster across the whole slide, and this integrated representation is fed into a clustering Vision Transformer (ViT) model that produces a prognostic risk score for the patient.

The interpretability step comes next. The image-derived risk scores were integrated with matched RNA-seq data, and differential expression and pathway enrichment analyses were used to identify risk-associated genes and biological pathways. In this way, the model does not merely assign a number to a slide; it nominates specific molecular players whose expression patterns track with the imaging-based risk estimate.

The training and evaluation cohort was substantial. A total of 559 cases from Huashan Hospital were used for training and internal testing. External validation cohort 1 comprised 309 cases from Shanghai International Hospital, and external validation cohort 2 comprised 1684 cases from TCGA.

The results were striking. Risk scores produced by the prognostic model showed a significant correlation with patient survival, with a log-rank P value below 0.001. When the scores were integrated into existing risk stratification frameworks, a nomogram combining clinical information with the risk score achieved a C-index of 0.90, outperforming a traditional clinical-molecular nomogram in survival prediction, with superior calibration and classification accuracy. The net reclassification improvement was 0.99, with P below 0.001, indicating improved risk classification performance compared with the traditional clinical-molecular nomogram.

The molecular findings held up in independent data. Pathogenomic analyses were performed in 567 TCGA cases that had both matched whole-slide images and RNA-seq data, and the prognostic value of the key genes was further examined in 657 cases from the CGGA dataset. Specific genes and pathways showed significant associations with the risk scores, and the mean expression levels of these key genes exhibited strong prognostic value across both the pathogenomics cohort and the CGGA dataset, again with log-rank P values below 0.001.

Among the pathways implicated, the authors highlight the tumor immune microenvironment. The study’s conclusion states that therapeutic strategies modulating the tumor immune microenvironment may be particularly beneficial for high-risk patients identified by this approach. This is a hypothesis-generating observation rather than a tested treatment recommendation, since the study was observational and did not evaluate any therapy directly.

The work is an early-access, peer-reviewed accepted article, citable under a permanent DOI and subject to further edits before the final version of record. As with any retrospective multicenter analysis, the reported associations describe prediction and correlation in the studied cohorts; prospective clinical validation would be needed before such risk scores guide treatment decisions. Even so, the study offers a concrete example of how foundation-model pathology features, once anchored to transcriptomic data, can yield both a high-performing prognostic tool and biologically interpretable leads for glioma research.

Subject of Research: AI-based pathology image analysis for prognostic risk stratification and gene pathway identification in glioma

Article Title: Prognostic risk stratification in glioma based on the foundation pathology model with identification of key genes and biological pathways associated with prognosis: a multicenter study

Article References: Pan, T., Shi, Z., Chen, J., Yu, J., & Wu, G. (2026). Prognostic risk stratification in glioma based on the foundation pathology model with identification of key genes and biological pathways associated with prognosis: a multicenter study. Acta Neuropathologica Communications. https://doi.org/10.1186/s40478-026-02456-y

Image Credits: AI Generated

DOI: 10.1186/s40478-026-02456-y

Keywords: glioma, whole slide imaging, Vision Transformer, foundation pathology model, risk stratification, prognosis, RNA-seq, tumor immune microenvironment, TCGA, nomogram, pathogenomics, multicenter study