A foundation model-based multi-instance learning framework for accurate prediction of lymph node metastasis in prostate cancer from whole slide images.
[BACKGROUND] Nodal involvement (N stage) is a key prognostic factor in prostate cancer (PCa).
APA
Zeng G, Li W, et al. (2026). A foundation model-based multi-instance learning framework for accurate prediction of lymph node metastasis in prostate cancer from whole slide images.. Frontiers in oncology, 16, 1775750. https://doi.org/10.3389/fonc.2026.1775750
MLA
Zeng G, et al.. "A foundation model-based multi-instance learning framework for accurate prediction of lymph node metastasis in prostate cancer from whole slide images.." Frontiers in oncology, vol. 16, 2026, pp. 1775750.
PMID
41858345
Abstract
[BACKGROUND] Nodal involvement (N stage) is a key prognostic factor in prostate cancer (PCa). Conventional imaging and histopathology often have limited sensitivity and inter-observer variability. AI-based computational pathology, using multi-instance learning (MIL) and foundation models, offers a promising approach for accurate and interpretable N stage prediction from H&E-stained whole slide images (WSIs).
[METHODS] In this multicenter retrospective study, we developed a weakly supervised deep learning framework integrating MIL with domain-adapted foundation model encoders (UNI-v2, CONCH, ResNet-50) to predict N stage. WSIs from 280 RHWU patients were used for training and 306 TCGA patients for external validation. Attention heatmaps enabled interpretability, while transcriptomic analyses explored molecular correlates via differential expression and bioinformation analysis.
[RESULTS] The UNI-v2-based model achieved the highest performance (AUC 0.879 in RHWU, 0.850 in TCGA), surpassing CONCH and ResNet-50. Attention heatmaps highlighted tumor-stromal interfaces and poorly differentiated tumor clusters. Transcriptomic analysis identified 94 differentially expressed genes; upregulated genes were enriched in cell cycle, and immune pathways, while downregulated genes involved ion transport and metabolism.
[CONCLUSIONS] This AI-MIL framework accurately predicts nodal involvement in PCa and provides biologically interpretable insights, supporting its potential as a precision oncology tool for risk stratification and treatment planning.
[METHODS] In this multicenter retrospective study, we developed a weakly supervised deep learning framework integrating MIL with domain-adapted foundation model encoders (UNI-v2, CONCH, ResNet-50) to predict N stage. WSIs from 280 RHWU patients were used for training and 306 TCGA patients for external validation. Attention heatmaps enabled interpretability, while transcriptomic analyses explored molecular correlates via differential expression and bioinformation analysis.
[RESULTS] The UNI-v2-based model achieved the highest performance (AUC 0.879 in RHWU, 0.850 in TCGA), surpassing CONCH and ResNet-50. Attention heatmaps highlighted tumor-stromal interfaces and poorly differentiated tumor clusters. Transcriptomic analysis identified 94 differentially expressed genes; upregulated genes were enriched in cell cycle, and immune pathways, while downregulated genes involved ion transport and metabolism.
[CONCLUSIONS] This AI-MIL framework accurately predicts nodal involvement in PCa and provides biologically interpretable insights, supporting its potential as a precision oncology tool for risk stratification and treatment planning.