AI and Cancer Research ── Brain ── 2026-10-26
Brain tumours: treatment, trials and AI
- Abstract
Purpose: Artificial intelligence(AI)-based feature extraction from whole slide images (WSIs) shows significant promise for glioma prognosis. However, while foundation pathology models have potential, their features typically lack biological interpretability and clear links to disease mechanisms, limiting their clinical translational value.
Methods: We propose a clustering Vision Transformer (ViT)–based framework for pathological image analysis. Tens of thousands of image patches extracted from each WSI were first clustered into 100 groups. The foundation-model features of the 100 cluster centroids, together with the proportional representation of each cluster across the whole slide, were then integrated and input into the ViT model for prognostic prediction. The image-derived risk scores were further integrated with matched RNA-seq data to identify risk-associated genes and biological pathways through differential expression and pathway enrichment analyses. Prognostic prediction was developed and evaluated in a total of 2552 cases, including 559 cases from Huashan Hospital for training and internal testing, 309 cases from Shanghai International Hospital as external validation cohort 1, and 1684 cases from TCGA as external validation cohort 2. Pathogenomic analyses were performed in 567 TCGA cases with matched WSI and RNA-seq data and further validated in 657 CGGA cases.
Results: Risk scores produced by the prognostic model showed a significant correlation with patient survival (log-rank P < 0.001). When integrated into existing risk stratification frameworks, these risk scores enhanced survival predictions via a nomogram based on clinical information and risk score compared to a traditional clinical-molecular nomogram, achieving a C-index of 0.90, superior calibration and classification accuracy (net reclassification improvement = 0.99, P < 0.001). Specific genes and pathways demonstrated significant associations with risk scores, and the mean expression levels of these key genes exhibited strong prognostic value across both the pathogenomics cohort and the CGGA dataset (log-rank P < 0.001).
Conclusion: These results demonstrate that risk scores derived from pathological imaging can effectively identify key genes and biological pathways critical for stratifying glioma patients into high- and low-risk subgroups. Consequently, therapeutic strategies that modulate the tumor immune microenvironment may be particularly beneficial for high-risk patients identified by this approach.
Journal IF-equivalent: 5.4 (OpenAlex 2-year mean citedness, value as of 2026-10-25, retrieved 2026-10-26; not the official Clarivate IF)Reference: Pan T, Shi Z, Chen J, Yu J, Wu G. 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 neuropathol commun. 2026 Oct 24 [Epub ahead of print]. doi:10.1186/s40478-026-02456-y.Checked: Abstract only - Abstract
BACKGROUND: Optimal surgical management of adult-type diffuse gliomas (ADGs) varies substantially by subtype, making accurate preoperative subtyping crucial for individualized surgical planning. This study aimed to develop a radiomics-based model for preoperative subtyping of ADG in accordance with the 2021 World Health Organization (WHO) Classification of Central Nervous System Tumors. METHODS: A total of 382 patients with ADGs were enrolled and randomly assigned to a training set (n = 267) and an independent test set (n = 115). Tumor masks were automatically segmented and refined by experienced neurosurgeons. Radiomics features were extracted from MRI after standardized preprocessing and dimensionality reduction. Fourteen machine learning classifiers were evaluated, and the best-performing model was optimized and validated on the independent test set. Model performance was assessed using AUC, accuracy, recall, precision, F1 score, Kappa coefficient, and Matthews correlation coefficient. RESULTS: After LASSO (Least Absolute Shrinkage and Selection Operator) regression, 50 key features were retained. Random Forest showed the best preliminary performance and was selected for optimization. On the independent test set, the optimized model achieved a macro-average AUC of 0.9084, with favorable results across other metrics. Texture features contributed most strongly to prediction. CONCLUSION: This radiomics model enables robust preoperative ADG subtyping under the 2021 WHO classification and may support individualized surgical planning using routine MRI features. Owing to its retrospective single-center design, multicenter prospective external validation is required prior to widespread clinical application.
Journal IF-equivalent: 6.3 (OpenAlex 2-year mean citedness, value as of 2026-10-25, retrieved 2026-10-26; not the official Clarivate IF)Reference: Liang Y, Shi J, Zhao C, Liu J, Li K, Yang J, et al. Radiomics‐Assisted Preoperative Subtyping of Adult‐Type Diffuse Gliomas: A Retrospective Study Aligned With the WHO 2021 Classification. CNS Neurosci Ther. 2026 Oct 23;32(10):e71225. doi:10.1002/cns.71225. PMID: 42924459.Checked: Abstract only